//! `query_log` and its `domains` dimension table in `querylog.db`.
//!
//! Two shapes live here. The free functions follow the milestone-4 repository
//! idiom — prepare, use, finalize — because retention runs them a handful of
//! times per day, and the API read layer at the bottom of the file runs once
//! per HTTP request. The flush loop is the one hot path in the program, so it
//! gets `BatchWriter`, which owns its three statements for its whole life
//! (`db.zig:360` names this file as the reason `db.zig` carries no statement
//! cache).
//!
//! Every string in a `Row` is borrowed for the duration of the call only:
//! `Stmt.bindText` binds with `SQLITE_TRANSIENT`, so SQLite copies before
//! `writeBatch` returns.
//!
//! The rows are expendable log data. Nothing here retries, and the caller
//! decides what a failed batch means.

const std = @import("std");
const Allocator = std.mem.Allocator;

const db = @import("../db.zig");
const provenance = @import("../provenance.zig");

/// One `query_log` row. The logger applies the privacy transforms of PLAN
/// §11.4 before it builds this, so every domain-bearing field is already
/// whatever the operator agreed to store.
///
/// A `null` text field is a fact the query did not have — no upstream was
/// attempted, no rule matched, no CNAME was uncloaked — and reaches the column
/// as NULL. The three closed enums have no such state: every logged query has a
/// policy verdict, a reason for it and a route, even when the verdict is
/// "not evaluated".
pub const Row = struct {
    timestamp: i64,
    domain: []const u8,
    client_ip: []const u8,
    qtype: ?u16,
    qclass: u16,
    /// Twelve bits: the EDNS extended RCODE the client saw. The column's
    /// `CHECK` bounds it to the same range, so a value this type cannot hold
    /// is one the schema would have refused anyway.
    rcode: u12,
    blocked: bool,
    response_time_us: ?i64,
    cache_hit: ?bool,
    upstream: ?[]const u8,
    group_id: ?i64,
    group_name: ?[]const u8,
    policy_action: provenance.PolicyAction,
    policy_reason: provenance.PolicyReason,
    matched: ?[]const u8,
    source_id: ?i64,
    source_name: ?[]const u8,
    cname_target: ?[]const u8,
    safe_search_target: ?[]const u8,
    route_kind: provenance.RouteKind,
    forward_zone: ?[]const u8,
};

/// The projection grain of the four `bucket_*` tables, in seconds.
///
/// 1800 divides every serving width the Overview offers at or above half an
/// hour (1800, 3600, 21600), which is what lets one grain answer the 24h, 7d
/// and 30d windows exactly. The 1h window is 60-second buckets and is served
/// from the raw rows instead.
pub const grain: i64 = 1800;

/// Floors a timestamp onto the projection grid.
///
/// `@divFloor`, never `@divTrunc` or SQLite's `/`: both truncate toward zero,
/// which for a negative timestamp names the bucket *after* the one the row
/// belongs to. Every producer and every consumer of a `bucket` value in this
/// file goes through this function or an equivalent floor expression.
pub fn bucketOf(timestamp: i64) i64 {
    return @divFloor(timestamp, grain) * grain;
}

/// `bucket_types.qtype` is NOT NULL, so the rows that carry no query type need
/// a value of their own. No qtype is a `u16`, so -1 cannot collide with one.
pub const null_qtype: i64 = -1;

/// The answering resolver's identity, as the routes breakdown defines it: the
/// upstream url for `upstream` rows, the zone for `forward_zone` rows, null
/// everywhere else. The Zig twin of `overview_raw_sql`'s CASE, and the only
/// place the writer decides what a route's source is.
fn routeSourceOf(row: Row) ?[]const u8 {
    return switch (row.route_kind) {
        .upstream => row.upstream,
        .forward_zone => row.forward_zone,
        else => null,
    };
}

const insert_domain_sql = "INSERT OR IGNORE INTO domains (domain) VALUES (?1)";

const select_domain_sql = "SELECT id FROM domains WHERE domain = ?1";

const insert_row_sql =
    \\INSERT INTO query_log
    \\  (timestamp, domain_id, client_ip, qtype, blocked,
    \\   response_time_us, cache_hit, upstream, qclass, rcode,
    \\   group_id, group_name, policy_action, policy_reason, matched,
    \\   source_id, source_name, cname_target, safe_search_target,
    \\   route_kind, forward_zone)
    \\VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10,
    \\        ?11, ?12, ?13, ?14, ?15, ?16, ?17, ?18, ?19, ?20, ?21)
;

/// One `bucket_totals` row's worth of change from a single batch.
const TotalsDelta = struct {
    queries: i64 = 0,
    blocked: i64 = 0,
    cached: i64 = 0,
    rt_sum: i64 = 0,
    rt_count: i64 = 0,
};

/// The delta-map keys borrow the batch's strings. They live only from the
/// aggregation pass to the end of the `writeBatch` that produced them, which is
/// inside the borrow `Row` already documents.
const ClientKey = struct { bucket: i64, client_ip: []const u8 };
const TypeKey = struct { bucket: i64, qtype: i64 };
const RouteKey = struct { bucket: i64, route: provenance.RouteKind, source: ?[]const u8 };

const ClientKeyContext = struct {
    pub fn hash(_: ClientKeyContext, key: ClientKey) u64 {
        var hasher: std.hash.Wyhash = .init(@bitCast(key.bucket));
        hasher.update(key.client_ip);
        return hasher.final();
    }
    pub fn eql(_: ClientKeyContext, a: ClientKey, b: ClientKey) bool {
        return a.bucket == b.bucket and std.mem.eql(u8, a.client_ip, b.client_ip);
    }
};

const RouteKeyContext = struct {
    pub fn hash(_: RouteKeyContext, key: RouteKey) u64 {
        var hasher: std.hash.Wyhash = .init(@bitCast(key.bucket));
        hasher.update(@tagName(key.route));
        hasher.update(&[_]u8{@intFromBool(key.source != null)});
        hasher.update(key.source orelse "");
        return hasher.final();
    }
    pub fn eql(_: RouteKeyContext, a: RouteKey, b: RouteKey) bool {
        return a.bucket == b.bucket and a.route == b.route and sameSource(a.source, b.source);
    }
};

const load_percentage = std.hash_map.default_max_load_percentage;

const upsert_totals_sql =
    \\INSERT INTO bucket_totals (bucket, queries, blocked, cached, rt_sum, rt_count)
    \\VALUES (?1, ?2, ?3, ?4, ?5, ?6)
    \\ON CONFLICT(bucket) DO UPDATE SET
    \\    queries = queries + excluded.queries,
    \\    blocked = blocked + excluded.blocked,
    \\    cached = cached + excluded.cached,
    \\    rt_sum = rt_sum + excluded.rt_sum,
    \\    rt_count = rt_count + excluded.rt_count
;

const upsert_clients_sql =
    \\INSERT INTO bucket_clients (bucket, client_ip, queries)
    \\VALUES (?1, ?2, ?3)
    \\ON CONFLICT(bucket, client_ip) DO UPDATE SET queries = queries + excluded.queries
;

const upsert_types_sql =
    \\INSERT INTO bucket_types (bucket, qtype, count)
    \\VALUES (?1, ?2, ?3)
    \\ON CONFLICT(bucket, qtype) DO UPDATE SET count = count + excluded.count
;

const upsert_routes_sql =
    \\INSERT INTO bucket_routes (bucket, route_kind, source_present, source_text, count)
    \\VALUES (?1, ?2, ?3, ?4, ?5)
    \\ON CONFLICT(bucket, route_kind, source_present, source_text)
    \\DO UPDATE SET count = count + excluded.count
;

/// Owns the prepared statements of the flush loop. Init once, reuse per batch.
///
/// `database` must outlive the writer and must not move: every `Stmt` holds a
/// `*Db`. Neither `Db` nor `Stmt` is thread-safe, so one writer belongs to one
/// task.
///
/// The writer also maintains the four `bucket_*` projections, in the same
/// transaction as the rows they summarise. `gpa` is owned for the writer's life
/// and is used for nothing but the per-batch delta maps, whose capacity is
/// retained across batches and whose contents are dropped at the end of every
/// `writeBatch`.
pub const BatchWriter = struct {
    gpa: Allocator,
    database: *db.Db,
    insert_domain: db.Stmt,
    select_domain: db.Stmt,
    insert_row: db.Stmt,
    upsert_totals: db.Stmt,
    upsert_clients: db.Stmt,
    upsert_types: db.Stmt,
    upsert_routes: db.Stmt,

    totals_deltas: std.AutoHashMapUnmanaged(i64, TotalsDelta) = .empty,
    client_deltas: std.HashMapUnmanaged(ClientKey, i64, ClientKeyContext, load_percentage) = .empty,
    type_deltas: std.AutoHashMapUnmanaged(TypeKey, i64) = .empty,
    route_deltas: std.HashMapUnmanaged(RouteKey, i64, RouteKeyContext, load_percentage) = .empty,

    pub fn init(gpa: Allocator, database: *db.Db) db.Error!BatchWriter {
        var insert_domain = try database.prepare(insert_domain_sql);
        errdefer insert_domain.deinit();
        var select_domain = try database.prepare(select_domain_sql);
        errdefer select_domain.deinit();
        var insert_row = try database.prepare(insert_row_sql);
        errdefer insert_row.deinit();
        var upsert_totals = try database.prepare(upsert_totals_sql);
        errdefer upsert_totals.deinit();
        var upsert_clients = try database.prepare(upsert_clients_sql);
        errdefer upsert_clients.deinit();
        var upsert_types = try database.prepare(upsert_types_sql);
        errdefer upsert_types.deinit();
        const upsert_routes = try database.prepare(upsert_routes_sql);
        return .{
            .gpa = gpa,
            .database = database,
            .insert_domain = insert_domain,
            .select_domain = select_domain,
            .insert_row = insert_row,
            .upsert_totals = upsert_totals,
            .upsert_clients = upsert_clients,
            .upsert_types = upsert_types,
            .upsert_routes = upsert_routes,
        };
    }

    pub fn deinit(self: *BatchWriter) void {
        self.route_deltas.deinit(self.gpa);
        self.type_deltas.deinit(self.gpa);
        self.client_deltas.deinit(self.gpa);
        self.totals_deltas.deinit(self.gpa);
        self.upsert_routes.deinit();
        self.upsert_types.deinit();
        self.upsert_clients.deinit();
        self.upsert_totals.deinit();
        self.insert_row.deinit();
        self.select_domain.deinit();
        self.insert_domain.deinit();
    }

    /// One transaction for the whole batch: the raw rows and every projection
    /// they touch commit together or not at all. Domains are interned through
    /// `INSERT OR IGNORE` followed by `SELECT id`.
    ///
    /// On any failure the transaction rolls back, so a batch is all or
    /// nothing, and the writer stays usable for the next batch.
    pub fn writeBatch(self: *BatchWriter, rows: []const Row) db.Error!void {
        if (rows.len == 0) return;

        // Declared before the aggregation so it runs after every `errdefer`
        // below, and placed before `BEGIN` so an `error.OutOfMemory` fails the
        // batch with no transaction outstanding.
        defer self.clearDeltas();
        try self.aggregate(rows);

        var tx = try db.Tx.begin(self.database);
        // `errdefer`s run in reverse: the statements are released before the
        // ROLLBACK, so no read cursor is still open when it runs.
        errdefer tx.rollback();
        errdefer self.resetAll();

        for (rows) |row| {
            const domain_id = try self.internDomain(row.domain);
            try self.write(row, domain_id);
        }
        try self.applyProjections();
        try tx.commit();
    }

    /// Folds the batch into per-key deltas, so the transaction below runs one
    /// UPSERT per touched key rather than per row. Any slice length: the
    /// logger's 100-row batching is its own policy, not this API's.
    fn aggregate(self: *BatchWriter, rows: []const Row) Allocator.Error!void {
        for (rows) |row| {
            const bucket = bucketOf(row.timestamp);

            const totals = try self.totals_deltas.getOrPut(self.gpa, bucket);
            if (!totals.found_existing) totals.value_ptr.* = .{};
            totals.value_ptr.queries += 1;
            if (row.blocked) totals.value_ptr.blocked += 1;
            if (row.cache_hit orelse false) totals.value_ptr.cached += 1;
            if (row.response_time_us) |us| {
                totals.value_ptr.rt_sum += us;
                totals.value_ptr.rt_count += 1;
            }

            try bump(&self.client_deltas, self.gpa, ClientKey{
                .bucket = bucket,
                .client_ip = row.client_ip,
            });
            try bump(&self.type_deltas, self.gpa, TypeKey{
                .bucket = bucket,
                .qtype = if (row.qtype) |v| @as(i64, v) else null_qtype,
            });
            try bump(&self.route_deltas, self.gpa, RouteKey{
                .bucket = bucket,
                .route = row.route_kind,
                .source = routeSourceOf(row),
            });
        }
    }

    /// One more row on `key`. `map` is any of the three count maps — they
    /// differ only in their key type and its hashing.
    fn bump(map: anytype, gpa: Allocator, key: anytype) Allocator.Error!void {
        const entry = try map.getOrPut(gpa, key);
        if (!entry.found_existing) entry.value_ptr.* = 0;
        entry.value_ptr.* += 1;
    }

    fn applyProjections(self: *BatchWriter) db.Error!void {
        var totals = self.totals_deltas.iterator();
        while (totals.next()) |entry| {
            const stmt = &self.upsert_totals;
            try stmt.reset();
            try stmt.bindInt(1, entry.key_ptr.*);
            try stmt.bindInt(2, entry.value_ptr.queries);
            try stmt.bindInt(3, entry.value_ptr.blocked);
            try stmt.bindInt(4, entry.value_ptr.cached);
            try stmt.bindInt(5, entry.value_ptr.rt_sum);
            try stmt.bindInt(6, entry.value_ptr.rt_count);
            try stmt.exec();
        }

        var clients = self.client_deltas.iterator();
        while (clients.next()) |entry| {
            const stmt = &self.upsert_clients;
            try stmt.reset();
            try stmt.bindInt(1, entry.key_ptr.bucket);
            try stmt.bindText(2, entry.key_ptr.client_ip);
            try stmt.bindInt(3, entry.value_ptr.*);
            try stmt.exec();
        }

        var types = self.type_deltas.iterator();
        while (types.next()) |entry| {
            const stmt = &self.upsert_types;
            try stmt.reset();
            try stmt.bindInt(1, entry.key_ptr.bucket);
            try stmt.bindInt(2, entry.key_ptr.qtype);
            try stmt.bindInt(3, entry.value_ptr.*);
            try stmt.exec();
        }

        var routes = self.route_deltas.iterator();
        while (routes.next()) |entry| {
            const stmt = &self.upsert_routes;
            try stmt.reset();
            try stmt.bindInt(1, entry.key_ptr.bucket);
            try stmt.bindText(2, @tagName(entry.key_ptr.route));
            try stmt.bindInt(3, @intFromBool(entry.key_ptr.source != null));
            try stmt.bindText(4, entry.key_ptr.source orelse "");
            try stmt.bindInt(5, entry.value_ptr.*);
            try stmt.exec();
        }
    }

    /// Capacity is retained; only the entries go. The keys borrow the batch's
    /// strings, so nothing may outlive the `writeBatch` that filled them.
    fn clearDeltas(self: *BatchWriter) void {
        self.totals_deltas.clearRetainingCapacity();
        self.client_deltas.clearRetainingCapacity();
        self.type_deltas.clearRetainingCapacity();
        self.route_deltas.clearRetainingCapacity();
    }

    fn internDomain(self: *BatchWriter, domain: []const u8) db.Error!i64 {
        try self.insert_domain.reset();
        try self.insert_domain.bindText(1, domain);
        try self.insert_domain.exec();

        try self.select_domain.reset();
        try self.select_domain.bindText(1, domain);
        // The insert above either created the row or found it already there,
        // so a miss means the table changed under this connection.
        if (!try self.select_domain.step()) return error.NotFound;
        const id = self.select_domain.columnInt(0);
        // A statement stopped on a row keeps its cursor open until it is
        // reset; the transaction must not carry that to the next row.
        try self.select_domain.reset();
        return id;
    }

    fn write(self: *BatchWriter, row: Row, domain_id: i64) db.Error!void {
        var stmt = &self.insert_row;
        try stmt.reset();
        try stmt.bindInt(1, row.timestamp);
        try stmt.bindInt(2, domain_id);
        try stmt.bindText(3, row.client_ip);
        try bindIntOrNull(stmt, 4, if (row.qtype) |v| @as(i64, v) else null);
        try stmt.bindBool(5, row.blocked);
        try bindIntOrNull(stmt, 6, row.response_time_us);
        try bindIntOrNull(stmt, 7, if (row.cache_hit) |v| @as(i64, @intFromBool(v)) else null);
        try stmt.bindTextOrNull(8, row.upstream);
        try stmt.bindInt(9, row.qclass);
        try stmt.bindInt(10, row.rcode);
        try bindIntOrNull(stmt, 11, row.group_id);
        try stmt.bindTextOrNull(12, row.group_name);
        try stmt.bindText(13, @tagName(row.policy_action));
        try stmt.bindText(14, @tagName(row.policy_reason));
        try stmt.bindTextOrNull(15, row.matched);
        try bindIntOrNull(stmt, 16, row.source_id);
        try stmt.bindTextOrNull(17, row.source_name);
        try stmt.bindTextOrNull(18, row.cname_target);
        try stmt.bindTextOrNull(19, row.safe_search_target);
        try stmt.bindText(20, @tagName(row.route_kind));
        try stmt.bindTextOrNull(21, row.forward_zone);
        try stmt.exec();
    }

    /// Best effort: this runs on the failure path, where the error that
    /// matters is the one already on its way to the caller.
    fn resetAll(self: *BatchWriter) void {
        self.upsert_routes.reset() catch {};
        self.upsert_types.reset() catch {};
        self.upsert_clients.reset() catch {};
        self.upsert_totals.reset() catch {};
        self.insert_row.reset() catch {};
        self.select_domain.reset() catch {};
        self.insert_domain.reset() catch {};
    }
};

fn bindIntOrNull(stmt: *db.Stmt, idx: c_int, value: ?i64) db.Error!void {
    if (value) |v| return stmt.bindInt(idx, v);
    return stmt.bindNull(idx);
}

/// What one prune did, and where coverage now begins.
pub const PruneResult = struct {
    deleted: i64,
    /// The watermark after the prune, which is what a later `availableSince`
    /// will return. Handed back so the caller need not re-read it.
    available_since: i64,
};

/// Deletes every `query_log` row strictly older than `cutoff_ts` and advances
/// the coverage watermark to the same cutoff, in one transaction.
///
/// **The two are one operation, not two.** The watermark is the promise that
/// every query since it is still in the file; a delete that commits without the
/// advance breaks that promise, and an advance that commits without the delete
/// hides rows the file still holds. Either failure rolls both back, and the
/// caller retries the whole thing on its next pass.
///
/// The watermark never moves backward: `max` is what makes a prune with a
/// cutoff older than the file's own creation a no-op on it rather than a
/// regression. Orphaned `domains` rows stay — it is a dimension table,
/// re-interning a name costs one indexed insert, and §11.3 asks for no
/// collection.
///
/// The four projections are pruned in the same transaction. Buckets entirely
/// before the cutoff's own bucket are deleted outright; a cutoff that does not
/// land on the grid leaves one straddling bucket, which is recomputed from the
/// rows that survived rather than reduced by an estimate. Any failure in any of
/// that rolls back the raw delete and the watermark advance with it.
pub fn pruneOlderThan(database: *db.Db, cutoff_ts: i64) db.Error!PruneResult {
    var tx = try db.Tx.begin(database);
    errdefer tx.rollback();

    var deleting = try database.prepare("DELETE FROM query_log WHERE timestamp < ?1");
    defer deleting.deinit();
    try deleting.bindInt(1, cutoff_ts);
    try deleting.exec();
    const deleted = database.changes();

    const grid = bucketOf(cutoff_ts);
    for (delete_before_bucket_sql) |sql| {
        var stmt = try database.prepare(sql);
        defer stmt.deinit();
        try stmt.bindInt(1, grid);
        try stmt.exec();
    }
    // On the grid there is nothing partial to fix: the delete above took whole
    // buckets and the loop took their projection rows.
    if (cutoff_ts != grid) try recomputeBucket(database, grid);

    var advancing = try database.prepare(
        "UPDATE querylog_meta SET available_since = max(available_since, ?1) WHERE id = 1",
    );
    defer advancing.deinit();
    try advancing.bindInt(1, cutoff_ts);
    try advancing.exec();

    const watermark = try database.queryInt("SELECT available_since FROM querylog_meta WHERE id = 1");
    try tx.commit();
    return .{ .deleted = deleted, .available_since = watermark };
}

const delete_before_bucket_sql = [_][]const u8{
    "DELETE FROM bucket_totals WHERE bucket < ?1",
    "DELETE FROM bucket_clients WHERE bucket < ?1",
    "DELETE FROM bucket_types WHERE bucket < ?1",
    "DELETE FROM bucket_routes WHERE bucket < ?1",
};

const delete_one_bucket_sql = [_][]const u8{
    "DELETE FROM bucket_totals WHERE bucket = ?1",
    "DELETE FROM bucket_clients WHERE bucket = ?1",
    "DELETE FROM bucket_types WHERE bucket = ?1",
    "DELETE FROM bucket_routes WHERE bucket = ?1",
};

/// `?1` is the bucket start and `?2` its exclusive end. Each statement rebuilds
/// one projection table's rows for that bucket from the raw rows still in it.
///
/// `HAVING count(*) > 0` on the totals statement is what keeps an emptied
/// bucket from being written back as a row of zeros: a bare aggregate always
/// produces one row, and a from-scratch recomputation produces none. The other
/// three group by a key, so an empty range already yields nothing.
const recompute_bucket_sql = [_][]const u8{
    \\INSERT INTO bucket_totals (bucket, queries, blocked, cached, rt_sum, rt_count)
    \\SELECT ?1, count(*), coalesce(sum(blocked <> 0), 0), coalesce(sum(cache_hit = 1), 0),
    \\       coalesce(sum(response_time_us), 0), count(response_time_us)
    \\  FROM query_log
    \\ WHERE timestamp >= ?1 AND timestamp < ?2
    \\HAVING count(*) > 0
    ,
    \\INSERT INTO bucket_clients (bucket, client_ip, queries)
    \\SELECT ?1, client_ip, count(*)
    \\  FROM query_log
    \\ WHERE timestamp >= ?1 AND timestamp < ?2
    \\ GROUP BY client_ip
    ,
    \\INSERT INTO bucket_types (bucket, qtype, count)
    \\SELECT ?1, coalesce(qtype, -1), count(*)
    \\  FROM query_log
    \\ WHERE timestamp >= ?1 AND timestamp < ?2
    \\ GROUP BY coalesce(qtype, -1)
    ,
    \\INSERT INTO bucket_routes (bucket, route_kind, source_present, source_text, count)
    \\SELECT ?1, route_kind, source IS NOT NULL, coalesce(source, ''), count(*)
    \\  FROM (SELECT route_kind,
    \\               CASE route_kind
    \\                    WHEN 'upstream' THEN upstream
    \\                    WHEN 'forward_zone' THEN forward_zone
    \\               END AS source
    \\          FROM query_log
    \\         WHERE timestamp >= ?1 AND timestamp < ?2)
    \\ GROUP BY route_kind, source
    ,
};

/// Replaces one bucket's projection rows with a recomputation from the raw rows
/// that are still in it. The caller owns the transaction.
fn recomputeBucket(database: *db.Db, bucket: i64) db.Error!void {
    for (delete_one_bucket_sql) |sql| {
        var stmt = try database.prepare(sql);
        defer stmt.deinit();
        try stmt.bindInt(1, bucket);
        try stmt.exec();
    }
    for (recompute_bucket_sql) |sql| {
        var stmt = try database.prepare(sql);
        defer stmt.deinit();
        try stmt.bindInt(1, bucket);
        try stmt.bindInt(2, bucket + grain);
        try stmt.exec();
    }
}

/// The oldest timestamp this file can still answer for. A query window that
/// starts before it is incomplete, and the API says so rather than charting the
/// gap as zero.
pub fn availableSince(database: *db.Db) db.Error!i64 {
    return database.queryInt("SELECT available_since FROM querylog_meta WHERE id = 1");
}

/// `PRAGMA wal_checkpoint(TRUNCATE)`: moves the WAL into the database and
/// truncates it to zero bytes, which is what keeps a day of log writes from
/// growing the WAL past the free space the disk monitor watches.
///
/// SQLite reports a checkpoint blocked by a concurrent reader in the row it
/// returns, not as an error code, so a blocked checkpoint is not an error
/// here. Retention checkpoints after every prune, so the next pass retries.
/// On a database that is not in WAL mode the pragma is a no-op.
pub fn checkpointTruncate(database: *db.Db) db.Error!void {
    return database.exec("PRAGMA wal_checkpoint(TRUNCATE);");
}

/// Rewrites the whole file. Retention runs this rarely by design — on an SD
/// card a full rewrite is the most expensive thing this program does.
pub fn vacuum(database: *db.Db) db.Error!void {
    return database.exec("VACUUM;");
}

pub fn countRows(database: *db.Db) db.Error!i64 {
    return database.queryInt("SELECT count(*) FROM query_log");
}

pub fn countDomains(database: *db.Db) db.Error!i64 {
    return database.queryInt("SELECT count(*) FROM domains");
}

// ---------------------------------------------------------------------------
// the API read layer (`GET /api/queries`, `GET /api/overview`)
// ---------------------------------------------------------------------------

/// One row of `GET /api/queries`, joined back through the `domains` dimension.
///
/// A summary projection, deliberately narrower than `QueryDetail`: the list is
/// a table the operator scans, and the full provenance of a row is one request
/// away at `GET /api/queries/{id}`.
///
/// The nullable text columns read a NULL as `""` — the same convention
/// `Stmt.columnText` already uses. None of them is ever written as an empty
/// string, so the mapping loses nothing and the API layer can treat `""` as
/// "absent".
pub const QueryRow = struct {
    id: i64,
    ts: i64,
    domain: []const u8,
    client_ip: []const u8,
    qtype: ?u16,
    qclass: u16,
    rcode: u12,
    blocked: bool,
    response_time_us: ?i64,
    cache_hit: ?bool,
    upstream: []const u8,
    policy_action: provenance.PolicyAction,
    policy_reason: provenance.PolicyReason,
    route_kind: provenance.RouteKind,
};

/// Everything one `query_log` row records about one query, for
/// `GET /api/queries/{id}`.
///
/// Same NULL-reads-as-`""` convention as `QueryRow`, and the same closed enums:
/// a stored value the schema does not define is `error.Mismatch`, never passed
/// through as text.
pub const QueryDetail = struct {
    id: i64,
    ts: i64,
    domain: []const u8,
    client_ip: []const u8,
    qtype: ?u16,
    qclass: u16,
    rcode: u12,
    blocked: bool,
    response_time_us: ?i64,
    cache_hit: ?bool,
    upstream: []const u8,
    group_id: ?i64,
    group_name: []const u8,
    policy_action: provenance.PolicyAction,
    policy_reason: provenance.PolicyReason,
    matched: []const u8,
    source_id: ?i64,
    source_name: []const u8,
    cname_target: []const u8,
    safe_search_target: []const u8,
    route_kind: provenance.RouteKind,
    forward_zone: []const u8,
};

/// Every field is an independent narrowing; `null` means "do not filter on it".
///
/// `since` is inclusive and `until` is exclusive, so adjacent windows tile
/// without double-counting a row on the boundary.
pub const QueryFilter = struct {
    limit: u32 = 100,
    /// Keyset cursor: only rows with a strictly smaller `id`. Rows come back
    /// newest-first, so this is the id of the last row of the previous page.
    before: ?i64 = null,
    /// Matched case-insensitively for ASCII, which is what SQLite's `LIKE`
    /// does and what a domain search wants.
    domain_substring: ?[]const u8 = null,
    client: ?[]const u8 = null,
    blocked: ?bool = null,
    since: ?i64 = null,
    until: ?i64 = null,
};

/// Ruling 11 caps the page at 1000; the repository enforces it too, so a caller
/// that forgets cannot ask this connection for the whole table.
pub const max_limit: u32 = 1000;

const select_head =
    \\SELECT q.id, q.timestamp, d.domain, q.client_ip, q.qtype, q.blocked,
    \\       q.response_time_us, q.cache_hit, q.upstream, q.qclass, q.rcode,
    \\       q.policy_action, q.policy_reason, q.route_kind
    \\  FROM query_log q JOIN domains d ON d.id = q.domain_id
;

/// The escape character of `where_domain`. SQLite does not give string literals
/// C escapes, so `'\'` in the SQL text is one backslash.
const like_escape = '\\';

const where_before = " q.id < ?";
const where_domain = " d.domain LIKE ? ESCAPE '\\'";
const where_client = " q.client_ip = ?";
const where_blocked = " q.blocked = ?";
const where_since = " q.timestamp >= ?";
const where_until = " q.timestamp < ?";
const select_tail = " ORDER BY q.id DESC LIMIT ?";

const where_keyword = " WHERE";
const and_keyword = " AND";

/// Assembles the statement from the fixed fragments above and nothing else.
///
/// **No value ever reaches this buffer.** Every filter contributes a `?` and is
/// bound afterwards, in the order the predicates were appended: an unnumbered
/// parameter takes the next free index, so append order and bind order are the
/// same single contract.
const Sql = struct {
    /// `where_keyword` is longer than `and_keyword` and is used at most once,
    /// so counting six of it bounds every reachable combination.
    const capacity = select_head.len + 6 * where_keyword.len + select_tail.len +
        where_before.len + where_domain.len + where_client.len +
        where_blocked.len + where_since.len + where_until.len;

    buf: [capacity]u8 = undefined,
    len: usize = 0,
    has_where: bool = false,

    fn put(self: *Sql, fragment: []const u8) void {
        @memcpy(self.buf[self.len..][0..fragment.len], fragment);
        self.len += fragment.len;
    }

    fn predicate(self: *Sql, fragment: []const u8) void {
        self.put(if (self.has_where) and_keyword else where_keyword);
        self.has_where = true;
        self.put(fragment);
    }

    fn text(self: *const Sql) []const u8 {
        return self.buf[0..self.len];
    }
};

/// Rows come back newest-first (`id DESC`). Every string is allocated from
/// `arena`, including the list's own storage, so the caller frees the whole
/// result by resetting the arena — there is nothing to unwind on failure.
pub fn selectQueries(database: *db.Db, arena: Allocator, filter: QueryFilter) db.Error!std.ArrayList(QueryRow) {
    var sql: Sql = .{};
    sql.put(select_head);
    if (filter.before != null) sql.predicate(where_before);
    if (filter.domain_substring != null) sql.predicate(where_domain);
    if (filter.client != null) sql.predicate(where_client);
    if (filter.blocked != null) sql.predicate(where_blocked);
    if (filter.since != null) sql.predicate(where_since);
    if (filter.until != null) sql.predicate(where_until);
    sql.put(select_tail);

    var stmt = try database.prepare(sql.text());
    defer stmt.deinit();

    var idx: c_int = 0;
    if (filter.before) |v| {
        idx += 1;
        try stmt.bindInt(idx, v);
    }
    if (filter.domain_substring) |v| {
        idx += 1;
        try stmt.bindText(idx, try likePattern(arena, v));
    }
    if (filter.client) |v| {
        idx += 1;
        try stmt.bindText(idx, v);
    }
    if (filter.blocked) |v| {
        idx += 1;
        try stmt.bindBool(idx, v);
    }
    if (filter.since) |v| {
        idx += 1;
        try stmt.bindInt(idx, v);
    }
    if (filter.until) |v| {
        idx += 1;
        try stmt.bindInt(idx, v);
    }
    idx += 1;
    try stmt.bindInt(idx, @min(filter.limit, max_limit));

    var out: std.ArrayList(QueryRow) = .empty;
    while (try stmt.step()) {
        try out.append(arena, .{
            .id = stmt.columnInt(0),
            .ts = stmt.columnInt(1),
            .domain = try stmt.columnTextAlloc(arena, 2),
            .client_ip = try stmt.columnTextAlloc(arena, 3),
            .qtype = if (stmt.isNull(4)) null else std.math.cast(u16, stmt.columnInt(4)) orelse
                return error.Mismatch,
            .blocked = stmt.columnBool(5),
            .response_time_us = if (stmt.isNull(6)) null else stmt.columnInt(6),
            .cache_hit = if (stmt.isNull(7)) null else stmt.columnBool(7),
            .upstream = try stmt.columnTextAlloc(arena, 8),
            .qclass = try columnU16(&stmt, 9),
            .rcode = try columnU12(&stmt, 10),
            .policy_action = try provenance.parse(provenance.PolicyAction, stmt.columnText(11)),
            .policy_reason = try provenance.parse(provenance.PolicyReason, stmt.columnText(12)),
            .route_kind = try provenance.parse(provenance.RouteKind, stmt.columnText(13)),
        });
    }
    return out;
}

/// A `NOT NULL` integer column that the schema bounds to 16 bits. A value
/// outside that range means the row came from something other than this schema.
fn columnU16(stmt: *db.Stmt, col: c_int) db.Error!u16 {
    return std.math.cast(u16, stmt.columnInt(col)) orelse error.Mismatch;
}

/// The `rcode` column, which the schema's `CHECK` bounds to twelve bits. This
/// build cannot write a wider value — the field is a `u12` all the way from the
/// handler — so a row that carries one was written by something else, and is
/// `error.Mismatch` rather than a value truncated into shape.
fn columnU12(stmt: *db.Stmt, col: c_int) db.Error!u12 {
    return std.math.cast(u12, stmt.columnInt(col)) orelse error.Mismatch;
}

const select_detail_sql =
    \\SELECT q.id, q.timestamp, d.domain, q.client_ip, q.qtype, q.blocked,
    \\       q.response_time_us, q.cache_hit, q.upstream, q.qclass, q.rcode,
    \\       q.group_id, q.group_name, q.policy_action, q.policy_reason,
    \\       q.matched, q.source_id, q.source_name, q.cname_target,
    \\       q.safe_search_target, q.route_kind, q.forward_zone
    \\  FROM query_log q JOIN domains d ON d.id = q.domain_id
    \\ WHERE q.id = ?1
;

/// One row's full provenance, or `null` when no row has that id — which is what
/// an id the operator kept from before a retention pass looks like, and is a
/// 404 rather than an error.
///
/// Every string is allocated from `arena`, on the same terms as
/// `selectQueries`.
pub fn detailById(database: *db.Db, arena: Allocator, id: i64) db.Error!?QueryDetail {
    var stmt = try database.prepare(select_detail_sql);
    defer stmt.deinit();
    try stmt.bindInt(1, id);
    if (!try stmt.step()) return null;

    return .{
        .id = stmt.columnInt(0),
        .ts = stmt.columnInt(1),
        .domain = try stmt.columnTextAlloc(arena, 2),
        .client_ip = try stmt.columnTextAlloc(arena, 3),
        .qtype = if (stmt.isNull(4)) null else try columnU16(&stmt, 4),
        .blocked = stmt.columnBool(5),
        .response_time_us = if (stmt.isNull(6)) null else stmt.columnInt(6),
        .cache_hit = if (stmt.isNull(7)) null else stmt.columnBool(7),
        .upstream = try stmt.columnTextAlloc(arena, 8),
        .qclass = try columnU16(&stmt, 9),
        .rcode = try columnU12(&stmt, 10),
        .group_id = if (stmt.isNull(11)) null else stmt.columnInt(11),
        .group_name = try stmt.columnTextAlloc(arena, 12),
        .policy_action = try provenance.parse(provenance.PolicyAction, stmt.columnText(13)),
        .policy_reason = try provenance.parse(provenance.PolicyReason, stmt.columnText(14)),
        .matched = try stmt.columnTextAlloc(arena, 15),
        .source_id = if (stmt.isNull(16)) null else stmt.columnInt(16),
        .source_name = try stmt.columnTextAlloc(arena, 17),
        .cname_target = try stmt.columnTextAlloc(arena, 18),
        .safe_search_target = try stmt.columnTextAlloc(arena, 19),
        .route_kind = try provenance.parse(provenance.RouteKind, stmt.columnText(20)),
        .forward_zone = try stmt.columnTextAlloc(arena, 21),
    };
}

/// Wraps `needle` in `%` and neutralises the two `LIKE` metacharacters, so a
/// user searching for `a_b` gets domains containing `a_b` and not domains
/// containing `axb`. The escape character escapes itself.
fn likePattern(arena: Allocator, needle: []const u8) Allocator.Error![]const u8 {
    var out: std.ArrayList(u8) = try .initCapacity(arena, needle.len * 2 + 2);
    out.appendAssumeCapacity('%');
    for (needle) |ch| {
        if (ch == '%' or ch == '_' or ch == like_escape) out.appendAssumeCapacity(like_escape);
        out.appendAssumeCapacity(ch);
    }
    out.appendAssumeCapacity('%');
    return out.items;
}

/// The Overview rollup for one window. `avg_response_time_us` is `null` when no
/// row in the window recorded a response time. The mean is derived from a sum
/// and a count rather than SQL's `avg`, which returns REAL: `Stmt` reads
/// integers, and integer microseconds are exact.
pub const StatsTotals = struct {
    queries: u64,
    blocked: u64,
    distinct_clients: u64,
    avg_response_time_us: ?i64,
};

/// `count` and `sum` over non-negative columns cannot go negative; a negative
/// value means the row came from something other than this schema.
fn countOf(value: i64) db.Error!u64 {
    if (value < 0) return error.Mismatch;
    return @intCast(value);
}

/// One bucket of the Overview timeseries. `ts` is the bucket's inclusive start.
pub const Bucket = struct {
    ts: i64,
    queries: u64,
    blocked: u64,
    cached: u64,
};

/// One row of the Overview type breakdown. `qtype` is nullable in the schema, so
/// the rows that carry no type group into a row of their own rather than
/// disappearing from a breakdown that claims to add up.
///
/// No name field: the only qtype-name table lives in the admin, and a second
/// copy here would drift out of agreement with it.
///
/// The list carries no zero rows: a type absent from the window is absent from
/// it. Its order is an order, not an identity — a caller keys on the `qtype`
/// value, never on a row's position, because a rank change between refreshes
/// moves rows and must not move what they mean.
pub const TypeCount = struct {
    qtype: ?u16,
    count: u64,
};

/// One row of the Overview route breakdown: how a slice of the window was
/// answered.
///
/// `source` is the answering resolver's identity and nothing else — the
/// upstream url for `upstream` rows, the zone for `forward_zone` rows, null
/// everywhere else. It is deliberately not `source_name`, which names the
/// blocklist a block came from and would read as an upstream here.
///
/// A null source on an `upstream` row is its own group, not a dropped row: it
/// is a real state of the log and the caller labels it.
pub const RouteCount = struct {
    route: provenance.RouteKind,
    source: ?[]const u8,
    count: u64,
};

/// How many clients the Overview names before the rest become `other`. Eight is
/// what one legend can carry without becoming a second table.
pub const max_client_series = 8;

/// One named client's series. `buckets` is always the caller's bucket count
/// long, zero-filled, and aligned exactly like `Overview.buckets`.
pub const ClientSeries = struct {
    client: []const u8,
    buckets: []const u64,
};

/// Named clients are ranked by in-window total, ties broken by address, so the
/// cut at `max_client_series` is the same cut on every read over the same data.
///
/// `other` is always present and always bucket-count sized — including for an
/// empty window and for a window with eight clients or fewer. A caller charting
/// a stack must not have to invent the residual series.
pub const ClientsBreakdown = struct {
    clients: []const ClientSeries,
    other: []const u64,
};

fn zeroedBuckets(arena: Allocator, bucket_count: u32) Allocator.Error![]u64 {
    const buckets = try arena.alloc(u64, bucket_count);
    @memset(buckets, 0);
    return buckets;
}

// ---------------------------------------------------------------------------
// the Overview read path (milestone 36)
// ---------------------------------------------------------------------------

/// Everything one Overview response reports about one window, from one read of
/// one already-open transaction.
///
/// Storage owns these scalars: `since`, `bucket_seconds` and `bucket_count` are
/// numbers, not a web module's period type, so nothing here has to know the
/// period grammar the handler enforces.
pub const Overview = struct {
    totals: StatsTotals,
    /// `bucket_count` entries, one per bucket of the window, zero-filled and
    /// aligned to `since`.
    buckets: []const Bucket,
    clients: ClientsBreakdown,
    types: []const TypeCount,
    routes: []const RouteCount,
};

/// One client's running window total beside its per-bucket series, so the cut
/// at `max_client_series` can be made after the whole window is folded rather
/// than by a second pass over the data.
const ClientAccum = struct {
    client: []const u8,
    total: u64,
    buckets: []u64,
};

/// The shared shape both paths fold into. Every allocation is from `arena`, so
/// there is nothing to unwind on failure.
const Accum = struct {
    arena: Allocator,
    since: i64,
    width: i64,
    bucket_count: u32,
    buckets: []Bucket,
    rt_sum: i64 = 0,
    rt_count: i64 = 0,
    clients: std.ArrayList(ClientAccum) = .empty,
    /// `client_ip` to its index in `clients`. The names are arena copies, so
    /// they outlive the statement that produced them.
    client_index: std.StringHashMapUnmanaged(u32) = .empty,
    types: std.ArrayList(TypeCount) = .empty,
    type_index: std.AutoHashMapUnmanaged(i64, u32) = .empty,
    /// Household cardinality: a handful of upstreams plus six route kinds. A
    /// linear scan is the whole index this needs.
    routes: std.ArrayList(RouteCount) = .empty,

    fn init(arena: Allocator, since: i64, width: i64, bucket_count: u32) Allocator.Error!Accum {
        const buckets = try arena.alloc(Bucket, bucket_count);
        for (buckets, 0..) |*bucket, i| bucket.* = .{
            .ts = since + width * @as(i64, @intCast(i)),
            .queries = 0,
            .blocked = 0,
            .cached = 0,
        };
        return .{ .arena = arena, .since = since, .width = width, .bucket_count = bucket_count, .buckets = buckets };
    }

    /// The serving-bucket index a grid bucket or a raw timestamp falls in.
    /// Both paths bound their reads by the window already; the check is what
    /// keeps a schema surprise from writing past the slice.
    fn indexOf(self: *const Accum, timestamp: i64) db.Error!usize {
        const offset = @divFloor(timestamp - self.since, self.width);
        const index = std.math.cast(usize, offset) orelse return error.Mismatch;
        if (index >= self.buckets.len) return error.Mismatch;
        return index;
    }

    fn addTotals(self: *Accum, index: usize, queries: u64, blocked: u64, cached: u64) void {
        self.buckets[index].queries += queries;
        self.buckets[index].blocked += blocked;
        self.buckets[index].cached += cached;
    }

    fn addClient(self: *Accum, client: []const u8, index: usize, count: u64) Allocator.Error!void {
        const slot = try self.client_index.getOrPut(self.arena, client);
        if (!slot.found_existing) {
            const owned = try self.arena.dupe(u8, client);
            slot.key_ptr.* = owned;
            slot.value_ptr.* = @intCast(self.clients.items.len);
            try self.clients.append(self.arena, .{
                .client = owned,
                .total = 0,
                .buckets = try zeroedBuckets(self.arena, self.bucket_count),
            });
        }
        const accum = &self.clients.items[slot.value_ptr.*];
        accum.total += count;
        accum.buckets[index] += count;
    }

    /// `qtype` is the stored encoding: `null_qtype` for the rows that carry no
    /// query type.
    fn addType(self: *Accum, qtype: i64, count: u64) db.Error!void {
        const slot = try self.type_index.getOrPut(self.arena, qtype);
        if (!slot.found_existing) {
            slot.value_ptr.* = @intCast(self.types.items.len);
            try self.types.append(self.arena, .{
                .qtype = if (qtype == null_qtype) null else std.math.cast(u16, qtype) orelse
                    return error.Mismatch,
                .count = 0,
            });
        }
        self.types.items[slot.value_ptr.*].count += count;
    }

    fn addRoute(self: *Accum, route: provenance.RouteKind, source: ?[]const u8, count: u64) Allocator.Error!void {
        for (self.routes.items) |*existing| {
            if (existing.route == route and sameSource(existing.source, source)) {
                existing.count += count;
                return;
            }
        }
        try self.routes.append(self.arena, .{
            .route = route,
            .source = if (source) |s| try self.arena.dupe(u8, s) else null,
            .count = count,
        });
    }

    fn finish(self: *Accum) Allocator.Error!Overview {
        var queries: u64 = 0;
        var blocked: u64 = 0;
        for (self.buckets) |bucket| {
            queries += bucket.queries;
            blocked += bucket.blocked;
        }

        // Every comparator below is a total order on a unique key, so an
        // unstable sort still produces one list for one window.
        std.sort.pdq(TypeCount, self.types.items, {}, typeBefore);
        std.sort.pdq(RouteCount, self.routes.items, {}, routeBefore);
        std.sort.pdq(ClientAccum, self.clients.items, {}, clientBefore);

        const named = @min(self.clients.items.len, max_client_series);
        const other = try zeroedBuckets(self.arena, self.bucket_count);
        for (self.clients.items[named..]) |folded| {
            for (other, folded.buckets) |*slot, count| slot.* += count;
        }
        const series = try self.arena.alloc(ClientSeries, named);
        for (series, self.clients.items[0..named]) |*entry, accum| {
            entry.* = .{ .client = accum.client, .buckets = accum.buckets };
        }

        return .{
            .totals = .{
                .queries = queries,
                .blocked = blocked,
                .distinct_clients = self.clients.items.len,
                .avg_response_time_us = if (self.rt_count == 0)
                    null
                else
                    @divTrunc(self.rt_sum, self.rt_count),
            },
            .buckets = self.buckets,
            .clients = .{ .clients = series, .other = other },
            .types = self.types.items,
            .routes = self.routes.items,
        };
    }
};

/// Absent and present are different sources, and so are two different present
/// ones: an `upstream` row the log recorded no resolver for is its own group,
/// exactly as `GROUP BY ... source` makes it.
fn sameSource(a: ?[]const u8, b: ?[]const u8) bool {
    const left = a orelse return b == null;
    const right = b orelse return false;
    return std.mem.eql(u8, left, right);
}

/// The oracle's `ORDER BY count(*) DESC, qtype IS NULL, qtype ASC`.
fn typeBefore(_: void, a: TypeCount, b: TypeCount) bool {
    if (a.count != b.count) return a.count > b.count;
    const a_type = a.qtype orelse return false;
    const b_type = b.qtype orelse return true;
    return a_type < b_type;
}

/// The oracle's `ORDER BY count(*) DESC, route_kind ASC, source IS NULL,
/// source ASC`. `route_kind` is compared as the stored text, not as the
/// enum's declaration order — the two disagree, and the stored text is what the
/// SQL sorted.
fn routeBefore(_: void, a: RouteCount, b: RouteCount) bool {
    if (a.count != b.count) return a.count > b.count;
    switch (std.mem.order(u8, @tagName(a.route), @tagName(b.route))) {
        .lt => return true,
        .gt => return false,
        .eq => {},
    }
    const a_source = a.source orelse return false;
    const b_source = b.source orelse return true;
    return std.mem.order(u8, a_source, b_source) == .lt;
}

/// The oracle's `ORDER BY count(*) DESC, client_ip ASC`, which is what makes
/// the cut at `max_client_series` the same cut on every read.
fn clientBefore(_: void, a: ClientAccum, b: ClientAccum) bool {
    if (a.total != b.total) return a.total > b.total;
    return std.mem.order(u8, a.client, b.client) == .lt;
}

const overview_projection_totals_sql =
    \\SELECT bucket, queries, blocked, cached, rt_sum, rt_count
    \\  FROM bucket_totals
    \\ WHERE bucket >= ?1 AND bucket < ?2
;

const overview_projection_clients_sql =
    \\SELECT bucket, client_ip, queries
    \\  FROM bucket_clients
    \\ WHERE bucket >= ?1 AND bucket < ?2
;

const overview_projection_types_sql =
    \\SELECT qtype, sum(count)
    \\  FROM bucket_types
    \\ WHERE bucket >= ?1 AND bucket < ?2
    \\ GROUP BY qtype
;

const overview_projection_routes_sql =
    \\SELECT route_kind, source_present, source_text, sum(count)
    \\  FROM bucket_routes
    \\ WHERE bucket >= ?1 AND bucket < ?2
    \\ GROUP BY route_kind, source_present, source_text
;

const overview_raw_sql =
    \\SELECT timestamp, client_ip, qtype, blocked, cache_hit, response_time_us,
    \\       route_kind,
    \\       CASE route_kind
    \\            WHEN 'upstream' THEN upstream
    \\            WHEN 'forward_zone' THEN forward_zone
    \\       END AS source
    \\  FROM query_log
    \\ WHERE timestamp >= ?1 AND timestamp < ?2
;

/// The whole Overview payload for `[since, since + bucket_seconds *
/// bucket_count)`, from one already-open read transaction — the caller owns the
/// transaction and the lock, as it does for every other read here.
///
/// Two implementations behind one entry point. At or above the projection grain
/// the four `bucket_*` tables answer the window, and the cost stops depending
/// on how many raw rows it holds; below it (the 1h window, on 60-second
/// buckets) one pass over the raw rows in the window does the same work in Zig.
/// Both produce the same contracts, which is what the equivalence test asserts.
///
/// Preconditions are caller bugs, not runtime conditions: a zero width or count
/// is `error.Misuse`, and so is a projection-path window that is not on the
/// grid, because the projections cannot express it. The handler's `window()`
/// guarantees all of them.
pub fn overview(
    database: *db.Db,
    arena: Allocator,
    since: i64,
    bucket_seconds: u32,
    bucket_count: u32,
) db.Error!Overview {
    if (bucket_seconds == 0 or bucket_count == 0) return error.Misuse;
    const width: i64 = bucket_seconds;
    const span = std.math.mul(i64, width, bucket_count) catch return error.Misuse;
    const until = std.math.add(i64, since, span) catch return error.Misuse;
    const from_projections = width >= grain;
    if (from_projections and (@mod(since, grain) != 0 or @rem(width, grain) != 0)) return error.Misuse;

    var accum = try Accum.init(arena, since, width, bucket_count);
    if (from_projections) {
        try readProjections(database, &accum, since, until);
    } else {
        try readRaw(database, &accum, since, until);
    }
    return try accum.finish();
}

fn readProjections(database: *db.Db, accum: *Accum, since: i64, until: i64) db.Error!void {
    {
        var stmt = try database.prepare(overview_projection_totals_sql);
        defer stmt.deinit();
        try stmt.bindInt(1, since);
        try stmt.bindInt(2, until);
        while (try stmt.step()) {
            const index = try accum.indexOf(stmt.columnInt(0));
            accum.addTotals(
                index,
                try countOf(stmt.columnInt(1)),
                try countOf(stmt.columnInt(2)),
                try countOf(stmt.columnInt(3)),
            );
            accum.rt_sum += stmt.columnInt(4);
            accum.rt_count += stmt.columnInt(5);
        }
    }
    {
        var stmt = try database.prepare(overview_projection_clients_sql);
        defer stmt.deinit();
        try stmt.bindInt(1, since);
        try stmt.bindInt(2, until);
        while (try stmt.step()) {
            const index = try accum.indexOf(stmt.columnInt(0));
            try accum.addClient(stmt.columnText(1), index, try countOf(stmt.columnInt(2)));
        }
    }
    {
        var stmt = try database.prepare(overview_projection_types_sql);
        defer stmt.deinit();
        try stmt.bindInt(1, since);
        try stmt.bindInt(2, until);
        while (try stmt.step()) {
            try accum.addType(stmt.columnInt(0), try countOf(stmt.columnInt(1)));
        }
    }
    {
        var stmt = try database.prepare(overview_projection_routes_sql);
        defer stmt.deinit();
        try stmt.bindInt(1, since);
        try stmt.bindInt(2, until);
        while (try stmt.step()) {
            const route = try provenance.parse(provenance.RouteKind, stmt.columnText(0));
            const source: ?[]const u8 = if (stmt.columnInt(1) == 0) null else stmt.columnText(2);
            try accum.addRoute(route, source, try countOf(stmt.columnInt(3)));
        }
    }
}

fn readRaw(database: *db.Db, accum: *Accum, since: i64, until: i64) db.Error!void {
    var stmt = try database.prepare(overview_raw_sql);
    defer stmt.deinit();
    try stmt.bindInt(1, since);
    try stmt.bindInt(2, until);

    while (try stmt.step()) {
        const index = try accum.indexOf(stmt.columnInt(0));
        const blocked: u64 = @intFromBool(stmt.columnBool(3));
        // `cache_hit = 1`, not "non-zero": the SQL aggregates this replaces
        // compare against the literal, and a NULL compares to neither.
        const cached: u64 = @intFromBool(!stmt.isNull(4) and stmt.columnInt(4) == 1);
        accum.addTotals(index, 1, blocked, cached);
        if (!stmt.isNull(5)) {
            accum.rt_sum += stmt.columnInt(5);
            accum.rt_count += 1;
        }
        try accum.addClient(stmt.columnText(1), index, 1);
        try accum.addType(if (stmt.isNull(2)) null_qtype else stmt.columnInt(2), 1);
        try accum.addRoute(
            try provenance.parse(provenance.RouteKind, stmt.columnText(6)),
            stmt.columnTextOrNull(7),
            1,
        );
    }
}

// ---------------------------------------------------------------------------
// tests
// ---------------------------------------------------------------------------

const logger = @import("../logger.zig");
const querylog_schema = @import("../querylog_schema.zig");

const testing = std.testing;

fn openLog() !db.Db {
    var database = try db.Db.open(":memory:", .{ .mode = .memory });
    errdefer database.close();
    try db.applyPragmas(&database, .{});
    try database.exec(querylog_schema.ddl);
    return database;
}

fn plainRow(timestamp: i64, domain: []const u8) Row {
    return .{
        .timestamp = timestamp,
        .domain = domain,
        .client_ip = "192.0.2.10",
        .qtype = 1,
        .qclass = 1,
        .rcode = 0,
        .blocked = false,
        .response_time_us = 1200,
        .cache_hit = false,
        .upstream = "9.9.9.9",
        .group_id = 1,
        .group_name = "default",
        .policy_action = .allow,
        .policy_reason = .no_match,
        .matched = null,
        .source_id = null,
        .source_name = null,
        .cname_target = null,
        .safe_search_target = null,
        .route_kind = .upstream,
        .forward_zone = null,
    };
}

fn domainIdOf(database: *db.Db, domain: []const u8) !i64 {
    var stmt = try database.prepare(select_domain_sql);
    defer stmt.deinit();
    try stmt.bindText(1, domain);
    try testing.expect(try stmt.step());
    return stmt.columnInt(0);
}

test "a foreign row with an rcode wider than twelve bits is refused, not truncated" {
    var database = try db.Db.open(":memory:", .{ .mode = .memory });
    defer database.close();
    try db.applyPragmas(&database, .{});

    // The shipped schema's `CHECK` makes this row impossible in a file this
    // build created, so the table is built without it. The read path's job is
    // to refuse a `querylog.db` that came from somewhere else rather than to
    // narrow a value it cannot represent.
    try database.exec(
        \\CREATE TABLE domains (id INTEGER PRIMARY KEY, domain TEXT NOT NULL UNIQUE);
        \\CREATE TABLE query_log (
        \\    id INTEGER PRIMARY KEY, timestamp INTEGER NOT NULL,
        \\    domain_id INTEGER NOT NULL, client_ip TEXT NOT NULL,
        \\    qtype INTEGER, blocked INTEGER NOT NULL, response_time_us INTEGER,
        \\    cache_hit INTEGER, upstream TEXT, qclass INTEGER NOT NULL,
        \\    rcode INTEGER NOT NULL, group_id INTEGER, group_name TEXT,
        \\    policy_action TEXT NOT NULL, policy_reason TEXT NOT NULL,
        \\    matched TEXT, source_id INTEGER, source_name TEXT,
        \\    cname_target TEXT, safe_search_target TEXT,
        \\    route_kind TEXT NOT NULL, forward_zone TEXT
        \\);
        \\INSERT INTO domains (id, domain) VALUES (1, 'a.example');
        \\INSERT INTO query_log
        \\  (id, timestamp, domain_id, client_ip, blocked, qclass, rcode,
        \\   policy_action, policy_reason, route_kind)
        \\VALUES (1, 10, 1, '192.0.2.10', 0, 1, 4096, 'allow', 'no_match', 'upstream');
    );

    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try testing.expectError(error.Mismatch, selectQueries(&database, arena, .{}));
    try testing.expectError(error.Mismatch, detailById(&database, arena, 1));
}

test "writeBatch inserts every row and interns each domain once" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{
        plainRow(100, "example.com"),
        plainRow(101, "example.com"),
        plainRow(102, "ads.example.net"),
    });

    try testing.expectEqual(@as(i64, 3), try countRows(&database));
    try testing.expectEqual(@as(i64, 2), try countDomains(&database));

    const first = try domainIdOf(&database, "example.com");
    try testing.expectEqual(
        @as(i64, 2),
        try database.queryInt("SELECT count(*) FROM query_log WHERE domain_id = 1"),
    );
    try testing.expectEqual(@as(i64, 1), first);
}

test "a second batch reuses the interned domain id" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{plainRow(100, "example.com")});
    const before = try domainIdOf(&database, "example.com");

    try writer.writeBatch(&.{ plainRow(200, "example.com"), plainRow(201, "other.example") });
    const after = try domainIdOf(&database, "example.com");

    try testing.expectEqual(before, after);
    try testing.expectEqual(@as(i64, 3), try countRows(&database));
    try testing.expectEqual(@as(i64, 2), try countDomains(&database));
    try testing.expectEqual(
        @as(i64, 2),
        try database.queryInt("SELECT count(*) FROM query_log WHERE domain_id = 1"),
    );
}

test "every column round-trips a value and a null" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{
        .{
            .timestamp = 10,
            .domain = "blocked.example",
            .client_ip = "2001:db8::1",
            .qtype = 28,
            .qclass = 1,
            .rcode = 3,
            .blocked = true,
            .response_time_us = 42,
            .cache_hit = true,
            .upstream = "https://dns.example/dns-query",
            .group_id = 7,
            .group_name = "kids",
            .policy_action = .block,
            .policy_reason = .blocklist_wildcard,
            .matched = "*.ads.example",
            .source_id = 3,
            .source_name = "steven black",
            .cname_target = "tracker.cdn.example",
            .safe_search_target = "forcesafesearch.google.com",
            .route_kind = .blocked,
            .forward_zone = "home.arpa",
        },
        // Every nullable column absent at once, which is the shape of a query
        // the pipeline answered before any of them applied.
        .{
            .timestamp = 11,
            .domain = "quiet.example",
            .client_ip = "hidden",
            .qtype = null,
            .qclass = 3,
            .rcode = 0,
            .blocked = false,
            .response_time_us = null,
            .cache_hit = null,
            .upstream = null,
            .group_id = null,
            .group_name = null,
            .policy_action = .not_evaluated,
            .policy_reason = .non_in_class,
            .matched = null,
            .source_id = null,
            .source_name = null,
            .cname_target = null,
            .safe_search_target = null,
            .route_kind = .upstream,
            .forward_zone = null,
        },
    });

    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    const full = (try detailById(&database, arena, 1)).?;
    try testing.expectEqualStrings("blocked.example", full.domain);
    try testing.expectEqualStrings("2001:db8::1", full.client_ip);
    try testing.expectEqual(@as(?u16, 28), full.qtype);
    try testing.expectEqual(@as(u16, 1), full.qclass);
    try testing.expectEqual(@as(u12, 3), full.rcode);
    try testing.expect(full.blocked);
    try testing.expectEqual(@as(?i64, 42), full.response_time_us);
    try testing.expectEqual(@as(?bool, true), full.cache_hit);
    try testing.expectEqualStrings("https://dns.example/dns-query", full.upstream);
    try testing.expectEqual(@as(?i64, 7), full.group_id);
    try testing.expectEqualStrings("kids", full.group_name);
    try testing.expectEqual(provenance.PolicyAction.block, full.policy_action);
    try testing.expectEqual(provenance.PolicyReason.blocklist_wildcard, full.policy_reason);
    try testing.expectEqualStrings("*.ads.example", full.matched);
    try testing.expectEqual(@as(?i64, 3), full.source_id);
    try testing.expectEqualStrings("steven black", full.source_name);
    try testing.expectEqualStrings("tracker.cdn.example", full.cname_target);
    try testing.expectEqualStrings("forcesafesearch.google.com", full.safe_search_target);
    try testing.expectEqual(provenance.RouteKind.blocked, full.route_kind);
    try testing.expectEqualStrings("home.arpa", full.forward_zone);

    // A NULL text column reads as the empty string, by the documented
    // convention; a NULL integer stays null, because 0 is a real id.
    const bare = (try detailById(&database, arena, 2)).?;
    try testing.expectEqual(@as(?u16, null), bare.qtype);
    try testing.expectEqual(@as(u16, 3), bare.qclass);
    try testing.expectEqual(@as(?i64, null), bare.response_time_us);
    try testing.expectEqual(@as(?bool, null), bare.cache_hit);
    try testing.expectEqualStrings("", bare.upstream);
    try testing.expectEqual(@as(?i64, null), bare.group_id);
    try testing.expectEqualStrings("", bare.group_name);
    try testing.expectEqual(provenance.PolicyAction.not_evaluated, bare.policy_action);
    try testing.expectEqual(provenance.PolicyReason.non_in_class, bare.policy_reason);
    try testing.expectEqualStrings("", bare.matched);
    try testing.expectEqual(@as(?i64, null), bare.source_id);
    try testing.expectEqualStrings("", bare.source_name);
    try testing.expectEqualStrings("", bare.cname_target);
    try testing.expectEqualStrings("", bare.safe_search_target);
    try testing.expectEqual(provenance.RouteKind.upstream, bare.route_kind);
    try testing.expectEqualStrings("", bare.forward_zone);
}

test "detailById returns null for an id no row has" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    try seed(&database, &.{plainRow(10, "a.example")});

    // An id from before a retention pass looks exactly like this, and is a 404
    // rather than an error.
    try testing.expectEqual(@as(?QueryDetail, null), try detailById(&database, arena_state.allocator(), 2));
    try testing.expectEqual(@as(?QueryDetail, null), try detailById(&database, arena_state.allocator(), 0));
    try testing.expect((try detailById(&database, arena_state.allocator(), 1)) != null);
}

test "a stored enum value the schema does not define is a data error, not a passthrough" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try seed(&database, &.{plainRow(10, "a.example")});
    try database.exec("UPDATE query_log SET policy_reason = 'whatever' WHERE id = 1;");

    try testing.expectError(error.Mismatch, detailById(&database, arena, 1));
    try testing.expectError(error.Mismatch, selectQueries(&database, arena, .{}));
}

test "an empty batch writes nothing and opens no transaction" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    // A transaction is already open, so a `BEGIN IMMEDIATE` from `writeBatch`
    // would fail: this is what proves the empty batch returns before it.
    var tx = try db.Tx.begin(&database);
    try writer.writeBatch(&.{});
    tx.rollback();

    try testing.expectEqual(@as(i64, 0), try countRows(&database));
    try testing.expectEqual(@as(i64, 0), try countDomains(&database));
}

test "pruneOlderThan deletes strictly older rows and returns the count" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{
        plainRow(100, "old.example"),
        plainRow(199, "old.example"),
        plainRow(200, "edge.example"),
        plainRow(300, "fresh.example"),
    });

    try testing.expectEqual(@as(i64, 2), (try pruneOlderThan(&database, 200)).deleted);
    try testing.expectEqual(@as(i64, 2), try countRows(&database));
    // The row exactly at the cutoff stays.
    try testing.expectEqual(
        @as(i64, 1),
        try database.queryInt("SELECT count(*) FROM query_log WHERE timestamp = 200"),
    );
    // A second pass over the same cutoff finds nothing left to do.
    try testing.expectEqual(@as(i64, 0), (try pruneOlderThan(&database, 200)).deleted);
}

test "a prune advances the coverage watermark to its own cutoff" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    // The seeded watermark is `created_at + 1`, which is now-ish; the cutoffs
    // below are all in the past, so they start out behind it.
    const start = try availableSince(&database);
    try writer.writeBatch(&.{ plainRow(start + 100, "a.example"), plainRow(start + 300, "b.example") });

    const first = try pruneOlderThan(&database, start + 200);
    try testing.expectEqual(@as(i64, 1), first.deleted);
    try testing.expectEqual(start + 200, first.available_since);
    try testing.expectEqual(start + 200, try availableSince(&database));

    // A prune that deletes nothing still advances: the window it swept is
    // covered whether or not it held rows.
    const second = try pruneOlderThan(&database, start + 250);
    try testing.expectEqual(@as(i64, 0), second.deleted);
    try testing.expectEqual(start + 250, try availableSince(&database));
}

test "the watermark never moves backward" {
    var database = try openLog();
    defer database.close();

    const start = try availableSince(&database);
    const advanced = try pruneOlderThan(&database, start + 1000);
    try testing.expectEqual(start + 1000, advanced.available_since);

    // A shortened `retention_days`, a clock that stepped back, a pass with a
    // stale cutoff: none of them may widen the promise the file makes.
    for ([_]i64{ start + 999, start, start - 100_000, 0 }) |older| {
        const result = try pruneOlderThan(&database, older);
        try testing.expectEqual(start + 1000, result.available_since);
        try testing.expectEqual(start + 1000, try availableSince(&database));
    }
}

test "a failed delete leaves both the rows and the watermark untouched" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const start = try availableSince(&database);
    try writer.writeBatch(&.{plainRow(start - 100, "old.example")});
    try database.exec(
        \\CREATE TRIGGER refuse_delete BEFORE DELETE ON query_log
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    try testing.expectError(error.Constraint, pruneOlderThan(&database, start + 1000));

    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(start, try availableSince(&database));
}

test "a failed watermark update leaves the rows it had already deleted" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const start = try availableSince(&database);
    try writer.writeBatch(&.{plainRow(start - 100, "old.example")});
    // The delete succeeds and the advance does not. Without one transaction
    // around the pair, this is the case that loses rows the watermark still
    // promises.
    try database.exec(
        \\CREATE TRIGGER refuse_advance BEFORE UPDATE ON querylog_meta
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    try testing.expectError(error.Constraint, pruneOlderThan(&database, start + 1000));

    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(start, try availableSince(&database));
}

test "a failed commit rolls back the delete and the watermark together" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const start = try availableSince(&database);
    try writer.writeBatch(&.{plainRow(start - 100, "old.example")});

    // Both statements succeed and COMMIT is what fails: the advance inserts a
    // `query_log` row whose `domain_id` references nothing, and
    // `defer_foreign_keys` holds that violation back until the commit checks
    // it (SQLite's documented semantics for the pragma; the assertions below
    // observe the rollback, not the moment the check ran).
    try database.exec(
        \\CREATE TRIGGER break_at_commit AFTER UPDATE ON querylog_meta
        \\BEGIN INSERT INTO query_log
        \\  (timestamp, domain_id, client_ip, blocked, qclass, rcode,
        \\   policy_action, policy_reason, route_kind)
        \\VALUES (1, 999999, 'x', 0, 1, 0, 'allow', 'no_match', 'upstream'); END;
    );
    try database.exec("PRAGMA defer_foreign_keys = ON;");

    try testing.expectError(error.Constraint, pruneOlderThan(&database, start + 1000));

    // Nothing survived: not the delete, not the advance, not the row the
    // trigger inserted.
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(start, try availableSince(&database));
    try testing.expectEqual(
        @as(i64, 0),
        try database.queryInt("SELECT count(*) FROM query_log WHERE client_ip = 'x'"),
    );
}

test "pruneOlderThan leaves the domains dimension table intact" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{ plainRow(10, "a.example"), plainRow(11, "b.example") });
    try testing.expectEqual(@as(i64, 2), (try pruneOlderThan(&database, 1000)).deleted);

    try testing.expectEqual(@as(i64, 0), try countRows(&database));
    try testing.expectEqual(@as(i64, 2), try countDomains(&database));
}

test "a failing row rolls the whole batch back and the writer survives it" {
    var database = try openLog();
    defer database.close();
    try database.exec(
        \\CREATE TRIGGER refuse_boom BEFORE INSERT ON query_log
        \\WHEN new.client_ip = 'boom'
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    var doomed = plainRow(20, "second.example");
    doomed.client_ip = "boom";
    try testing.expectError(error.Constraint, writer.writeBatch(&.{
        plainRow(10, "first.example"),
        doomed,
    }));

    // The interned domain of the row that did insert is gone with it.
    try testing.expectEqual(@as(i64, 0), try countRows(&database));
    try testing.expectEqual(@as(i64, 0), try countDomains(&database));

    try writer.writeBatch(&.{plainRow(30, "third.example")});
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(@as(i64, 1), try countDomains(&database));
}

test "countRows and countDomains agree with what the batches wrote" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try testing.expectEqual(@as(i64, 0), try countRows(&database));
    try testing.expectEqual(@as(i64, 0), try countDomains(&database));

    var rows: [50]Row = undefined;
    var names: [50][16]u8 = undefined;
    for (&rows, &names, 0..) |*row, *name, i| {
        const written = std.fmt.bufPrint(name, "d{d}.example", .{i % 7}) catch unreachable;
        row.* = plainRow(@intCast(i), written);
    }
    try writer.writeBatch(&rows);

    try testing.expectEqual(@as(i64, 50), try countRows(&database));
    try testing.expectEqual(@as(i64, 7), try countDomains(&database));
}

// `PRAGMA wal_checkpoint` needs a real WAL, which an in-memory database cannot
// have. `std.testing.tmpDir` creates its directory under `.zig-cache/tmp/`
// relative to the process working directory, which is also how SQLite's VFS
// resolves the filename it is handed (`storage_integration_test.zig:44`).
const tmp_prefix = ".zig-cache/tmp/";
const sub_path_len = @typeInfo(@FieldType(testing.TmpDir, "sub_path")).array.len;

test "checkpointTruncate and vacuum run against a WAL file database" {
    var tmp = testing.tmpDir(.{});
    defer tmp.cleanup();

    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
    const path = try std.fmt.bufPrintZ(&path_buf, "{s}{s}/querylog.db", .{ tmp_prefix, &tmp.sub_path });

    var database = try db.Db.open(path, .{ .mode = .read_write_create });
    defer database.close();
    try db.applyPragmas(&database, .{});
    {
        var stmt = try database.prepare("PRAGMA journal_mode");
        defer stmt.deinit();
        try testing.expect(try stmt.step());
        // `columnText` is borrowed until the next call on the statement, so it
        // is compared here rather than carried out of this block.
        try testing.expectEqualStrings("wal", stmt.columnText(0));
    }
    try database.exec(querylog_schema.ddl);

    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();
    try writer.writeBatch(&.{ plainRow(10, "a.example"), plainRow(20, "b.example") });

    try checkpointTruncate(&database);
    try testing.expectEqual(@as(i64, 1), (try pruneOlderThan(&database, 20)).deleted);
    try checkpointTruncate(&database);
    try vacuum(&database);

    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(@as(i64, 2), try countDomains(&database));
}

// --- the read layer -------------------------------------------------------

/// `BatchWriter` assigns `query_log.id` in the order it is handed the rows, so
/// every test below knows the id of each seeded row: the nth row of the nth
/// batch has id n.
fn seed(database: *db.Db, rows: []const Row) !void {
    var writer = try BatchWriter.init(testing.allocator, database);
    defer writer.deinit();
    try writer.writeBatch(rows);
}

fn ids(rows: []const QueryRow, out: []i64) []const i64 {
    for (rows, 0..) |row, i| out[i] = row.id;
    return out[0..rows.len];
}

test "selectQueries returns the newest row first and reads every column" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    try seed(&database, &.{
        .{
            .timestamp = 10,
            .domain = "ads.example.net",
            .client_ip = "192.0.2.10",
            .qtype = 28,
            .qclass = 1,
            .rcode = 0,
            .blocked = true,
            .response_time_us = 4200,
            .cache_hit = true,
            .upstream = "https://dns.example/dns-query",
            .group_id = 2,
            .group_name = "kids",
            .policy_action = .block,
            .policy_reason = .blocklist_domain,
            .matched = "ads.example.net",
            .source_id = 5,
            .source_name = "steven black",
            .cname_target = null,
            .safe_search_target = null,
            .route_kind = .blocked,
            .forward_zone = null,
        },
        .{
            .timestamp = 20,
            .domain = "quiet.example",
            .client_ip = "hidden",
            .qtype = null,
            .qclass = 1,
            .rcode = 2,
            .blocked = false,
            .response_time_us = null,
            .cache_hit = null,
            .upstream = null,
            .group_id = null,
            .group_name = null,
            .policy_action = .not_evaluated,
            .policy_reason = .snapshot_unavailable,
            .matched = null,
            .source_id = null,
            .source_name = null,
            .cname_target = null,
            .safe_search_target = null,
            .route_kind = .rejected,
            .forward_zone = null,
        },
    });

    const rows = try selectQueries(&database, arena_state.allocator(), .{});
    try testing.expectEqual(@as(usize, 2), rows.items.len);

    const newest = rows.items[0];
    try testing.expectEqual(@as(i64, 2), newest.id);
    try testing.expectEqual(@as(i64, 20), newest.ts);
    try testing.expectEqualStrings("quiet.example", newest.domain);
    try testing.expectEqualStrings("hidden", newest.client_ip);
    try testing.expectEqual(@as(?u16, null), newest.qtype);
    try testing.expectEqual(@as(u16, 1), newest.qclass);
    try testing.expectEqual(@as(u12, 2), newest.rcode);
    try testing.expect(!newest.blocked);
    try testing.expectEqual(@as(?i64, null), newest.response_time_us);
    try testing.expectEqual(@as(?bool, null), newest.cache_hit);
    // A NULL text column reads as the empty string, by documented convention.
    try testing.expectEqualStrings("", newest.upstream);
    try testing.expectEqual(provenance.PolicyAction.not_evaluated, newest.policy_action);
    try testing.expectEqual(provenance.PolicyReason.snapshot_unavailable, newest.policy_reason);
    try testing.expectEqual(provenance.RouteKind.rejected, newest.route_kind);

    const oldest = rows.items[1];
    try testing.expectEqual(@as(i64, 1), oldest.id);
    try testing.expectEqual(@as(i64, 10), oldest.ts);
    try testing.expectEqualStrings("ads.example.net", oldest.domain);
    try testing.expectEqualStrings("192.0.2.10", oldest.client_ip);
    try testing.expectEqual(@as(?u16, 28), oldest.qtype);
    try testing.expectEqual(@as(u16, 1), oldest.qclass);
    try testing.expectEqual(@as(u12, 0), oldest.rcode);
    try testing.expect(oldest.blocked);
    try testing.expectEqual(@as(?i64, 4200), oldest.response_time_us);
    try testing.expectEqual(@as(?bool, true), oldest.cache_hit);
    try testing.expectEqualStrings("https://dns.example/dns-query", oldest.upstream);
    try testing.expectEqual(provenance.PolicyAction.block, oldest.policy_action);
    try testing.expectEqual(provenance.PolicyReason.blocklist_domain, oldest.policy_reason);
    try testing.expectEqual(provenance.RouteKind.blocked, oldest.route_kind);
}

test "selectQueries honours the limit and caps it at max_limit" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    var rows: [1005]Row = undefined;
    for (&rows, 0..) |*row, i| row.* = plainRow(@intCast(i), "example.com");
    try seed(&database, &rows);

    const few = try selectQueries(&database, arena, .{ .limit = 3 });
    try testing.expectEqual(@as(usize, 3), few.items.len);

    // Asked for more than the cap, and for more rows than the cap, so the cap
    // is what bounds the answer rather than the table.
    const capped = try selectQueries(&database, arena, .{ .limit = 5000 });
    try testing.expectEqual(@as(usize, max_limit), capped.items.len);

    const none = try selectQueries(&database, arena, .{ .limit = 0 });
    try testing.expectEqual(@as(usize, 0), none.items.len);
}

test "keyset paging walks every row exactly once across the page boundaries" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    var seeded: [7]Row = undefined;
    for (&seeded, 0..) |*row, i| row.* = plainRow(@intCast(i), "example.com");
    try seed(&database, &seeded);

    var seen: std.ArrayList(i64) = .empty;
    defer seen.deinit(testing.allocator);

    var before: ?i64 = null;
    var pages: usize = 0;
    while (pages < 10) : (pages += 1) {
        const page = try selectQueries(&database, arena, .{ .limit = 3, .before = before });
        if (page.items.len == 0) break;
        for (page.items) |row| try seen.append(testing.allocator, row.id);
        before = page.items[page.items.len - 1].id;
    }

    // Two full pages and one short page; the fourth call returns nothing and
    // breaks before the counter, which is how the walk knows it is done.
    try testing.expectEqual(@as(usize, 3), pages);
    try testing.expectEqualSlices(i64, &.{ 7, 6, 5, 4, 3, 2, 1 }, seen.items);
}

test "each filter narrows the result on its own" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    var blocked_row = plainRow(200, "ads.example.net");
    blocked_row.client_ip = "192.0.2.20";
    blocked_row.blocked = true;
    blocked_row.policy_action = .block;
    blocked_row.policy_reason = .blocklist_domain;
    blocked_row.route_kind = .blocked;
    try seed(&database, &.{
        plainRow(100, "one.example.com"),
        blocked_row,
        plainRow(300, "two.example.com"),
    });

    var buf: [8]i64 = undefined;

    const by_domain = try selectQueries(&database, arena, .{ .domain_substring = "example.com" });
    try testing.expectEqualSlices(i64, &.{ 3, 1 }, ids(by_domain.items, &buf));

    const by_client = try selectQueries(&database, arena, .{ .client = "192.0.2.20" });
    try testing.expectEqualSlices(i64, &.{2}, ids(by_client.items, &buf));

    // An exact match, not a prefix: the seeded clients share the first octets.
    const no_client = try selectQueries(&database, arena, .{ .client = "192.0.2" });
    try testing.expectEqual(@as(usize, 0), no_client.items.len);

    const only_blocked = try selectQueries(&database, arena, .{ .blocked = true });
    try testing.expectEqualSlices(i64, &.{2}, ids(only_blocked.items, &buf));

    const only_allowed = try selectQueries(&database, arena, .{ .blocked = false });
    try testing.expectEqualSlices(i64, &.{ 3, 1 }, ids(only_allowed.items, &buf));

    // Every filter at once, all satisfied by the one blocked row.
    const combined = try selectQueries(&database, arena, .{
        .limit = 10,
        .before = 3,
        .domain_substring = "ads",
        .client = "192.0.2.20",
        .blocked = true,
        .since = 200,
        .until = 300,
    });
    try testing.expectEqualSlices(i64, &.{2}, ids(combined.items, &buf));
}

test "since is inclusive, until is exclusive, and an empty range selects nothing" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try seed(&database, &.{
        plainRow(100, "a.example"),
        plainRow(200, "b.example"),
        plainRow(300, "c.example"),
    });

    var buf: [8]i64 = undefined;

    const window = try selectQueries(&database, arena, .{ .since = 100, .until = 300 });
    try testing.expectEqualSlices(i64, &.{ 2, 1 }, ids(window.items, &buf));

    const after = try selectQueries(&database, arena, .{ .since = 300 });
    try testing.expectEqualSlices(i64, &.{3}, ids(after.items, &buf));

    const empty = try selectQueries(&database, arena, .{ .since = 300, .until = 300 });
    try testing.expectEqual(@as(usize, 0), empty.items.len);

    const beyond = try selectQueries(&database, arena, .{ .since = 1000 });
    try testing.expectEqual(@as(usize, 0), beyond.items.len);
}

test "a domain substring matches % and _ as literal characters" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try seed(&database, &.{
        plainRow(10, "a_b.example"),
        plainRow(20, "axb.example"),
        plainRow(30, "a%b.example"),
        plainRow(40, "azzb.example"),
        plainRow(50, "back\\slash.example"),
    });

    var buf: [8]i64 = undefined;

    // Unescaped, `_` is LIKE's single-character wildcard and would also match
    // "axb"; escaped, it matches only the underscore.
    const underscore = try selectQueries(&database, arena, .{ .domain_substring = "a_b" });
    try testing.expectEqualSlices(i64, &.{1}, ids(underscore.items, &buf));

    // Unescaped, `%` would match everything from "a" to "b", so "azzb" too.
    const percent = try selectQueries(&database, arena, .{ .domain_substring = "a%b" });
    try testing.expectEqualSlices(i64, &.{3}, ids(percent.items, &buf));

    // The escape character escapes itself, so it is searchable as well.
    const backslash = try selectQueries(&database, arena, .{ .domain_substring = "k\\s" });
    try testing.expectEqualSlices(i64, &.{5}, ids(backslash.items, &buf));

    // An empty needle is `%%`, which matches every row rather than none.
    const all = try selectQueries(&database, arena, .{ .domain_substring = "" });
    try testing.expectEqual(@as(usize, 5), all.items.len);
}

test "the built SQL never carries a filter value and fits its buffer" {
    var sql: Sql = .{};
    sql.put(select_head);
    sql.predicate(where_before);
    sql.predicate(where_domain);
    sql.predicate(where_client);
    sql.predicate(where_blocked);
    sql.predicate(where_since);
    sql.predicate(where_until);
    sql.put(select_tail);

    // Every predicate present is the longest reachable statement.
    try testing.expect(sql.len <= Sql.capacity);
    try testing.expectEqual(@as(usize, 1), std.mem.count(u8, sql.text(), " WHERE"));
    try testing.expectEqual(@as(usize, 5), std.mem.count(u8, sql.text(), " AND"));
    // Six filters plus the LIMIT, each a bare parameter.
    try testing.expectEqual(@as(usize, 7), std.mem.count(u8, sql.text(), "?"));
}

test "likePattern wraps the needle and neutralises every metacharacter" {
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try testing.expectEqualStrings("%plain%", try likePattern(arena, "plain"));
    try testing.expectEqualStrings("%a\\_b%", try likePattern(arena, "a_b"));
    try testing.expectEqualStrings("%a\\%b%", try likePattern(arena, "a%b"));
    try testing.expectEqualStrings("%a\\\\b%", try likePattern(arena, "a\\b"));
    try testing.expectEqualStrings("%%", try likePattern(arena, ""));
}

// ---------------------------------------------------------------------------
// period aggregations (milestone 30)
// ---------------------------------------------------------------------------

const agg_since: i64 = 1_700_000_000;
const agg_width: u32 = 60;
const agg_buckets: u32 = 10;

/// One row of the aggregation fixtures. Everything the three breakdowns read
/// is a parameter; everything else is the same on every row, so a test that
/// changes an outcome names the reason it changed.
fn aggRow(offset: i64, client: []const u8, qtype: ?u16, kind: provenance.RouteKind, source: ?[]const u8) Row {
    return .{
        .timestamp = agg_since + offset,
        .domain = "example.com",
        .client_ip = client,
        .qtype = qtype,
        .qclass = 1,
        .rcode = 0,
        .blocked = kind == .blocked,
        .response_time_us = 1000,
        .cache_hit = kind == .cache,
        .upstream = if (kind == .upstream) source else null,
        .group_id = 1,
        .group_name = "default",
        .policy_action = if (kind == .blocked) .block else .allow,
        .policy_reason = if (kind == .blocked) .blocklist_domain else .no_match,
        .matched = null,
        .source_id = null,
        // Blocklist provenance, deliberately set on every row: the routes
        // breakdown must never group by it.
        .source_name = "StevenBlack",
        .cname_target = null,
        .safe_search_target = null,
        .route_kind = kind,
        .forward_zone = if (kind == .forward_zone) source else null,
    };
}

test "the type breakdown groups by qtype, keeps the null row and orders it last" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try seed(&database, &.{
        aggRow(0, "192.0.2.10", 1, .upstream, "9.9.9.9"),
        aggRow(1, "192.0.2.10", 1, .upstream, "9.9.9.9"),
        aggRow(2, "192.0.2.10", 1, .upstream, "9.9.9.9"),
        aggRow(3, "192.0.2.10", 28, .upstream, "9.9.9.9"),
        aggRow(4, "192.0.2.10", 28, .upstream, "9.9.9.9"),
        // Ties with qtype 28, so the tie-break puts the lower code first.
        aggRow(5, "192.0.2.10", 16, .upstream, "9.9.9.9"),
        aggRow(6, "192.0.2.10", 16, .upstream, "9.9.9.9"),
        // A row with no type at all: its own group, never a dropped row.
        aggRow(7, "192.0.2.10", null, .upstream, "9.9.9.9"),
        aggRow(8, "192.0.2.10", null, .upstream, "9.9.9.9"),
        // Outside the window.
        aggRow(-1, "192.0.2.10", 255, .upstream, "9.9.9.9"),
    });

    const rows = (try overview(&database, arena, agg_since, agg_width, agg_buckets)).types;
    try testing.expectEqual(@as(usize, 4), rows.len);
    try testing.expectEqual(@as(?u16, 1), rows[0].qtype);
    try testing.expectEqual(@as(u64, 3), rows[0].count);
    try testing.expectEqual(@as(?u16, 16), rows[1].qtype);
    try testing.expectEqual(@as(?u16, 28), rows[2].qtype);
    try testing.expectEqual(@as(?u16, null), rows[3].qtype);
    try testing.expectEqual(@as(u64, 2), rows[3].count);
}

test "an empty window has no type rows at all" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    const rows = (try overview(&database, arena_state.allocator(), agg_since, agg_width, agg_buckets)).types;
    try testing.expectEqual(@as(usize, 0), rows.len);
}

test "the route breakdown keys on the answering resolver, not on blocklist provenance" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    try seed(&database, &.{
        aggRow(0, "192.0.2.10", 1, .upstream, "https://a.example/dns-query"),
        aggRow(1, "192.0.2.10", 1, .upstream, "https://a.example/dns-query"),
        aggRow(2, "192.0.2.10", 1, .upstream, "https://b.example/dns-query"),
        // An upstream row whose resolver the log did not record: its own group.
        aggRow(3, "192.0.2.10", 1, .upstream, null),
        aggRow(4, "192.0.2.10", 1, .forward_zone, "lan"),
        aggRow(5, "192.0.2.10", 1, .blocked, null),
        aggRow(6, "192.0.2.10", 1, .cache, null),
        aggRow(7, "192.0.2.10", 1, .local, null),
        aggRow(8, "192.0.2.10", 1, .rejected, null),
    });

    const rows = (try overview(&database, arena, agg_since, agg_width, agg_buckets)).routes;
    // Two upstreams, one null-source upstream, one forward zone and four
    // source-less kinds. Every row carries the same `source_name`, so a
    // breakdown that grouped by it would collapse to one row.
    try testing.expectEqual(@as(usize, 8), rows.len);

    try testing.expectEqual(provenance.RouteKind.upstream, rows[0].route);
    try testing.expectEqualStrings("https://a.example/dns-query", rows[0].source.?);
    try testing.expectEqual(@as(u64, 2), rows[0].count);

    // The seven remaining rows all count 1, so the tie-break orders them:
    // route ascending, then source ascending with nulls last.
    for (rows[1..]) |row| try testing.expectEqual(@as(u64, 1), row.count);
    try testing.expectEqual(provenance.RouteKind.blocked, rows[1].route);
    try testing.expectEqual(@as(?[]const u8, null), rows[1].source);
    try testing.expectEqual(provenance.RouteKind.cache, rows[2].route);
    try testing.expectEqual(provenance.RouteKind.forward_zone, rows[3].route);
    try testing.expectEqualStrings("lan", rows[3].source.?);
    try testing.expectEqual(provenance.RouteKind.local, rows[4].route);
    try testing.expectEqual(provenance.RouteKind.rejected, rows[5].route);
    try testing.expectEqual(provenance.RouteKind.upstream, rows[6].route);
    try testing.expectEqualStrings("https://b.example/dns-query", rows[6].source.?);
    try testing.expectEqual(provenance.RouteKind.upstream, rows[7].route);
    try testing.expectEqual(@as(?[]const u8, null), rows[7].source);
}

test "an empty window has no route rows at all" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    const rows = (try overview(&database, arena_state.allocator(), agg_since, agg_width, agg_buckets)).routes;
    try testing.expectEqual(@as(usize, 0), rows.len);
}

test "an empty window still has a zero-filled other series and no named clients" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    const result = (try overview(&database, arena_state.allocator(), agg_since, agg_width, agg_buckets)).clients;
    try testing.expectEqual(@as(usize, 0), result.clients.len);
    try testing.expectEqual(@as(usize, agg_buckets), result.other.len);
    for (result.other) |count| try testing.expectEqual(@as(u64, 0), count);
}

test "client series are bucket-aligned, zero-filled and ranked by in-window total" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    try seed(&database, &.{
        aggRow(0, "192.0.2.20", 1, .upstream, "9.9.9.9"),
        aggRow(1, "192.0.2.20", 1, .upstream, "9.9.9.9"),
        aggRow(agg_width * 3, "192.0.2.20", 1, .upstream, "9.9.9.9"),
        aggRow(agg_width * 3, "192.0.2.10", 1, .upstream, "9.9.9.9"),
        // Outside the window on both sides.
        aggRow(-1, "192.0.2.20", 1, .upstream, "9.9.9.9"),
        aggRow(agg_width * agg_buckets, "192.0.2.10", 1, .upstream, "9.9.9.9"),
    });

    const result = (try overview(&database, arena_state.allocator(), agg_since, agg_width, agg_buckets)).clients;
    try testing.expectEqual(@as(usize, 2), result.clients.len);
    try testing.expectEqualStrings("192.0.2.20", result.clients[0].client);
    try testing.expectEqualStrings("192.0.2.10", result.clients[1].client);

    for (result.clients) |series| try testing.expectEqual(@as(usize, agg_buckets), series.buckets.len);
    try testing.expectEqual(@as(u64, 2), result.clients[0].buckets[0]);
    try testing.expectEqual(@as(u64, 1), result.clients[0].buckets[3]);
    try testing.expectEqual(@as(u64, 0), result.clients[0].buckets[9]);
    try testing.expectEqual(@as(u64, 1), result.clients[1].buckets[3]);
    for (result.other) |count| try testing.expectEqual(@as(u64, 0), count);
}

test "the ninth client folds into other and the cut is the same on every read" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    // Nine clients, each with one more query than the next, so the ranking is
    // total and the ninth is unambiguously the one that folds.
    var address: [16]u8 = undefined;
    var index: u32 = 0;
    while (index < 9) : (index += 1) {
        const client = try std.fmt.bufPrint(&address, "192.0.2.{d}", .{100 + index});
        var repeat: u32 = 0;
        while (repeat <= index) : (repeat += 1) {
            try seed(&database, &.{aggRow(@intCast(repeat), client, 1, .upstream, "9.9.9.9")});
        }
    }

    const result = (try overview(&database, arena, agg_since, agg_width, agg_buckets)).clients;
    try testing.expectEqual(@as(usize, max_client_series), result.clients.len);
    // The busiest is 192.0.2.108 with nine rows; the lone folded client is
    // 192.0.2.100 with one.
    try testing.expectEqualStrings("192.0.2.108", result.clients[0].client);
    for (result.clients) |series| {
        try testing.expect(!std.mem.eql(u8, "192.0.2.100", series.client));
    }

    var other_total: u64 = 0;
    for (result.other) |count| other_total += count;
    try testing.expectEqual(@as(u64, 1), other_total);
}

test "the three breakdowns conserve the window's total" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    // A matrix that exercises every branch the aggregations partition on: null
    // and non-null qtypes, every route kind, a null resolver identity, and ten
    // clients so the top-eight cut has a residual to carry.
    const kinds = [_]provenance.RouteKind{ .blocked, .local, .forward_zone, .upstream, .cache, .rejected };
    var address: [16]u8 = undefined;
    var index: u32 = 0;
    while (index < 40) : (index += 1) {
        const client = try std.fmt.bufPrint(&address, "192.0.2.{d}", .{100 + index % 10});
        const kind = kinds[index % kinds.len];
        try seed(&database, &.{aggRow(
            @intCast(index % (agg_width * agg_buckets)),
            client,
            if (index % 7 == 0) null else @intCast(1 + index % 3),
            kind,
            if (index % 11 == 0) null else "9.9.9.9",
        )});
    }

    const result = try overview(&database, arena, agg_since, agg_width, agg_buckets);
    try testing.expect(result.totals.queries > 0);

    var typed: u64 = 0;
    for (result.types) |row| typed += row.count;
    try testing.expectEqual(result.totals.queries, typed);

    var routed: u64 = 0;
    for (result.routes) |row| routed += row.count;
    try testing.expectEqual(result.totals.queries, routed);

    // Per bucket, not just in total: a series misaligned by one bucket would
    // still sum correctly over the window.
    for (result.buckets, 0..) |bucket, at| {
        var summed: u64 = result.clients.other[at];
        for (result.clients.clients) |series| summed += series.buckets[at];
        try testing.expectEqual(bucket.queries, summed);
    }
}

test "the aggregations pass a redacted client through as the log stored it" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();

    // `hide_client_ips` is applied by the logger before the row is written, so
    // the read path has nothing to transform — and must not invent anything
    // either. The marker is the client's whole identity here.
    try seed(&database, &.{
        aggRow(0, logger.hidden_marker, 1, .upstream, "9.9.9.9"),
        aggRow(1, logger.hidden_marker, 1, .upstream, "9.9.9.9"),
    });

    const result = (try overview(&database, arena_state.allocator(), agg_since, agg_width, agg_buckets)).clients;
    try testing.expectEqual(@as(usize, 1), result.clients.len);
    try testing.expectEqualStrings(logger.hidden_marker, result.clients[0].client);
    try testing.expectEqual(@as(u64, 2), result.clients[0].buckets[0]);
}

test "a prune committed mid-read is invisible to the reader's transaction" {
    var tmp = testing.tmpDir(.{});
    defer tmp.cleanup();

    var path_buf: [tmp_prefix.len + sub_path_len + 32]u8 = undefined;
    const path = try std.fmt.bufPrintZ(&path_buf, "{s}{s}/querylog.db", .{ tmp_prefix, &tmp.sub_path });

    var reader = try db.Db.open(path, .{ .mode = .read_write_create });
    defer reader.close();
    try db.applyPragmas(&reader, .{});
    try reader.exec(querylog_schema.ddl);
    // The seeded watermark is the file's creation second, which is now; the
    // fixtures below are in 2023, so the prune's cutoff would never advance it.
    try reader.exec(
        \\UPDATE querylog_meta SET created_at = 1700000000, available_since = 1700000000 WHERE id = 1
    );
    try seed(&reader, &.{
        plainRow(agg_since, "a.example"),
        plainRow(agg_since + 1, "b.example"),
        plainRow(agg_since + 300, "c.example"),
    });

    // A second connection to the same file, as retention has in production.
    var pruner = try db.Db.open(path, .{ .mode = .read_write_create });
    defer pruner.close();
    try db.applyPragmas(&pruner, .{});

    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    var tx = try db.ReadTx.begin(&reader);
    const before = (try overview(&reader, arena, agg_since, 1000, 1)).totals;
    const watermark_before = try availableSince(&reader);
    try testing.expectEqual(@as(u64, 3), before.queries);

    // The prune commits while the reader's transaction is open.
    const pruned = try pruneOlderThan(&pruner, agg_since + 200);
    try testing.expectEqual(@as(i64, 2), pruned.deleted);
    try testing.expect(pruned.available_since > watermark_before);

    // Neither half of the answer moved: the rows the reader would report and
    // the watermark it would tag them with still describe one state.
    const during = (try overview(&reader, arena, agg_since, 1000, 1)).totals;
    try testing.expectEqual(before.queries, during.queries);
    try testing.expectEqual(watermark_before, try availableSince(&reader));
    try tx.commit();

    // The next response sees the prune — both halves of it.
    const after = (try overview(&reader, arena, agg_since, 1000, 1)).totals;
    try testing.expectEqual(@as(u64, 1), after.queries);
    try testing.expectEqual(pruned.available_since, try availableSince(&reader));
}

// ---------------------------------------------------------------------------
// the projections and the Overview read path (milestone 36)
// ---------------------------------------------------------------------------

/// Floors a `query_log.timestamp` onto the projection grid in SQL. SQLite's `/`
/// truncates toward zero, so a negative timestamp needs the bias — the same
/// reason `bucketOf` is `@divFloor`.
const bucket_expr = "(CASE WHEN timestamp >= 0 THEN timestamp / 1800 * 1800" ++
    " ELSE (timestamp - 1799) / 1800 * 1800 END)";

/// Each projection table beside a from-scratch recomputation of it. Both sides
/// have unique keys, so `EXCEPT` in both directions is an exact equality test
/// rather than a containment one.
const recompute_checks = [_]struct { projection: []const u8, recompute: []const u8 }{
    .{
        .projection = "SELECT bucket, queries, blocked, cached, rt_sum, rt_count FROM bucket_totals",
        .recompute = "SELECT " ++ bucket_expr ++ ", count(*), coalesce(sum(blocked <> 0), 0)," ++
            " coalesce(sum(cache_hit = 1), 0), coalesce(sum(response_time_us), 0)," ++
            " count(response_time_us) FROM query_log GROUP BY 1",
    },
    .{
        .projection = "SELECT bucket, client_ip, queries FROM bucket_clients",
        .recompute = "SELECT " ++ bucket_expr ++ ", client_ip, count(*) FROM query_log GROUP BY 1, 2",
    },
    .{
        .projection = "SELECT bucket, qtype, count FROM bucket_types",
        .recompute = "SELECT " ++ bucket_expr ++ ", coalesce(qtype, -1), count(*) FROM query_log GROUP BY 1, 2",
    },
    .{
        .projection = "SELECT bucket, route_kind, source_present, source_text, count FROM bucket_routes",
        .recompute = "SELECT " ++ bucket_expr ++ ", route_kind, source IS NOT NULL, coalesce(source, ''), count(*)" ++
            " FROM (SELECT timestamp, route_kind, CASE route_kind WHEN 'upstream' THEN upstream" ++
            " WHEN 'forward_zone' THEN forward_zone END AS source FROM query_log) GROUP BY 1, 2, 3, 4",
    },
};

/// Exported for `querylog_migrations.zig`'s fixture and migration tests: the
/// authority on projection coherence is this file, and a second copy of the
/// recompute SQL there would be free to drift from the writer it checks.
pub fn expectProjectionsMatchRecompute(database: *db.Db) !void {
    for (recompute_checks) |check| {
        var buf: [4096]u8 = undefined;
        const sql = try std.fmt.bufPrint(
            &buf,
            "SELECT (SELECT count(*) FROM ({s} EXCEPT {s})) + (SELECT count(*) FROM ({s} EXCEPT {s}))",
            .{ check.projection, check.recompute, check.recompute, check.projection },
        );
        try testing.expectEqual(@as(i64, 0), try database.queryInt(sql));
    }
}

test "bucketOf floors onto the grid on both sides of the epoch" {
    try testing.expectEqual(@as(i64, 0), bucketOf(0));
    try testing.expectEqual(@as(i64, 0), bucketOf(1799));
    try testing.expectEqual(@as(i64, grain), bucketOf(grain));
    try testing.expectEqual(@as(i64, grain), bucketOf(grain + 1));
    // Truncation toward zero would name bucket 0 for all three of these, which
    // is the bucket *after* the one the row belongs to.
    try testing.expectEqual(@as(i64, -grain), bucketOf(-1));
    try testing.expectEqual(@as(i64, -grain), bucketOf(-grain));
    try testing.expectEqual(@as(i64, -2 * grain), bucketOf(-grain - 1));
}

test "a batch maintains all four projections in the transaction that writes the rows" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    var blocked_row = aggRow(0, "192.0.2.20", null, .blocked, null);
    blocked_row.response_time_us = null;
    try writer.writeBatch(&.{
        aggRow(0, "192.0.2.10", 1, .upstream, "9.9.9.9"),
        aggRow(5, "192.0.2.10", 1, .cache, null),
        blocked_row,
        // The next grid bucket, so the floor is what separates them.
        aggRow(grain, "192.0.2.10", 28, .upstream, null),
    });

    // Two grid buckets: three rows floored into the first, one into the second.
    try testing.expectEqual(@as(i64, 2), try database.queryInt("SELECT count(*) FROM bucket_totals"));
    try testing.expectEqual(bucketOf(agg_since), try database.queryInt("SELECT min(bucket) FROM bucket_totals"));
    try testing.expectEqual(@as(i64, 3), try database.queryInt("SELECT queries FROM bucket_totals ORDER BY bucket LIMIT 1"));
    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT blocked FROM bucket_totals ORDER BY bucket LIMIT 1"));
    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT cached FROM bucket_totals ORDER BY bucket LIMIT 1"));
    // The blocked row carries no response time, so two of the three are timed.
    try testing.expectEqual(@as(i64, 2), try database.queryInt("SELECT rt_count FROM bucket_totals ORDER BY bucket LIMIT 1"));
    try expectProjectionsMatchRecompute(&database);

    // A second batch accumulates onto the same keys rather than replacing them.
    try writer.writeBatch(&.{aggRow(7, "192.0.2.10", 1, .upstream, "9.9.9.9")});
    try testing.expectEqual(@as(i64, 4), try database.queryInt("SELECT queries FROM bucket_totals ORDER BY bucket LIMIT 1"));
    try testing.expectEqual(
        @as(i64, 3),
        try database.queryInt("SELECT queries FROM bucket_clients ORDER BY bucket, client_ip LIMIT 1"),
    );
    try expectProjectionsMatchRecompute(&database);
}

test "a NULL qtype and a NULL route source are stored losslessly" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{
        aggRow(0, "192.0.2.10", null, .upstream, null),
        aggRow(1, "192.0.2.10", 1, .upstream, ""),
    });

    try testing.expectEqual(@as(i64, null_qtype), try database.queryInt("SELECT min(qtype) FROM bucket_types"));
    // A row whose resolver the log did not record, and a row whose resolver is
    // the empty string, are different keys — which is what `source_present`
    // buys over "empty means absent".
    try testing.expectEqual(@as(i64, 2), try database.queryInt("SELECT count(*) FROM bucket_routes"));
    try testing.expectEqual(
        @as(i64, 1),
        try database.queryInt("SELECT count(*) FROM bucket_routes WHERE source_present = 0 AND source_text = ''"),
    );
    try testing.expectEqual(
        @as(i64, 1),
        try database.queryInt("SELECT count(*) FROM bucket_routes WHERE source_present = 1 AND source_text = ''"),
    );
    try expectProjectionsMatchRecompute(&database);
}

test "a failed projection update rolls the whole batch back and the writer survives it" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    try writer.writeBatch(&.{aggRow(0, "192.0.2.10", 1, .upstream, "9.9.9.9")});
    try database.exec(
        \\CREATE TRIGGER refuse_projection BEFORE INSERT ON bucket_clients
        \\WHEN new.client_ip = 'boom'
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    try testing.expectError(error.Constraint, writer.writeBatch(&.{
        aggRow(1, "boom", 1, .upstream, "9.9.9.9"),
    }));

    // The raw row went in before the projection statement failed; neither
    // survived, because they were one transaction.
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try expectProjectionsMatchRecompute(&database);

    try database.exec("DROP TRIGGER refuse_projection;");
    try writer.writeBatch(&.{aggRow(2, "192.0.2.11", 1, .upstream, "9.9.9.9")});
    try testing.expectEqual(@as(i64, 2), try countRows(&database));
    try expectProjectionsMatchRecompute(&database);
}

test "a prune on the grid drops whole projection buckets and leaves the rest" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const base = bucketOf(agg_since);
    try seedAt(&database, &.{ base, base + grain, base + 2 * grain });

    _ = try pruneOlderThan(&database, base + 2 * grain);
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM bucket_totals"));
    try testing.expectEqual(base + 2 * grain, try database.queryInt("SELECT bucket FROM bucket_totals"));
    try expectProjectionsMatchRecompute(&database);
}

test "a prune off the grid recomputes the straddling bucket instead of estimating it" {
    var database = try openLog();
    defer database.close();

    const base = bucketOf(agg_since);
    // Four rows in one bucket; the cutoff falls between the second and third.
    try seedAt(&database, &.{ base, base + 100, base + 500, base + 900 });

    _ = try pruneOlderThan(&database, base + 400);
    try testing.expectEqual(@as(i64, 2), try countRows(&database));
    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT count(*) FROM bucket_totals"));
    try testing.expectEqual(@as(i64, 2), try database.queryInt("SELECT queries FROM bucket_totals"));
    try expectProjectionsMatchRecompute(&database);

    // Emptying the straddling bucket removes its row rather than writing zeros.
    _ = try pruneOlderThan(&database, base + grain - 1);
    try testing.expectEqual(@as(i64, 0), try countRows(&database));
    try testing.expectEqual(@as(i64, 0), try database.queryInt("SELECT count(*) FROM bucket_totals"));
    try expectProjectionsMatchRecompute(&database);
}

test "a failed projection prune rolls back the raw delete and the watermark with it" {
    var database = try openLog();
    defer database.close();

    const base = bucketOf(agg_since);
    try seedAt(&database, &.{ base, base + grain });
    const watermark = try availableSince(&database);

    try database.exec(
        \\CREATE TRIGGER refuse_bucket_prune BEFORE DELETE ON bucket_clients
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    try testing.expectError(error.Constraint, pruneOlderThan(&database, base + grain));

    try testing.expectEqual(@as(i64, 2), try countRows(&database));
    try testing.expectEqual(watermark, try availableSince(&database));
    try expectProjectionsMatchRecompute(&database);

    try database.exec("DROP TRIGGER refuse_bucket_prune;");
    _ = try pruneOlderThan(&database, base + grain);
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try expectProjectionsMatchRecompute(&database);
}

test "a failed straddling-bucket replacement rolls back the raw delete and the watermark with it" {
    var database = try openLog();
    defer database.close();

    const base = bucketOf(agg_since);
    // One whole bucket to delete outright, and one the cutoff cuts in half.
    try seedAt(&database, &.{ base, base + grain, base + grain + 100, base + grain + 900 });
    const watermark = try availableSince(&database);
    const straddling_totals = "SELECT queries FROM bucket_totals ORDER BY bucket DESC LIMIT 1";
    const straddling_queries = try database.queryInt(straddling_totals);

    // Only the recompute inserts into a projection table; the bulk prune above
    // it deletes. So this trigger fires during the straddling-bucket
    // replacement and nowhere else in the pass.
    try database.exec(
        \\CREATE TRIGGER refuse_recompute BEFORE INSERT ON bucket_totals
        \\BEGIN SELECT RAISE(ABORT, 'refused'); END;
    );

    try testing.expectError(error.Constraint, pruneOlderThan(&database, base + grain + 500));

    // The raw delete, the watermark advance, the whole-bucket projection
    // deletes and the straddling bucket's own delete all went back together.
    try testing.expectEqual(@as(i64, 4), try countRows(&database));
    try testing.expectEqual(watermark, try availableSince(&database));
    try testing.expectEqual(@as(i64, 2), try database.queryInt("SELECT count(*) FROM bucket_totals"));
    try testing.expectEqual(straddling_queries, try database.queryInt(straddling_totals));
    try expectProjectionsMatchRecompute(&database);

    try database.exec("DROP TRIGGER refuse_recompute;");
    _ = try pruneOlderThan(&database, base + grain + 500);
    try testing.expectEqual(@as(i64, 1), try countRows(&database));
    try testing.expectEqual(@as(i64, 1), try database.queryInt("SELECT queries FROM bucket_totals"));
    try expectProjectionsMatchRecompute(&database);
}

/// One row per timestamp, all in the same batch, with the fixture defaults.
fn seedAt(database: *db.Db, timestamps: []const i64) !void {
    var writer = try BatchWriter.init(testing.allocator, database);
    defer writer.deinit();

    var rows: [16]Row = undefined;
    for (rows[0..timestamps.len], timestamps) |*row, ts| {
        row.* = plainRow(ts, "example.com");
    }
    try writer.writeBatch(rows[0..timestamps.len]);
}

test "an arbitrary interleaving of batches and prunes leaves every projection exact" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    var prng: std.Random.DefaultPrng = .init(0x36a5_0000_36a5);
    const random = prng.random();

    const kinds = [_]provenance.RouteKind{ .blocked, .local, .forward_zone, .upstream, .cache, .rejected };
    const sources = [_]?[]const u8{ "9.9.9.9", "https://a.example/dns-query", null, "" };
    const base = bucketOf(agg_since);

    var rows: [24]Row = undefined;
    var addresses: [24][20]u8 = undefined;

    var step: usize = 0;
    while (step < 60) : (step += 1) {
        if (step % 5 == 4) {
            // Half the cutoffs land off the grid on purpose: the straddling
            // bucket is the only case a delete alone cannot get right.
            const offset = random.intRangeAtMost(i64, 0, 6 * grain);
            const skew: i64 = if (random.boolean()) 0 else random.intRangeLessThan(i64, 1, grain);
            _ = try pruneOlderThan(&database, base + offset + skew);
        } else {
            const count = random.intRangeAtMost(usize, 1, rows.len);
            for (rows[0..count], addresses[0..count]) |*row, *address| {
                const client = try std.fmt.bufPrint(address, "192.0.2.{d}", .{random.intRangeAtMost(u8, 1, 5)});
                const kind = kinds[random.uintLessThan(usize, kinds.len)];
                row.* = plainRow(base + random.intRangeAtMost(i64, 0, 8 * grain), "example.com");
                row.client_ip = client;
                row.qtype = if (random.boolean()) null else random.intRangeAtMost(u16, 1, 3);
                row.blocked = kind == .blocked;
                row.cache_hit = if (random.boolean()) null else kind == .cache;
                row.response_time_us = if (random.boolean()) null else random.intRangeAtMost(i64, 0, 5000);
                row.route_kind = kind;
                const source = sources[random.uintLessThan(usize, sources.len)];
                row.upstream = if (kind == .upstream) source else null;
                row.forward_zone = if (kind == .forward_zone) source else null;
                row.policy_action = if (kind == .blocked) .block else .allow;
                row.policy_reason = if (kind == .blocked) .blocklist_domain else .no_match;
            }
            try writer.writeBatch(rows[0..count]);
        }
        try expectProjectionsMatchRecompute(&database);
    }

    // The walk has to have exercised both halves of it.
    try testing.expect((try countRows(&database)) > 0);
}

// --- the Overview equivalence oracle --------------------------------------

/// A test-only copy of the five SQL aggregates `overview` replaces, kept here
/// so it survives their deletion. It is the independent statement of the
/// contract: if `overview` and this disagree on any window, one of them is
/// wrong, and this one is the one the goldens were written against.
const oracle = struct {
    const totals_sql =
        \\SELECT count(*),
        \\       coalesce(sum(blocked <> 0), 0),
        \\       count(DISTINCT client_ip),
        \\       coalesce(sum(response_time_us), 0),
        \\       count(response_time_us)
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?2
    ;

    const buckets_sql =
        \\SELECT (timestamp - ?1) / ?2,
        \\       count(*),
        \\       coalesce(sum(blocked <> 0), 0),
        \\       coalesce(sum(cache_hit = 1), 0)
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?3
        \\ GROUP BY 1
    ;

    const types_sql =
        \\SELECT qtype, count(*)
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?2
        \\ GROUP BY qtype
        \\ ORDER BY count(*) DESC, qtype IS NULL, qtype ASC
    ;

    const routes_sql =
        \\SELECT route_kind,
        \\       CASE route_kind
        \\            WHEN 'upstream' THEN upstream
        \\            WHEN 'forward_zone' THEN forward_zone
        \\       END AS source,
        \\       count(*)
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?2
        \\ GROUP BY route_kind, source
        \\ ORDER BY count(*) DESC, route_kind ASC, source IS NULL, source ASC
    ;

    const clients_rank_sql =
        \\SELECT client_ip
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?2
        \\ GROUP BY client_ip
        \\ ORDER BY count(*) DESC, client_ip ASC
        \\ LIMIT ?3
    ;

    const clients_buckets_sql =
        \\SELECT client_ip, (timestamp - ?1) / ?2, count(*)
        \\  FROM query_log
        \\ WHERE timestamp >= ?1 AND timestamp < ?3
        \\ GROUP BY 1, 2
    ;

    fn build(
        database: *db.Db,
        arena: Allocator,
        since: i64,
        bucket_seconds: u32,
        bucket_count: u32,
    ) !Overview {
        const width: i64 = bucket_seconds;
        const until = since + width * bucket_count;

        var totals: StatsTotals = undefined;
        {
            var stmt = try database.prepare(totals_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, until);
            try testing.expect(try stmt.step());
            const timed = stmt.columnInt(4);
            totals = .{
                .queries = try countOf(stmt.columnInt(0)),
                .blocked = try countOf(stmt.columnInt(1)),
                .distinct_clients = try countOf(stmt.columnInt(2)),
                .avg_response_time_us = if (timed == 0) null else @divTrunc(stmt.columnInt(3), timed),
            };
        }

        const buckets = try arena.alloc(Bucket, bucket_count);
        for (buckets, 0..) |*bucket, i| bucket.* = .{
            .ts = since + width * @as(i64, @intCast(i)),
            .queries = 0,
            .blocked = 0,
            .cached = 0,
        };
        {
            var stmt = try database.prepare(buckets_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, width);
            try stmt.bindInt(3, until);
            while (try stmt.step()) {
                const index: usize = @intCast(stmt.columnInt(0));
                buckets[index].queries = try countOf(stmt.columnInt(1));
                buckets[index].blocked = try countOf(stmt.columnInt(2));
                buckets[index].cached = try countOf(stmt.columnInt(3));
            }
        }

        var types: std.ArrayList(TypeCount) = .empty;
        {
            var stmt = try database.prepare(types_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, until);
            while (try stmt.step()) {
                try types.append(arena, .{
                    .qtype = if (stmt.isNull(0)) null else try columnU16(&stmt, 0),
                    .count = try countOf(stmt.columnInt(1)),
                });
            }
        }

        var routes: std.ArrayList(RouteCount) = .empty;
        {
            var stmt = try database.prepare(routes_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, until);
            while (try stmt.step()) {
                try routes.append(arena, .{
                    .route = try provenance.parse(provenance.RouteKind, stmt.columnText(0)),
                    .source = try stmt.columnTextAllocOrNull(arena, 1),
                    .count = try countOf(stmt.columnInt(2)),
                });
            }
        }

        var names: std.ArrayList([]const u8) = .empty;
        var series: std.ArrayList([]u64) = .empty;
        {
            var stmt = try database.prepare(clients_rank_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, until);
            try stmt.bindInt(3, max_client_series);
            while (try stmt.step()) {
                try names.append(arena, try stmt.columnTextAlloc(arena, 0));
                try series.append(arena, try zeroedBuckets(arena, bucket_count));
            }
        }
        const other = try zeroedBuckets(arena, bucket_count);
        {
            var stmt = try database.prepare(clients_buckets_sql);
            defer stmt.deinit();
            try stmt.bindInt(1, since);
            try stmt.bindInt(2, width);
            try stmt.bindInt(3, until);
            while (try stmt.step()) {
                const index: usize = @intCast(stmt.columnInt(1));
                const count = try countOf(stmt.columnInt(2));
                const client = stmt.columnText(0);
                const target = for (names.items, series.items) |name, target_buckets| {
                    if (std.mem.eql(u8, name, client)) break target_buckets;
                } else other;
                target[index] += count;
            }
        }
        const clients = try arena.alloc(ClientSeries, names.items.len);
        for (clients, names.items, series.items) |*entry, name, entry_buckets| {
            entry.* = .{ .client = name, .buckets = entry_buckets };
        }

        return .{
            .totals = totals,
            .buckets = buckets,
            .clients = .{ .clients = clients, .other = other },
            .types = types.items,
            .routes = routes.items,
        };
    }
};

fn expectOverviewEqual(expected: Overview, actual: Overview) !void {
    try testing.expectEqual(expected.totals.queries, actual.totals.queries);
    try testing.expectEqual(expected.totals.blocked, actual.totals.blocked);
    try testing.expectEqual(expected.totals.distinct_clients, actual.totals.distinct_clients);
    try testing.expectEqual(expected.totals.avg_response_time_us, actual.totals.avg_response_time_us);
    try testing.expectEqualSlices(Bucket, expected.buckets, actual.buckets);

    try testing.expectEqual(expected.types.len, actual.types.len);
    for (expected.types, actual.types) |want, got| {
        try testing.expectEqual(want.qtype, got.qtype);
        try testing.expectEqual(want.count, got.count);
    }

    try testing.expectEqual(expected.routes.len, actual.routes.len);
    for (expected.routes, actual.routes) |want, got| {
        try testing.expectEqual(want.route, got.route);
        try testing.expectEqual(want.count, got.count);
        if (want.source) |source| {
            try testing.expectEqualStrings(source, got.source orelse return error.TestExpectedEqual);
        } else {
            try testing.expectEqual(@as(?[]const u8, null), got.source);
        }
    }

    try testing.expectEqual(expected.clients.clients.len, actual.clients.clients.len);
    for (expected.clients.clients, actual.clients.clients) |want, got| {
        try testing.expectEqualStrings(want.client, got.client);
        try testing.expectEqualSlices(u64, want.buckets, got.buckets);
    }
    try testing.expectEqualSlices(u64, expected.clients.other, actual.clients.other);
}

/// Every window the equivalence test walks. The first is the raw path (60 s
/// buckets, the 1h period); the rest are the projection path at each serving
/// width the period grammar offers.
const overview_windows = [_]struct { bucket_seconds: u32, bucket_count: u32 }{
    .{ .bucket_seconds = 60, .bucket_count = 60 },
    .{ .bucket_seconds = 1800, .bucket_count = 48 },
    .{ .bucket_seconds = 3600, .bucket_count = 168 },
    .{ .bucket_seconds = 21600, .bucket_count = 120 },
};

fn expectOverviewMatchesOracle(database: *db.Db, since: i64) !void {
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    for (overview_windows) |window| {
        const want = try oracle.build(database, arena, since, window.bucket_seconds, window.bucket_count);
        const got = try overview(database, arena, since, window.bucket_seconds, window.bucket_count);
        expectOverviewEqual(want, got) catch |err| {
            std.debug.print("overview mismatch at bucket_seconds={d}\n", .{window.bucket_seconds});
            return err;
        };
    }
}

test "overview equals the oracle on an empty database" {
    var database = try openLog();
    defer database.close();
    try expectOverviewMatchesOracle(&database, bucketOf(agg_since));
}

test "overview equals the oracle over a window with no rows in it" {
    var database = try openLog();
    defer database.close();
    try seedAt(&database, &.{ agg_since, agg_since + 10 });

    // Two full 30-day spans later: every panel is empty and the axis is still
    // whole.
    try expectOverviewMatchesOracle(&database, bucketOf(agg_since) + 240 * grain);
}

test "overview equals the oracle across nulls, ties and a bucket still in progress" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const since = bucketOf(agg_since);
    const kinds = [_]provenance.RouteKind{ .blocked, .local, .forward_zone, .upstream, .cache, .rejected };

    // Ten clients so the top-eight cut has a residual, and the counts are
    // deliberately equal in pairs so the `client_ip` tie-break decides.
    var rows: [120]Row = undefined;
    var addresses: [120][20]u8 = undefined;
    for (&rows, &addresses, 0..) |*row, *address, i| {
        const client = try std.fmt.bufPrint(address, "192.0.2.{d}", .{100 + (i / 2) % 10});
        const kind = kinds[i % kinds.len];
        row.* = plainRow(since + @as(i64, @intCast(i)) * 137, "example.com");
        row.client_ip = client;
        // Every seventh row carries no query type, and the types collide in
        // count so the null-last tie-break is exercised.
        row.qtype = if (i % 7 == 0) null else @intCast(1 + i % 3);
        row.blocked = kind == .blocked;
        row.cache_hit = if (i % 5 == 0) null else kind == .cache;
        row.response_time_us = if (i % 4 == 0) null else @intCast(100 + i);
        row.route_kind = kind;
        // Every eleventh upstream row records no resolver: the NULL-source
        // group the routes breakdown must keep rather than drop.
        const source: ?[]const u8 = if (i % 11 == 0) null else if (i % 2 == 0) "9.9.9.9" else "https://a.example/dns-query";
        row.upstream = if (kind == .upstream) source else null;
        row.forward_zone = if (kind == .forward_zone) source else null;
        row.policy_action = if (kind == .blocked) .block else .allow;
        row.policy_reason = if (kind == .blocked) .blocklist_domain else .no_match;
    }
    try writer.writeBatch(&rows);

    try expectOverviewMatchesOracle(&database, since);

    // A window whose last bucket is only partly filled: `since` moved so the
    // newest rows land inside the final bucket rather than closing it.
    try expectOverviewMatchesOracle(&database, since - 40 * grain);

    // And after a prune off the grid, so the projections the read path uses are
    // the recomputed ones.
    _ = try pruneOlderThan(&database, since + 900);
    try expectProjectionsMatchRecompute(&database);
    try expectOverviewMatchesOracle(&database, since);
}

test "overview equals the oracle after an interleaving of batches and prunes" {
    var database = try openLog();
    defer database.close();
    var writer = try BatchWriter.init(testing.allocator, &database);
    defer writer.deinit();

    const since = bucketOf(agg_since);
    var prng: std.Random.DefaultPrng = .init(0x0e17_3a99);
    const random = prng.random();

    var rows: [16]Row = undefined;
    var addresses: [16][20]u8 = undefined;
    var step: usize = 0;
    while (step < 12) : (step += 1) {
        const count = random.intRangeAtMost(usize, 1, rows.len);
        for (rows[0..count], addresses[0..count]) |*row, *address| {
            const client = try std.fmt.bufPrint(address, "10.0.0.{d}", .{random.intRangeAtMost(u8, 1, 12)});
            row.* = plainRow(since + random.intRangeAtMost(i64, 0, 20 * grain), "example.com");
            row.client_ip = client;
            row.qtype = if (random.boolean()) null else random.intRangeAtMost(u16, 1, 4);
            row.response_time_us = if (random.boolean()) null else random.intRangeAtMost(i64, 0, 9000);
            row.cache_hit = if (random.boolean()) null else random.boolean();
        }
        try writer.writeBatch(rows[0..count]);
        if (step % 4 == 3) {
            _ = try pruneOlderThan(&database, since + random.intRangeAtMost(i64, 0, 10 * grain));
        }
        try expectOverviewMatchesOracle(&database, since);
    }
}

test "overview refuses a window the projections cannot express" {
    var database = try openLog();
    defer database.close();
    var arena_state: std.heap.ArenaAllocator = .init(testing.allocator);
    defer arena_state.deinit();
    const arena = arena_state.allocator();

    const since = bucketOf(agg_since);

    try testing.expectError(error.Misuse, overview(&database, arena, since, 0, 48));
    try testing.expectError(error.Misuse, overview(&database, arena, since, 1800, 0));
    try testing.expectError(error.Misuse, overview(&database, arena, std.math.maxInt(i64) - 1, 3600, 48));

    // On the projection path the grid is part of the contract: an unaligned
    // start or a width that is not a whole number of grains cannot be answered
    // from 30-minute rows, and guessing would be worse than refusing.
    try testing.expectError(error.Misuse, overview(&database, arena, since + 1, 1800, 48));
    try testing.expectError(error.Misuse, overview(&database, arena, since, 2700, 48));

    // Below the grain none of that applies: the raw rows carry every second.
    _ = try overview(&database, arena, since + 1, 60, 60);
}