@camstack/system 1.2.390 → 1.2.392

package/dist/builtins/sqlite-storage/query-profiler.d.ts

@@ -1,242 +1,242 @@
-import { IScopedLogger } from '@camstack/types';
+import type { IScopedLogger } from "@camstack/types";
 /**
- * Per-operation timing for the synchronous SQLite engine.
- *
- * ## Why this exists
- *
- * `hub/sqlite-settings` is a forked runner with ONE JS thread and a
- * **synchronous** SQLite binding: every statement it executes stops the event
- * loop for the whole of its duration, and every configuration read in the
- * cluster queues behind it. On 2026-08-25 that runner reported
- * `loopMax=68283ms` while the container had zero memory pressure
- * (`memory.events max` 0/min, cgroup 16-20 GB of 24) and the NVMe was idle
- * (`%util` 1.5-8%, `r_await` 0.14 ms) — so neither of the two previously known
- * causes applied, and there was **not one line in the log saying what the
- * runner was doing**. Diagnosis fell back to sampling `/proc/<pid>/syscall`
- * from the host, which is not a thing an operator can be asked to do.
- *
- * A 120 s census of the runner's main thread (2 216 samples at 54 ms) read
- * `D=41% R=50% S=9%` against 13.9 s of CPU: the thread was blocked in
- * `pread64 → filemap_read → folio_wait_bit_common` — synchronous 4 KiB page-ins
- * on `camstack.db` — for two fifths of wall-clock, at a sustained 418 read
- * syscalls per second. What it could NOT say is which query issued them.
- *
- * ## Why an aggregate, and not only a slow-query log
- *
- * The 68 s stall was **not** one 68 s statement. At 418 page reads per second
- * the cost is spread across many statements that are each individually
- * unremarkable, so a threshold-only log would have reported nothing at all
- * while the loop sat still. {@link SqliteOpProfiler.report} therefore prints
- * the window's busiest `(op, collection)` pairs and the share of wall-clock
- * the engine spent inside SQLite — the number that says "this runner is
- * saturated" even when no single call is.
- *
- * The slow-call WARN is kept for the other shape of the same failure: one
- * statement that really does take seconds. Both are needed; neither is
- * sufficient.
- *
- * ## Why the SQL shape, and why the plan
- *
- * `(op, collection)` names a TABLE, not a call site. On 2026-08-25 the busiest
- * line of a 90 s operator stall was `query pipeline-analytics:tracks — 1005
- * calls, max 1615 ms` and eleven different reads in one addon are spelled
- * exactly that way; the aggregate accused all of them and cleared none. So the
- * shape of the prepared statement (`statement-shape.ts`) is part of the
- * aggregate KEY, and the report prints it.
- *
- * A shape says WHICH query. It does not say why it is slow — "1.6 s for 501
- * rows" is a full scan and "1.6 s for 501 rows" is also a cold disk, and those
- * are the same line. `EXPLAIN QUERY PLAN` is the only thing that separates
- * them, and it is attached to the slow-call WARN because that is where the
- * question is asked. It is asked ONCE PER SHAPE PER WINDOW: a scan that fires
- * the WARN fires it repeatedly, and a plan re-printed forty times is how the
- * one line that mattered stops being read.
- *
- * ## Cost
- *
- * Two `Date.now()` calls and one `Map` lookup per statement. At the measured
- * live rate (a few hundred statements/s) that is far below the noise floor of
- * the work being measured — and unlike a sampling profiler it is always on,
- * which is the point: the stall that matters is the one nobody was watching.
- */
+* Per-operation timing for the synchronous SQLite engine.
+*
+* ## Why this exists
+*
+* `hub/sqlite-settings` is a forked runner with ONE JS thread and a
+* **synchronous** SQLite binding: every statement it executes stops the event
+* loop for the whole of its duration, and every configuration read in the
+* cluster queues behind it. On 2026-08-25 that runner reported
+* `loopMax=68283ms` while the container had zero memory pressure
+* (`memory.events max` 0/min, cgroup 16-20 GB of 24) and the NVMe was idle
+* (`%util` 1.5-8%, `r_await` 0.14 ms) — so neither of the two previously known
+* causes applied, and there was **not one line in the log saying what the
+* runner was doing**. Diagnosis fell back to sampling `/proc/<pid>/syscall`
+* from the host, which is not a thing an operator can be asked to do.
+*
+* A 120 s census of the runner's main thread (2 216 samples at 54 ms) read
+* `D=41% R=50% S=9%` against 13.9 s of CPU: the thread was blocked in
+* `pread64 → filemap_read → folio_wait_bit_common` — synchronous 4 KiB page-ins
+* on `camstack.db` — for two fifths of wall-clock, at a sustained 418 read
+* syscalls per second. What it could NOT say is which query issued them.
+*
+* ## Why an aggregate, and not only a slow-query log
+*
+* The 68 s stall was **not** one 68 s statement. At 418 page reads per second
+* the cost is spread across many statements that are each individually
+* unremarkable, so a threshold-only log would have reported nothing at all
+* while the loop sat still. {@link SqliteOpProfiler.report} therefore prints
+* the window's busiest `(op, collection)` pairs and the share of wall-clock
+* the engine spent inside SQLite — the number that says "this runner is
+* saturated" even when no single call is.
+*
+* The slow-call WARN is kept for the other shape of the same failure: one
+* statement that really does take seconds. Both are needed; neither is
+* sufficient.
+*
+* ## Why the SQL shape, and why the plan
+*
+* `(op, collection)` names a TABLE, not a call site. On 2026-08-25 the busiest
+* line of a 90 s operator stall was `query pipeline-analytics:tracks — 1005
+* calls, max 1615 ms` and eleven different reads in one addon are spelled
+* exactly that way; the aggregate accused all of them and cleared none. So the
+* shape of the prepared statement (`statement-shape.ts`) is part of the
+* aggregate KEY, and the report prints it.
+*
+* A shape says WHICH query. It does not say why it is slow — "1.6 s for 501
+* rows" is a full scan and "1.6 s for 501 rows" is also a cold disk, and those
+* are the same line. `EXPLAIN QUERY PLAN` is the only thing that separates
+* them, and it is attached to the slow-call WARN because that is where the
+* question is asked. It is asked ONCE PER SHAPE PER WINDOW: a scan that fires
+* the WARN fires it repeatedly, and a plan re-printed forty times is how the
+* one line that mattered stops being read.
+*
+* ## Cost
+*
+* Two `Date.now()` calls and one `Map` lookup per statement. At the measured
+* live rate (a few hundred statements/s) that is far below the noise floor of
+* the work being measured — and unlike a sampling profiler it is always on,
+* which is the point: the stall that matters is the one nobody was watching.
+*/
 /**
- * A single statement whose duration is being recorded.
- *
- * `rows` is "rows touched" — returned by a read, affected by a write. It is
- * what separates "this query is slow because it is scanning" from "this query
- * is slow because the disk is".
- */
+* A single statement whose duration is being recorded.
+*
+* `rows` is "rows touched" — returned by a read, affected by a write. It is
+* what separates "this query is slow because it is scanning" from "this query
+* is slow because the disk is".
+*/
 export interface SqliteOpSample {
-    /** Backend method that ran the statement: `query`, `count`, `set`, … */
-    readonly op: string;
-    /** Scoped table name (`<namespace>:<collection>`), never a device name. */
-    readonly collection: string;
-    readonly durationMs: number;
-    readonly rows: number;
-    /**
-     * Numeric device id when the statement was issued on behalf of one. Every
-     * log line about a device carries `tags: { deviceId }` — a stall is always
-     * asked per-camera, and a line without the tag cannot answer it.
-     */
-    readonly deviceId?: number;
-    /** See {@link SqliteOpLabel.owner}. */
-    readonly owner?: string;
-    /** See {@link SqliteOpLabel.sql}. */
-    readonly sql?: string;
-    /** See {@link SqliteOpLabel.params}. */
-    readonly params?: readonly unknown[];
+	/** Backend method that ran the statement: `query`, `count`, `set`, … */
+	readonly op: string;
+	/** Scoped table name (`<namespace>:<collection>`), never a device name. */
+	readonly collection: string;
+	readonly durationMs: number;
+	readonly rows: number;
+	/**
+	* Numeric device id when the statement was issued on behalf of one. Every
+	* log line about a device carries `tags: { deviceId }` — a stall is always
+	* asked per-camera, and a line without the tag cannot answer it.
+	*/
+	readonly deviceId?: number;
+	/** See {@link SqliteOpLabel.owner}. */
+	readonly owner?: string;
+	/** See {@link SqliteOpLabel.sql}. */
+	readonly sql?: string;
+	/** See {@link SqliteOpLabel.params}. */
+	readonly params?: readonly unknown[];
 }
 /** One `(op, collection, owner, shape)` tuple's totals over a report window. */
 export interface SqliteOpStat {
-    readonly op: string;
-    readonly collection: string;
-    readonly owner?: string;
-    /** `statement-shape.ts` fingerprint; absent for a fixed-shape statement the
-     *  engine did not hand to the profiler. */
-    readonly shape?: string;
-    readonly calls: number;
-    readonly totalMs: number;
-    readonly maxMs: number;
-    readonly rows: number;
+	readonly op: string;
+	readonly collection: string;
+	readonly owner?: string;
+	/** `statement-shape.ts` fingerprint; absent for a fixed-shape statement the
+	*  engine did not hand to the profiler. */
+	readonly shape?: string;
+	readonly calls: number;
+	readonly totalMs: number;
+	readonly maxMs: number;
+	readonly rows: number;
 }
 /**
- * Runs `EXPLAIN QUERY PLAN` for a statement on the connection that executed it.
- *
- * Injected rather than reached for: the profiler must not own a database
- * handle, and the plan is only trustworthy from the SAME connection — a second
- * handle can differ in `sqlite_stat1` visibility and in temp-schema state, and
- * a plan taken from a connection that is not the one that stalled describes a
- * query nobody ran.
- *
- * Returns one string per plan row (SQLite's `detail` column). Must not throw —
- * an unexplainable statement yields `[]`.
- */
+* Runs `EXPLAIN QUERY PLAN` for a statement on the connection that executed it.
+*
+* Injected rather than reached for: the profiler must not own a database
+* handle, and the plan is only trustworthy from the SAME connection — a second
+* handle can differ in `sqlite_stat1` visibility and in temp-schema state, and
+* a plan taken from a connection that is not the one that stalled describes a
+* query nobody ran.
+*
+* Returns one string per plan row (SQLite's `detail` column). Must not throw —
+* an unexplainable statement yields `[]`.
+*/
 export type SqlExplainer = (sql: string, params: readonly unknown[]) => readonly string[];
 /**
- * Duration at which one statement is worth a WARN on its own, in ms.
- *
- * 1 000 ms is a full second of every cluster-wide configuration read queued
- * behind one statement. Below that the aggregate report is the right place to
- * see it; above it, the statement is the story and deserves its own line.
- */
-export declare const SQLITE_SLOW_CALL_MS = 1000;
+* Duration at which one statement is worth a WARN on its own, in ms.
+*
+* 1 000 ms is a full second of every cluster-wide configuration read queued
+* behind one statement. Below that the aggregate report is the right place to
+* see it; above it, the statement is the story and deserves its own line.
+*/
+export declare const SQLITE_SLOW_CALL_MS = 1e3;
 /** How often the aggregate report is emitted, in ms. Matched to the runner's
- *  own `[mem]` heartbeat so the two lines can be read side by side. */
-export declare const SQLITE_PROFILE_REPORT_INTERVAL_MS = 60000;
+*  own `[mem]` heartbeat so the two lines can be read side by side. */
+export declare const SQLITE_PROFILE_REPORT_INTERVAL_MS = 6e4;
 /**
- * Engine-busy time below which a window is not worth a line, in ms.
- *
- * 250 ms in a 60 s window is 0.4% — an engine doing its job. Reporting below
- * this would be a line a minute forever describing a runner that is fine,
- * which is how a log stops being read.
- */
+* Engine-busy time below which a window is not worth a line, in ms.
+*
+* 250 ms in a 60 s window is 0.4% — an engine doing its job. Reporting below
+* this would be a line a minute forever describing a runner that is fine,
+* which is how a log stops being read.
+*/
 export declare const SQLITE_PROFILE_REPORT_FLOOR_MS = 250;
 /** Pairs named in one report line. Enough to see the shape, short enough to read. */
 export declare const SQLITE_PROFILE_TOP_N = 5;
 /**
- * Share of window wall-clock at which ONE shape is a storm, not a busy engine.
- *
- * Measured 2026-08-31: the OFFSET census on `pipeline-analytics:media` was
- * 88.7 % of wall for six hours. A healthy bounded census (20 s budget in a
- * 60 s window) is 33 % and must not warn. Half the loop is the line between
- * "the engine is working" and "the cluster is queuing behind one query".
- *
- * Cap-usage rate cannot name this class — see `sqliteEngineStorm`.
- */
-export declare const SQLITE_ENGINE_STORM_WALL_SHARE = 0.5;
+* Share of window wall-clock at which ONE shape is a storm, not a busy engine.
+*
+* Measured 2026-08-31: the OFFSET census on `pipeline-analytics:media` was
+* 88.7 % of wall for six hours. A healthy bounded census (20 s budget in a
+* 60 s window) is 33 % and must not warn. Half the loop is the line between
+* "the engine is working" and "the cluster is queuing behind one query".
+*
+* Cap-usage rate cannot name this class — see `sqliteEngineStorm`.
+*/
+export declare const SQLITE_ENGINE_STORM_WALL_SHARE = .5;
 /**
- * The dominating shape if it held {@link SQLITE_ENGINE_STORM_WALL_SHARE} of
- * the window, otherwise `null`.
- *
- * Share is of WINDOW wall, not of busy time: an engine that is 90 % busy with
- * ten equal shapes is healthy; one shape at 50 % of wall is the 2026-08-31
- * store saturation regardless of what else ran. `stats` must already be
- * busiest-first (as {@link SqliteOpProfiler.drain} returns them).
- */
+* The dominating shape if it held {@link SQLITE_ENGINE_STORM_WALL_SHARE} of
+* the window, otherwise `null`.
+*
+* Share is of WINDOW wall, not of busy time: an engine that is 90 % busy with
+* ten equal shapes is healthy; one shape at 50 % of wall is the 2026-08-31
+* store saturation regardless of what else ran. `stats` must already be
+* busiest-first (as {@link SqliteOpProfiler.drain} returns them).
+*/
 export declare function sqliteEngineStorm(stats: readonly SqliteOpStat[], windowMs: number): SqliteOpStat | null;
 export interface SqliteOpProfilerOptions {
-    readonly logger: IScopedLogger;
-    /** Injected for tests; production reads the wall clock. */
-    readonly now?: () => number;
-    readonly slowCallMs?: number;
-    readonly reportIntervalMs?: number;
-    readonly reportFloorMs?: number;
-    readonly topN?: number;
-    /** Absent ⇒ slow-call WARNs carry no plan. See {@link SqlExplainer}. */
-    readonly explain?: SqlExplainer;
+	readonly logger: IScopedLogger;
+	/** Injected for tests; production reads the wall clock. */
+	readonly now?: () => number;
+	readonly slowCallMs?: number;
+	readonly reportIntervalMs?: number;
+	readonly reportFloorMs?: number;
+	readonly topN?: number;
+	/** Absent ⇒ slow-call WARNs carry no plan. See {@link SqlExplainer}. */
+	readonly explain?: SqlExplainer;
 }
 /** Identity of a measured call, without the numbers `measure` supplies itself. */
 export interface SqliteOpLabel {
-    readonly op: string;
-    readonly collection: string;
-    readonly deviceId?: number;
-    /**
-     * The addon a scoped row belongs to, when the filter named one (see
-     * `scopeOwnerOf`). Part of the aggregate KEY, not decoration: without it the
-     * report can say a collection is being hammered and never by whom, which is
-     * exactly where the 67 660-call storm of 2026-08-25 stopped.
-     */
-    readonly owner?: string;
-    /**
-     * The statement the engine prepared, verbatim.
-     *
-     * Supplied only by the ops that BUILD their SQL from a filter — those are the
-     * ones whose cost varies by call site. A fixed-shape statement (`get`, `set`)
-     * is already fully described by `(op, collection)`.
-     *
-     * Used for two things and nothing else: the shape that becomes part of the
-     * aggregate key, and the `EXPLAIN QUERY PLAN` on a slow call. Never logged
-     * raw — see `statement-shape.ts` for what is dropped.
-     */
-    readonly sql?: string;
-    /**
-     * The values bound to {@link sql}.
-     *
-     * Needed because `EXPLAIN QUERY PLAN` on a statement with unbound parameters
-     * is not the plan that ran: SQLite's planner reads bound values for `LIKE`
-     * prefixes and for index-selection on `IN`. Never logged, in whole or in
-     * part — they carry device ids, keys and user data.
-     */
-    readonly params?: readonly unknown[];
+	readonly op: string;
+	readonly collection: string;
+	readonly deviceId?: number;
+	/**
+	* The addon a scoped row belongs to, when the filter named one (see
+	* `scopeOwnerOf`). Part of the aggregate KEY, not decoration: without it the
+	* report can say a collection is being hammered and never by whom, which is
+	* exactly where the 67 660-call storm of 2026-08-25 stopped.
+	*/
+	readonly owner?: string;
+	/**
+	* The statement the engine prepared, verbatim.
+	*
+	* Supplied only by the ops that BUILD their SQL from a filter — those are the
+	* ones whose cost varies by call site. A fixed-shape statement (`get`, `set`)
+	* is already fully described by `(op, collection)`.
+	*
+	* Used for two things and nothing else: the shape that becomes part of the
+	* aggregate key, and the `EXPLAIN QUERY PLAN` on a slow call. Never logged
+	* raw — see `statement-shape.ts` for what is dropped.
+	*/
+	readonly sql?: string;
+	/**
+	* The values bound to {@link sql}.
+	*
+	* Needed because `EXPLAIN QUERY PLAN` on a statement with unbound parameters
+	* is not the plan that ran: SQLite's planner reads bound values for `LIKE`
+	* prefixes and for index-selection on `IN`. Never logged, in whole or in
+	* part — they carry device ids, keys and user data.
+	*/
+	readonly params?: readonly unknown[];
 }
 export declare class SqliteOpProfiler {
-    private readonly logger;
-    private readonly now;
-    private readonly slowCallMs;
-    private readonly reportIntervalMs;
-    private readonly reportFloorMs;
-    private readonly topN;
-    private readonly explain;
-    private readonly stats;
-    /** Shapes already explained in this window — see the class doc on why the
-     *  plan is printed once and not on every WARN. Cleared by `drain`. */
-    private readonly explainedShapes;
-    private windowStartedAt;
-    private timer;
-    constructor(options: SqliteOpProfilerOptions);
-    /**
-     * Time a synchronous SQLite call.
-     *
-     * The elapsed time is recorded on the way out of BOTH exits. A statement
-     * that threw still held the event loop for as long as it ran, and a
-     * profiler that only counts successes reports an idle engine during exactly
-     * the incident worth explaining.
-     */
-    measure<T>(label: SqliteOpLabel, run: () => T): T;
-    record(sample: SqliteOpSample): void;
-    /**
-     * `EXPLAIN QUERY PLAN` for a statement that just held the loop, or nothing.
-     *
-     * Nothing when: no explainer was injected, the op carries no SQL, or this
-     * shape has already been explained in this window. The explainer is
-     * contracted not to throw and is guarded anyway — a profiler that can turn a
-     * slow query into a crash is worse than one that cannot explain it.
-     */
-    private planFor;
-    /** Totals for the window, busiest first, and reset. */
-    drain(): readonly SqliteOpStat[];
-    /** Emit one aggregate line for the window, unless the engine was idle. */
-    report(): void;
-    start(): void;
-    stop(): void;
+	private readonly logger;
+	private readonly now;
+	private readonly slowCallMs;
+	private readonly reportIntervalMs;
+	private readonly reportFloorMs;
+	private readonly topN;
+	private readonly explain;
+	private readonly stats;
+	/** Shapes already explained in this window — see the class doc on why the
+	*  plan is printed once and not on every WARN. Cleared by `drain`. */
+	private readonly explainedShapes;
+	private windowStartedAt;
+	private timer;
+	constructor(options: SqliteOpProfilerOptions);
+	/**
+	* Time a synchronous SQLite call.
+	*
+	* The elapsed time is recorded on the way out of BOTH exits. A statement
+	* that threw still held the event loop for as long as it ran, and a
+	* profiler that only counts successes reports an idle engine during exactly
+	* the incident worth explaining.
+	*/
+	measure<T>(label: SqliteOpLabel, run: () => T): T;
+	record(sample: SqliteOpSample): void;
+	/**
+	* `EXPLAIN QUERY PLAN` for a statement that just held the loop, or nothing.
+	*
+	* Nothing when: no explainer was injected, the op carries no SQL, or this
+	* shape has already been explained in this window. The explainer is
+	* contracted not to throw and is guarded anyway — a profiler that can turn a
+	* slow query into a crash is worse than one that cannot explain it.
+	*/
+	private planFor;
+	/** Totals for the window, busiest first, and reset. */
+	drain(): readonly SqliteOpStat[];
+	/** Emit one aggregate line for the window, unless the engine was idle. */
+	report(): void;
+	start(): void;
+	stop(): void;
 }