@ultimat3/db 1.1.0 → 2.0.0

package/src/client.ts

@@ -4,8 +4,12 @@
 // never opens a socket (the CLI imports it to print help).
 
 import { type Role, resolveRole } from '@ultimat3/core';
-import { dbUnavailable } from './errors';
+import { statementAttribution } from './attribution';
+import { DbError, dbUnavailable, driverError, poolAcquireTimeout, poolMaxInvalid } from './errors';
+import { expectedQueryLoopReason } from './expected-loop';
+import { statementObserver } from './observe';
 import { type SqlFragment, sql } from './sql';
+import { withStatementSpan } from './statement-span';
 import { currentTx } from './transaction';
 
 export interface DbClient {
@@ -15,8 +19,13 @@ export interface DbClient {
   execute(fragment: SqlFragment): Promise<number>;
 }
 
-/** A connection pinned out of the pool. `withTransaction` needs one so BEGIN/COMMIT agree. */
-export interface DbConnection extends DbClient {
+/**
+ * A connection pinned out of the pool. `withTransaction` needs one so BEGIN/COMMIT agree.
+ * `Disposable`, so `using connection = await client.reserve()` gives the pin back on every exit
+ * path — the hand-rolled `finally` is what forgets it on the one path nobody wrote a test for.
+ */
+export interface DbConnection extends DbClient, Disposable {
+  /** Idempotent, and `[Symbol.dispose]` is the same call: releasing twice releases once. */
   release(): void;
 }
 
@@ -33,22 +42,94 @@ export interface PoolProfile {
   /** 0 disables the timeout — only `migrate`, which is allowed to take as long as it takes. */
   readonly statementTimeoutMs: number;
   readonly idleTimeoutMs: number;
+  /**
+   * How long a statement may **wait for a lock** before `55P03`, distinct from how long it may run.
+   * 0 everywhere but `migrate`, which is the only role that takes `ACCESS EXCLUSIVE`: an `alter
+   * table` queued behind a long `SELECT` puts every later query on that table behind it too,
+   * because Postgres' lock queue is FIFO — and `migrate` runs `statement_timeout = 0`, so nothing
+   * else would ever end the wait. Read by `migrate()` as a `SET LOCAL`, never by the pool.
+   */
+  readonly lockTimeoutMs: number;
+  /**
+   * How long `reserve()` may wait for a free connection before `X_DB_POOL_EXHAUSTED`. 0 waits
+   * forever, which is what a run-once role wants and what a request-serving one must never do:
+   * queueing turns exhaustion into a hang, `/readyz`'s `select 1` joins the same queue, the kubelet
+   * kills the pod, and the replacement inherits the same saturated database.
+   */
+  readonly acquireTimeoutMs: number;
 }
 
 /** Sized per role because the failure modes differ: RPS bursts vs. queue depth vs. run-once. */
 export const POOL_PROFILES: Readonly<Record<Role, PoolProfile>> = Object.freeze({
-  web: { max: 20, statementTimeoutMs: 10_000, idleTimeoutMs: 30_000 },
-  sync: { max: 10, statementTimeoutMs: 10_000, idleTimeoutMs: 60_000 },
-  worker: { max: 8, statementTimeoutMs: 120_000, idleTimeoutMs: 30_000 },
-  scheduler: { max: 2, statementTimeoutMs: 15_000, idleTimeoutMs: 60_000 },
-  migrate: { max: 1, statementTimeoutMs: 0, idleTimeoutMs: 10_000 },
-  replicator: { max: 4, statementTimeoutMs: 0, idleTimeoutMs: 60_000 },
+  web: {
+    max: 20,
+    statementTimeoutMs: 10_000,
+    idleTimeoutMs: 30_000,
+    lockTimeoutMs: 0,
+    acquireTimeoutMs: 5_000,
+  },
+  sync: {
+    max: 10,
+    statementTimeoutMs: 10_000,
+    idleTimeoutMs: 60_000,
+    lockTimeoutMs: 0,
+    acquireTimeoutMs: 5_000,
+  },
+  worker: {
+    max: 8,
+    statementTimeoutMs: 120_000,
+    idleTimeoutMs: 30_000,
+    lockTimeoutMs: 0,
+    acquireTimeoutMs: 10_000,
+  },
+  scheduler: {
+    max: 2,
+    statementTimeoutMs: 15_000,
+    idleTimeoutMs: 60_000,
+    lockTimeoutMs: 0,
+    acquireTimeoutMs: 10_000,
+  },
+  // `migrate` waits: its pool is `max: 1` and the advisory-lock pin holds it for the whole run, so
+  // a deadline here would refuse the migration's own session. The wait that needed bounding is the
+  // advisory lock's, and `MIGRATION_LOCK_WAIT_MS` bounds it.
+  migrate: {
+    max: 1,
+    statementTimeoutMs: 0,
+    idleTimeoutMs: 10_000,
+    lockTimeoutMs: 3_000,
+    acquireTimeoutMs: 0,
+  },
+  replicator: {
+    max: 4,
+    statementTimeoutMs: 0,
+    idleTimeoutMs: 60_000,
+    lockTimeoutMs: 0,
+    acquireTimeoutMs: 0,
+  },
 });
 
 export function poolProfileFor(role: Role = resolveRole()): PoolProfile {
   return POOL_PROFILES[role];
 }
 
+/** The one pool knob an operator can turn without a rebuild. Layered over the role default. */
+export const POOL_MAX_ENV = 'DATABASE_POOL_MAX';
+
+/**
+ * `DATABASE_POOL_MAX`, or nothing. `POOL_PROFILES` is frozen into the build, so before this the
+ * only way to change a fleet's connection count was to ship a new image — and 400 `web` pods at
+ * `max: 20` is 8,000 backends against a `max_connections` of 450. An unparseable value **refuses**
+ * rather than falling back: a fleet that ignored the number it was given is the failure the
+ * variable exists to prevent, and it would only be found in `pg_stat_activity` at 3am.
+ */
+function poolMaxFromEnv(): Partial<PoolProfile> {
+  const raw = process.env[POOL_MAX_ENV];
+  if (raw === undefined || raw.trim() === '') return {};
+  const max = Number(raw);
+  if (!Number.isSafeInteger(max) || max < 1) throw poolMaxInvalid(raw);
+  return { max };
+}
+
 /** One connection pinned out of `Bun.SQL`'s pool, released back by hand. */
 interface BunSqlReserved {
   unsafe(text: string, values?: readonly unknown[]): Promise<unknown>;
@@ -103,10 +184,57 @@ function rowsOf<T>(result: unknown): readonly T[] {
   return Array.isArray(result) ? (result as readonly T[]) : [];
 }
 
+// The command tag only when it counted something, exactly like `rowsOf` in `pglite.ts` — one rule
+// across both drivers, so `execute()` and the observer's event cannot answer differently for the
+// same statement depending on which database is behind them. A driver that tags a read `0` while
+// returning rows would otherwise report 0 here and the row count there.
 function affectedBy(result: unknown): number {
   if (!Array.isArray(result)) return 0;
   const count = (result as { count?: unknown }).count;
-  return typeof count === 'number' ? count : result.length;
+  return typeof count === 'number' && count > 0 ? count : result.length;
+}
+
+/**
+ * `pool.reserve()` under a deadline. Without one an exhausted pool does not fail, it **queues** —
+ * so a slow endpoint filling all 20 slots turns every later request, `/readyz`'s `select 1`
+ * included, into a wait with no end and no error, and the pod is killed for being unready rather
+ * than answering 503 for the requests it cannot serve.
+ *
+ * The losing reservation is released, never dropped: the pool hands out a connection whenever one
+ * frees, deadline or no deadline, and a pin nobody holds is a connection nobody gets back. That is
+ * the whole reason this is not a bare `Promise.race`.
+ */
+async function reserveWithin(
+  pool: Pick<BunSqlDriver, 'reserve'>,
+  profile: PoolProfile,
+): Promise<BunSqlReserved> {
+  const budget = profile.acquireTimeoutMs;
+  if (budget <= 0) return pool.reserve();
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  let expired = false;
+  const pending = pool.reserve();
+  try {
+    return await Promise.race([
+      pending,
+      new Promise<never>((_resolve, reject) => {
+        timer = setTimeout(() => {
+          expired = true;
+          reject(poolAcquireTimeout(budget, profile.max));
+        }, budget);
+        // The deadline must not be what keeps a finished process alive.
+        timer.unref?.();
+      }),
+    ]);
+  } finally {
+    if (timer !== undefined) clearTimeout(timer);
+    // Attached unconditionally so a rejection arriving after we gave up is handled, not unhandled.
+    void pending.then(
+      (late) => {
+        if (expired) late.release();
+      },
+      () => undefined,
+    );
+  }
 }
 
 export interface PostgresClient extends ReservableClient {
@@ -129,17 +257,76 @@ export function createPostgresClient(options: PostgresClientOptions = {}): Postg
     return driver;
   }
 
-  async function runOn(
+  /** The send itself: one statement on one handle, every driver failure typed on the way out. */
+  async function sendOn(
     driver: Pick<BunSqlDriver, 'unsafe'>,
     fragment: SqlFragment,
   ): Promise<unknown> {
     try {
       return await driver.unsafe(fragment.text, fragment.values);
     } catch (error) {
-      throw dbUnavailable(`statement failed: ${fragment.text.slice(0, 120)}`, error);
+      // `driverError`, not `dbUnavailable`: the SQLSTATE has always been on this error and nothing
+      // read it, so a `23505` from two clicks racing a signup told the operator the database was
+      // unreachable and paged on-call for an outage that never happened. Everything the table does
+      // not classify is still `X_DB_UNAVAILABLE`, byte for byte.
+      throw driverError(`statement failed: ${fragment.text.slice(0, 120)}`, error);
     }
   }
 
+  /**
+   * The funnel — pooled and pinned statements both arrive here, which is why the observer hangs
+   * off this one function and nowhere else. Uninstalled it costs one property read and one
+   * branch: no clock read, no span, no event object, and `sendOn` receives exactly the call `runOn`
+   * made before the seam existed (axiom 6).
+   */
+  async function runOn(
+    driver: Pick<BunSqlDriver, 'unsafe'>,
+    fragment: SqlFragment,
+  ): Promise<unknown> {
+    const observer = statementObserver();
+    if (observer === undefined) return sendOn(driver, fragment);
+    // Read here, not by the consumer: the scope is gone by the time a per-request detector judges
+    // what it collected, so the reason has to be captured with the statement it defends.
+    const expected = expectedQueryLoopReason();
+    // Same moment, same argument: `postgresRepo` is several frames and a microtask above this one,
+    // and what it knows — the entity and the operation — is what turns fifty identical `select`s
+    // into "50× findById on members". Absent for hand-written SQL, a migration, a health probe.
+    const attribution = statementAttribution();
+    const started = performance.now();
+    let result: unknown;
+    try {
+      // The span wraps the send and nothing else, so its duration is the statement's and the
+      // observer's own work is not charged to the database.
+      result = await withStatementSpan(fragment.text, () => sendOn(driver, fragment));
+    } catch (error) {
+      // A statement that failed is still a statement: fifty identical timeouts are an N+1 of
+      // timeouts. The error is already `X_DB_UNAVAILABLE`, so the event carries what the caller
+      // is about to be thrown — and an observer that throws here replaces it, which is why
+      // `observe.ts` says a reporting-only observer must not throw.
+      observer.onStatement({
+        text: fragment.text,
+        values: fragment.values,
+        durationMs: performance.now() - started,
+        rows: 0,
+        error,
+        attribution,
+        expected,
+      });
+      throw error;
+    }
+    // Outside the `try` deliberately: a throw from `onStatement` is the observer's, not the
+    // database's, and catching it above would report a statement that succeeded as failed.
+    observer.onStatement({
+      text: fragment.text,
+      values: fragment.values,
+      durationMs: performance.now() - started,
+      rows: affectedBy(result),
+      attribution,
+      expected,
+    });
+    return result;
+  }
+
   async function run(fragment: SqlFragment): Promise<unknown> {
     return runOn(connect(), fragment);
   }
@@ -164,30 +351,52 @@ export function createPostgresClient(options: PostgresClientOptions = {}): Postg
       const pool = connect();
       let reserved: BunSqlReserved;
       try {
-        reserved = await pool.reserve();
+        reserved = await reserveWithin(pool, profile);
       } catch (error) {
         // Acquiring the pin is the one step that runs outside `runOn`, so an exhausted or
         // unreachable pool would escape as an untyped driver error — and `readOnlyQuery` reaches
         // this line before its first statement, which is how MCP ends up returning something
-        // other than X_DB_UNAVAILABLE.
-        throw dbUnavailable('could not reserve a connection from the pool', error);
+        // other than X_DB_UNAVAILABLE. Our own deadline is already typed; only the driver's own
+        // failure needs classifying, and `53300` from the server lands as X_DB_POOL_EXHAUSTED too.
+        if (error instanceof DbError) throw error;
+        throw driverError('could not reserve a connection from the pool', error);
       }
+      let held = true;
+      // Direct only while the pin is held. `release()` hands this physical connection back, and
+      // the pool may already have given it to another unit of work mid-transaction — a statement
+      // issued on the stale handle would land inside theirs, committed or rolled back with it and
+      // no error anywhere to explain the row. So a late statement takes its own connection out of
+      // the pool, exactly like any other caller. Same rule as `pglite.ts`, one driver down.
+      const on = (fragment: SqlFragment): Promise<unknown> =>
+        held ? runOn(reserved, fragment) : run(fragment);
+      // Idempotent because two owners already exist on one exit path: `withTransaction` releases
+      // in a `finally` and `[Symbol.dispose]` fires on the same scope. A second `release()` on a
+      // handle already back in the pool frees whoever holds that connection now.
+      const release = (): void => {
+        if (!held) return;
+        held = false;
+        reserved.release();
+      };
       return {
-        query: async <T>(fragment: SqlFragment) => rowsOf<T>(await runOn(reserved, fragment)),
-        one: async <T>(fragment: SqlFragment) =>
-          rowsOf<T>(await runOn(reserved, fragment))[0] ?? null,
-        execute: async (fragment: SqlFragment) => affectedBy(await runOn(reserved, fragment)),
-        release: () => {
-          reserved.release();
-        },
+        query: async <T>(fragment: SqlFragment) => rowsOf<T>(await on(fragment)),
+        one: async <T>(fragment: SqlFragment) => rowsOf<T>(await on(fragment))[0] ?? null,
+        execute: async (fragment: SqlFragment) => affectedBy(await on(fragment)),
+        release,
+        [Symbol.dispose]: release,
       };
     },
     async ping(): Promise<void> {
       await client.query(sql`select 1`);
     },
     async close(): Promise<void> {
-      await driver?.close();
+      // Read-then-clear, the same shape as `pglite.ts`: a `close()` that rejects has still torn
+      // the pool down, so caching it would hand the next `connect()` a corpse and every statement
+      // after it would fail for a reason no caller can see. Clearing first also means a
+      // `connect()` racing the await opens a fresh pool instead of joining the one draining. The
+      // rejection still reaches the caller — a shutdown that could not drain wants to know.
+      const pool = driver;
       driver = undefined;
+      await pool?.close();
     },
   };
   return client;
@@ -200,9 +409,16 @@ export function setDbClient(client: DbClient | undefined): void {
   ambient = client;
 }
 
-/** The pool, ignoring any open transaction. `withTransaction` must not re-enter `db()`. */
+/**
+ * The pool, ignoring any open transaction. `withTransaction` must not re-enter `db()`.
+ *
+ * The role default is layered under `DATABASE_POOL_MAX`, because this is the one place the process
+ * builds its own client and therefore the only place an operator's value can reach one:
+ * `createPostgresClient` has always taken a `profile` override and nothing in a running app passed
+ * it, so `POOL_PROFILES` was the last word in a deployed image.
+ */
 export function baseClient(): DbClient {
-  if (ambient === undefined) ambient = createPostgresClient();
+  if (ambient === undefined) ambient = createPostgresClient({ profile: poolMaxFromEnv() });
   return ambient;
 }
 

package/src/destructive.ts

@@ -0,0 +1,126 @@
+// Single responsibility: decide whether a migration's `up` half destroys data, and whether the
+// file admits it. The strong-migrations idea, enforced rather than documented — a drop is allowed,
+// an *undeclared* drop is not. Only `up` is ever asked: reversing `create table` is `drop table`,
+// so a rail reading `down` would mark every migration ever generated, and a mark on all is none.
+
+import { stripSqlNoise } from './sql-noise';
+import { noiseAt } from './sql-scan';
+import { statementsOf } from './statement-split';
+
+/** The line a migration carries to declare that applying it destroys data. */
+export const DESTRUCTIVE_MARKER = '-- destructive: true';
+
+/**
+ * The whole comment, so a file that merely *mentions* the marker — `-- destructive: true is
+ * required for a drop` — has not declared anything. `\r` is allowed because an editor writing
+ * CRLF must not turn a declared drop back into a gate failure.
+ */
+const MARKER_COMMENT = /^--[ \t]*destructive:[ \t]*true[ \t]*\r?\n?$/i;
+
+/** Whether only spaces and tabs separate `index` from the start of its line. */
+function startsLine(sql: string, index: number): boolean {
+  for (let at = index - 1; at >= 0; at -= 1) {
+    const char = sql[at];
+    if (char === '\n') return true;
+    if (char !== ' ' && char !== '\t') return false;
+  }
+  return true;
+}
+
+/**
+ * Declared only where a reader would see it: a top-level `--` line of its own.
+ *
+ * A regex over the raw file matched the marker inside `/* … *\/` and inside a dollar-quoted body,
+ * where it declares nothing and no reviewer reading the diff would call it a declaration — yet it
+ * bought the file past `x verify`. Scanning is what tells a comment from a comment about a
+ * comment; the marker is a lexical fact, not a substring.
+ */
+export function hasDestructiveMarker(sql: string): boolean {
+  let index = 0;
+  while (index < sql.length) {
+    const noise = noiseAt(sql, index);
+    if (noise === null) {
+      index += 1;
+      continue;
+    }
+    const text = sql.slice(index, noise.end);
+    if (noise.kind === 'line-comment' && startsLine(sql, index) && MARKER_COMMENT.test(text)) {
+      return true;
+    }
+    index = noise.end;
+  }
+  return false;
+}
+
+export type DestructiveKind = 'drop-table' | 'drop-column' | 'retype-column' | 'truncate';
+
+/** What each kind does to the rows, in the words `X_MIGRATION_DESTRUCTIVE` prints. */
+export const DESTRUCTIVE_CAUSE: Readonly<Record<DestructiveKind, string>> = {
+  'drop-table': 'drops a table',
+  'drop-column': 'drops a column',
+  'retype-column': 'rewrites a column to another type',
+  truncate: 'truncates a table',
+};
+
+export interface DestructiveStatement {
+  readonly kind: DestructiveKind;
+  /** The statement as written, on one line, without the comments that preceded it. */
+  readonly statement: string;
+}
+
+/**
+ * The closed list. Four kinds, because a rail that tried to enumerate every Postgres foot-gun
+ * would be a second SQL parser competing with the server's own — and every one of these is a
+ * statement `generateMigration` can emit, so each has a generated case to hold it honest.
+ *
+ * First match wins: one statement is one operation, and a second finding on the same line would
+ * repeat an instruction that is already one marker for the whole file.
+ */
+const RULES: readonly (readonly [DestructiveKind, RegExp])[] = [
+  ['drop-table', /\bdrop\s+(?:foreign\s+)?table\b/],
+  ['truncate', /\btruncate\b/],
+  // Inside an `alter table`, a bare `drop <name>` is a column: every sub-clause that drops
+  // something the database can rebuild names itself, and all of them are listed here.
+  [
+    'drop-column',
+    /\balter\s+table\b[\s\S]*?\bdrop\s+(?!constraint\b|default\b|not\b|identity\b|expression\b|generated\b)/,
+  ],
+  // Not "narrowing". Whether the new type is narrower is knowable only against the old one, and a
+  // rewrite that fails on one row fails the whole migration whichever direction it went.
+  ['retype-column', /\balter\s+column\b[\s\S]*?\btype\b/],
+];
+
+/**
+ * One capped line — an error prints this, not a whole script. Only the comments *preceding* the
+ * statement come off, the ones `statementsOf` carries in from the file header; the SQL itself stays
+ * verbatim, because `stripSqlNoise` blanks quoted identifiers and `drop table ""` names nothing an
+ * author can act on. Blanking is for deciding, never for reporting.
+ */
+function excerpt(statement: string): string {
+  const line = statement
+    .replace(/^(?:\s*(?:--[^\n]*|\/\*[\s\S]*?\*\/)\s*)+/, '')
+    .replace(/\s+/g, ' ')
+    .trim();
+  return line.length > 120 ? `${line.slice(0, 117)}...` : line;
+}
+
+/**
+ * Every destructive statement in `up`, in apply order.
+ *
+ * `statementsOf` cuts on a `;` that is not inside a literal, an identifier, a dollar-quoted body
+ * or a comment, and `stripSqlNoise` blanks all four before a keyword is looked for — so
+ * `-- drop table users` and `insert into audit values ('drop table users')` are prose and data,
+ * not operations. A naive keyword scan reports both, which is how a rail earns being ignored.
+ */
+export function destructiveStatements(up: string): readonly DestructiveStatement[] {
+  const found: DestructiveStatement[] = [];
+  for (const statement of statementsOf(up)) {
+    const bare = stripSqlNoise(statement).toLowerCase();
+    const rule = RULES.find(([, pattern]) => pattern.test(bare));
+    if (rule !== undefined) found.push({ kind: rule[0], statement: excerpt(statement) });
+  }
+  return found;
+}
+
+/** Whether `up` destroys data at all — what `x db gen` writes the marker from. */
+export const isDestructive = (up: string): boolean => destructiveStatements(up).length > 0;