perf(config): batch + cache engine.getConfig to kill ~85 round-trips per query

Takeover of #1694: a single search fires ~85 serial getConfig() reads
(loadConfigWithEngine x2 plus the mode/cache/intent/rerank/graph-signals
resolvers); on a remote pooler each read is a round-trip, dominating query
latency and risking cli.ts's 10s disconnect force-exit truncating stdout.

The first read now batch-loads the whole config table into a process-
lifetime Map (single-flight under concurrency); setConfig/unsetConfig
write through; a 30s TTL bounds multi-writer staleness and
GBRAIN_CONFIG_CACHE_TTL_MS=0 restores per-key reads.

Rebased onto current master: unlike the original diff, both the batch
load and the TTL=0 per-key fallback stay inside connRetry() so the
#1603/#1891 retry+reconnect posture (pooler-drop self-heal) is preserved.
PGLite stays uncached (in-process, zero round-trips; raw-SQL test
fixtures rely on fresh reads). E2E helpers pin the cache off since those
suites seed config via raw SQL.

Co-authored-by: Omerbahari <Omerbahari@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/core/postgres-engine.ts

@@ -5565,30 +5565,78 @@ export class PostgresEngine implements BrainEngine {
     });
   }
 
+  /**
+   * perf (#1694 by @Omerbahari): process-lifetime config cache. A single
+   * search fires ~85 getConfig() reads (loadConfigWithEngine x2, plus
+   * mode/cache/intent/rerank/graph-signals resolvers). On a remote pooler
+   * each read is a round-trip; serial they dominate query latency and can
+   * push the op handler past cli.ts's 10s disconnect force-exit, truncating
+   * stdout. First read batch-loads the whole `config` table into this Map
+   * (inside the same connRetry posture as the per-key read — #1603/#1891);
+   * setConfig/unsetConfig write through. TTL bounds staleness for
+   * multi-writer processes; GBRAIN_CONFIG_CACHE_TTL_MS=0 disables.
+   * Only present keys are stored — Map.has() distinguishes known-absent.
+   */
+  private _configCache: Map<string, string> | null = null;
+  private _configCacheLoadedAt = 0;
+  private _configCacheLoad: Promise<void> | null = null;
+  private get _configCacheTtlMs(): number {
+    const raw = process.env.GBRAIN_CONFIG_CACHE_TTL_MS;
+    if (raw !== undefined) {
+      const n = parseInt(raw, 10);
+      if (Number.isFinite(n) && n >= 0) return n;
+    }
+    return 30_000;
+  }
+
   async getConfig(key: string): Promise<string | null> {
     // #1603: a transient pooler drop on this read used to throw / fall through
     // to defaults silently — which on remote Postgres surfaces as the wrong
-    // search mode/knobs and empty-stdout queries.
-    return this.connRetry(async () => {
-      const rows = await this.sql`SELECT value FROM config WHERE key = ${key}`;
-      return rows.length > 0 ? (rows[0].value as string) : null;
-    });
+    // search mode/knobs and empty-stdout queries. Both the batch load and the
+    // cache-off per-key read keep the connRetry reconnect posture.
+    const ttl = this._configCacheTtlMs;
+    if (ttl === 0) {
+      return this.connRetry(async () => {
+        const rows = await this.sql`SELECT value FROM config WHERE key = ${key}`;
+        return rows.length > 0 ? (rows[0].value as string) : null;
+      });
+    }
+    if (this._configCache === null || Date.now() - this._configCacheLoadedAt >= ttl) {
+      // Single-flight: concurrent cold reads share one batch load.
+      this._configCacheLoad ??= this.connRetry(async () => {
+        const rows = await this.sql`SELECT key, value FROM config` as unknown as
+          Array<{ key: string; value: string | null }>;
+        const map = new Map<string, string>();
+        for (const r of rows) if (r.value != null) map.set(r.key, r.value);
+        this._configCache = map;
+        this._configCacheLoadedAt = Date.now();
+      }).finally(() => {
+        this._configCacheLoad = null;
+      });
+      await this._configCacheLoad;
+    }
+    return this._configCache!.has(key) ? this._configCache!.get(key)! : null;
   }
 
   async setConfig(key: string, value: string): Promise<void> {
-    return this.connRetry(async () => {
+    await this.connRetry(async () => {
       await this.sql`
         INSERT INTO config (key, value) VALUES (${key}, ${value})
         ON CONFLICT (key) DO UPDATE SET value = EXCLUDED.value
       `;
     });
+    // Write-through so a long-lived process never serves stale config.
+    this._configCache?.set(key, value);
   }
 
   async unsetConfig(key: string): Promise<number> {
-    return this.connRetry(async () => {
+    const count = await this.connRetry(async () => {
       const result = await this.sql`DELETE FROM config WHERE key = ${key}` as unknown as { count: number };
       return result.count ?? 0;
     });
+    // Write-through: known-absent, so the cache doesn't serve a stale value.
+    this._configCache?.delete(key);
+    return count;
   }
 
   async listConfigKeys(prefix: string): Promise<string[]> {

test/e2e/helpers.ts

@@ -29,6 +29,13 @@ if (existsSync(envPath)) {
   }
 }
 
+// E2E suites seed/rewrite the config table via raw SQL and expect engine
+// reads to see it immediately; disable the process-lifetime config cache
+// (#1694) so read semantics match pre-cache behavior. Spawned CLI
+// subprocesses inherit this. Cache semantics are pinned by
+// test/postgres-engine-config-cache.test.ts.
+process.env.GBRAIN_CONFIG_CACHE_TTL_MS ??= '0';
+
 const DATABASE_URL = process.env.DATABASE_URL;
 const FIXTURES_DIR = resolve(import.meta.dir, 'fixtures');
 

test/postgres-engine-config-cache.test.ts

@@ -0,0 +1,120 @@
+/**
+ * Process-lifetime config cache (#1694 takeover, by @Omerbahari).
+ *
+ * A single search fires ~85 getConfig() reads; on a remote pooler each is a
+ * round-trip. The first read now batch-loads the whole `config` table into a
+ * Map; setConfig/unsetConfig write through; TTL bounds multi-writer
+ * staleness; GBRAIN_CONFIG_CACHE_TTL_MS=0 restores per-key reads.
+ *
+ * Pure: stubs `_sql` with a call-counting fake; no real DB.
+ */
+
+import { describe, it, expect, beforeEach, afterEach } from 'bun:test';
+import { PostgresEngine } from '../src/core/postgres-engine.ts';
+
+const FAST_RETRY = { maxRetries: 3, delayMs: 1, delayMaxMs: 1, jitter: 'none' as const };
+
+/** Engine whose `sql` records every query's template strings and returns `rows`. */
+function makeEngine(rows: unknown[]) {
+  const e = new PostgresEngine();
+  const calls: string[] = [];
+  (e as unknown as { _connectionStyle: string })._connectionStyle = 'instance';
+  (e as unknown as { _bulkRetryOptsCache: unknown })._bulkRetryOptsCache = FAST_RETRY;
+  (e as unknown as { _sql: unknown })._sql = (strings: TemplateStringsArray) => {
+    calls.push(strings.join('?'));
+    return Promise.resolve(rows);
+  };
+  return { engine: e, calls };
+}
+
+const savedTtl = process.env.GBRAIN_CONFIG_CACHE_TTL_MS;
+
+beforeEach(() => {
+  delete process.env.GBRAIN_CONFIG_CACHE_TTL_MS;
+});
+
+afterEach(() => {
+  if (savedTtl === undefined) delete process.env.GBRAIN_CONFIG_CACHE_TTL_MS;
+  else process.env.GBRAIN_CONFIG_CACHE_TTL_MS = savedTtl;
+});
+
+describe('PostgresEngine config cache (#1694)', () => {
+  it('batch-loads once and serves repeat reads from the cache', async () => {
+    const { engine, calls } = makeEngine([
+      { key: 'search.mode', value: 'balanced' },
+      { key: 'embedding_multimodal', value: 'true' },
+    ]);
+    expect(await engine.getConfig('search.mode')).toBe('balanced');
+    expect(await engine.getConfig('embedding_multimodal')).toBe('true');
+    expect(await engine.getConfig('search.mode')).toBe('balanced');
+    // One SELECT total — this is the whole point of the fix.
+    expect(calls.length).toBe(1);
+    expect(calls[0]).toContain('SELECT key, value FROM config');
+  });
+
+  it('returns null for a known-absent key without an extra round-trip', async () => {
+    const { engine, calls } = makeEngine([{ key: 'a', value: '1' }]);
+    expect(await engine.getConfig('missing.key')).toBeNull();
+    expect(await engine.getConfig('missing.key')).toBeNull();
+    expect(calls.length).toBe(1);
+  });
+
+  it('setConfig writes through so subsequent reads see the new value', async () => {
+    const { engine, calls } = makeEngine([{ key: 'k', value: 'old' }]);
+    expect(await engine.getConfig('k')).toBe('old');
+    await engine.setConfig('k', 'new');
+    expect(await engine.getConfig('k')).toBe('new');
+    expect(calls.length).toBe(2); // batch load + upsert; no re-read
+  });
+
+  it('unsetConfig writes through so subsequent reads see absence', async () => {
+    const { engine } = makeEngine([{ key: 'k', value: 'v' }]);
+    expect(await engine.getConfig('k')).toBe('v');
+    await engine.unsetConfig('k');
+    expect(await engine.getConfig('k')).toBeNull();
+  });
+
+  it('concurrent cold reads share a single batch load (single-flight)', async () => {
+    const { engine, calls } = makeEngine([{ key: 'k', value: 'v' }]);
+    const [a, b, c] = await Promise.all([
+      engine.getConfig('k'),
+      engine.getConfig('k'),
+      engine.getConfig('other'),
+    ]);
+    expect([a, b, c]).toEqual(['v', 'v', null]);
+    expect(calls.length).toBe(1);
+  });
+
+  it('GBRAIN_CONFIG_CACHE_TTL_MS=0 disables the cache (per-key reads)', async () => {
+    process.env.GBRAIN_CONFIG_CACHE_TTL_MS = '0';
+    const { engine, calls } = makeEngine([{ value: 'v' }]);
+    expect(await engine.getConfig('k')).toBe('v');
+    expect(await engine.getConfig('k')).toBe('v');
+    expect(calls.length).toBe(2);
+    expect(calls[0]).toContain('SELECT value FROM config WHERE key =');
+  });
+
+  it('an expired TTL reloads from the database', async () => {
+    process.env.GBRAIN_CONFIG_CACHE_TTL_MS = '1';
+    const { engine, calls } = makeEngine([{ key: 'k', value: 'v' }]);
+    expect(await engine.getConfig('k')).toBe('v');
+    await new Promise((r) => setTimeout(r, 5));
+    expect(await engine.getConfig('k')).toBe('v');
+    expect(calls.length).toBe(2); // two batch loads
+  });
+
+  it('the batch load keeps the connRetry reconnect posture (#1603/#1891)', async () => {
+    const e = new PostgresEngine();
+    (e as unknown as { _connectionStyle: string })._connectionStyle = 'instance';
+    (e as unknown as { _sql: unknown })._sql = null; // torn-down pool → retryable
+    (e as unknown as { _bulkRetryOptsCache: unknown })._bulkRetryOptsCache = FAST_RETRY;
+    let reconnects = 0;
+    (e as unknown as { reconnect: () => Promise<void> }).reconnect = async () => {
+      reconnects++;
+      (e as unknown as { _sql: unknown })._sql = () =>
+        Promise.resolve([{ key: 'k', value: 'v' }]);
+    };
+    expect(await e.getConfig('k')).toBe('v');
+    expect(reconnects).toBe(1);
+  });
+});

test/postgres-engine-config-reconnect.test.ts

@@ -43,7 +43,9 @@ function makeTornDownEngine(poolResult: unknown): { engine: PostgresEngine; reco
 
 describe('PostgresEngine non-batch config accessors self-heal (PR #1891 takeover)', () => {
   it('getConfig reconnects + retries a null instance pool, then returns the value', async () => {
-    const { engine, reconnects } = makeTornDownEngine([{ value: 'live-value' }]);
+    // Rows carry `key` too: getConfig's default cached path batch-loads
+    // `SELECT key, value FROM config` (#1694) through the same connRetry.
+    const { engine, reconnects } = makeTornDownEngine([{ key: 'some.key', value: 'live-value' }]);
     expect(await engine.getConfig('some.key')).toBe('live-value');
     expect(reconnects()).toBe(1); // exactly one reconnect closed the gap
   });