root/fusion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

test(core): cap pg-gate fork fan-out for DB-bound suite

Browse Source 
The test:pg-gate suite runs only *.pg.test.ts files, each building/copying a
per-file schema-template database (heavy CREATE/DROP DATABASE DDL serialized by
the single shared Postgres). Worker count derived from CPU cores over-scales on
high-core machines (6 forks on a 28-core box), oversubscribing the one Postgres
until every beforeAll exceeds the 15s hookTimeout (23/23 hook timeouts). CI's
low-core runners stay near 2 forks and pass, so it only bites high-core locals.

Add a maxCap clamp to computeMaxWorkers and a dedicated vitest.pg.config.ts
(maxCap=4) for the pg-gate, right-sizing concurrency to the actual constraint (a
single shared Postgres) rather than raising the timeout (forbidden appeasement).
Low-core machines keep their smaller CPU-derived count via min(4, cpuCap).

Verified: full test:pg-gate now passes 23 files / 126 tests on a 28-core host.
This commit is contained in:
gsxdsm
2026-07-20 18:16:58 -07:00
parent 2884bf76b1
commit caf425eaea
4 changed files with 101 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
packages/core/package.json
View File

@@ -43,7 +43,7 @@
"typecheck": "tsc --noEmit",
"test": "vitest run --silent=passed-only --reporter=dot",
"test:embedded-postgres": "vitest run src/__tests__/postgres/embedded-lifecycle.test.ts --silent=passed-only --reporter=dot",
"test:pg-gate": "vitest run src/__tests__/postgres/handoff-to-review-atomicity.pg.test.ts src/__tests__/postgres/store-list.pg.test.ts src/__tests__/postgres/task-lifecycle-e2e.pg.test.ts src/__tests__/postgres/soft-delete-resurrection-FN-5233.pg.test.ts src/__tests__/postgres/agent-logs-and-monitor.pg.test.ts src/__tests__/postgres/todo-store.pg.test.ts src/__tests__/postgres/workflow-definitions.pg.test.ts src/__tests__/postgres/message-store.pg.test.ts src/__tests__/postgres/insight-store.pg.test.ts src/__tests__/postgres/insight-run-execution.pg.test.ts src/__tests__/postgres/research-store.pg.test.ts src/__tests__/postgres/mission-store.pg.test.ts src/__tests__/postgres/goal-store.pg.test.ts src/__tests__/postgres/artifacts-documents-evals.pg.test.ts src/__tests__/postgres/command-center-analytics.pg.test.ts src/__tests__/postgres/command-center-remaining-analytics.pg.test.ts src/__tests__/postgres/research-execution.pg.test.ts src/__tests__/postgres/async-store-events.pg.test.ts src/__tests__/postgres/signal-ingestion.pg.test.ts src/__tests__/postgres/mission-autopilot.pg.test.ts src/__tests__/postgres/workflow-create.pg.test.ts src/__tests__/postgres/monitor-trait-storm-guard.pg.test.ts src/__tests__/postgres/agent-wake-getagent.pg.test.ts --silent=passed-only --reporter=dot"
"test:pg-gate": "vitest run --config vitest.pg.config.ts src/__tests__/postgres/handoff-to-review-atomicity.pg.test.ts src/__tests__/postgres/store-list.pg.test.ts src/__tests__/postgres/task-lifecycle-e2e.pg.test.ts src/__tests__/postgres/soft-delete-resurrection-FN-5233.pg.test.ts src/__tests__/postgres/agent-logs-and-monitor.pg.test.ts src/__tests__/postgres/todo-store.pg.test.ts src/__tests__/postgres/workflow-definitions.pg.test.ts src/__tests__/postgres/message-store.pg.test.ts src/__tests__/postgres/insight-store.pg.test.ts src/__tests__/postgres/insight-run-execution.pg.test.ts src/__tests__/postgres/research-store.pg.test.ts src/__tests__/postgres/mission-store.pg.test.ts src/__tests__/postgres/goal-store.pg.test.ts src/__tests__/postgres/artifacts-documents-evals.pg.test.ts src/__tests__/postgres/command-center-analytics.pg.test.ts src/__tests__/postgres/command-center-remaining-analytics.pg.test.ts src/__tests__/postgres/research-execution.pg.test.ts src/__tests__/postgres/async-store-events.pg.test.ts src/__tests__/postgres/signal-ingestion.pg.test.ts src/__tests__/postgres/mission-autopilot.pg.test.ts src/__tests__/postgres/workflow-create.pg.test.ts src/__tests__/postgres/monitor-trait-storm-guard.pg.test.ts src/__tests__/postgres/agent-wake-getagent.pg.test.ts --silent=passed-only --reporter=dot"
},
"devDependencies": {
"@earendil-works/pi-coding-agent": "0.80.10",

23
packages/core/src/__test-utils__/vitest-workers.ts
View File

@@ -2,6 +2,22 @@ import { cpus } from "node:os";
interface ComputeMaxWorkersOptions {
defaultCap?: number;
/*
* FNXC:PgTestWorkerCap 2026-07-18-18:00:
* Hard upper clamp applied to the FINAL resolved worker count, AFTER every
* other resolution path (explicit VITEST_MAX_WORKERS, workspace budget, CPU
* default) and the cpuCap clamp. Purpose: a suite bound by a single shared
* external resource (the pg-gate's one Postgres server, which serializes
* CREATE/DROP DATABASE DDL) must NOT scale its fork fan-out with CPU count.
* On a 28-core dev box the CPU-derived default is 6 forks; 6 concurrent forks
* each running the DDL-heavy per-file schema-template build oversubscribe the
* one Postgres until every beforeAll exceeds the 15s hookTimeout (the whole
* pg-gate then reports 23/23 hook timeouts). CI's low-core runners stay near 2
* and pass, which is why this only bites high-core local machines. `maxCap`
* lets the pg vitest config pin a DB-safe ceiling while low-core machines keep
* using their smaller CPU-derived count. See vitest.pg.config.ts.
*/
maxCap?: number;
}
function computeDefaultCap(cpuCap: number): number {
@@ -26,7 +42,7 @@ function parsePositiveInt(value: string | undefined): number | undefined {
// All paths clamp to (cpus - 1) so we never oversubscribe.
export function computeMaxWorkers(options: ComputeMaxWorkersOptions = {}): number {
const cpuCap = Math.max(1, cpus().length - 1);
const { defaultCap = computeDefaultCap(cpuCap) } = options;
const { defaultCap = computeDefaultCap(cpuCap), maxCap } = options;
const explicit = parsePositiveInt(process.env.VITEST_MAX_WORKERS);
const totalBudget = parsePositiveInt(process.env.FUSION_TEST_TOTAL_WORKERS);
@@ -49,6 +65,11 @@ export function computeMaxWorkers(options: ComputeMaxWorkersOptions = {}): numbe
}
workers = Math.min(workers, cpuCap);
// DB-bound suites clamp below the CPU-derived count so fork fan-out never
// oversubscribes a single shared external resource (see maxCap docs).
if (maxCap !== undefined && maxCap > 0) {
workers = Math.max(1, Math.min(workers, maxCap));
}
process.env.VITEST_MAX_WORKERS = String(workers);
return workers;
}

38
packages/core/src/__tests__/vitest-workers.test.ts
View File

@@ -54,6 +54,44 @@ describe("computeMaxWorkers", () => {
expect(process.env.VITEST_MAX_WORKERS).toBe("2");
});
// FNXC:PgTestWorkerCap 2026-07-18-18:00: maxCap must clamp the FINAL count for
// DB-bound suites (pg-gate) so fork fan-out never oversubscribes one Postgres,
// and it must win even over an explicit VITEST_MAX_WORKERS above the cap.
it("clamps the final worker count to maxCap for DB-bound suites", () => {
delete process.env.VITEST_MAX_WORKERS;
delete process.env.FUSION_TEST_TOTAL_WORKERS;
delete process.env.FUSION_TEST_CONCURRENCY;
const cpuCap = Math.max(1, cpus().length - 1);
const workers = computeMaxWorkers({ defaultCap: 6, maxCap: 4 });
expect(workers).toBe(Math.min(4, cpuCap));
expect(process.env.VITEST_MAX_WORKERS).toBe(String(workers));
});
it("applies maxCap even when explicit VITEST_MAX_WORKERS exceeds it", () => {
process.env.VITEST_MAX_WORKERS = "12";
delete process.env.FUSION_TEST_TOTAL_WORKERS;
delete process.env.FUSION_TEST_CONCURRENCY;
const cpuCap = Math.max(1, cpus().length - 1);
const workers = computeMaxWorkers({ maxCap: 4 });
expect(workers).toBe(Math.min(4, cpuCap));
});
it("does not raise workers below maxCap on low-core machines", () => {
delete process.env.VITEST_MAX_WORKERS;
delete process.env.FUSION_TEST_TOTAL_WORKERS;
delete process.env.FUSION_TEST_CONCURRENCY;
// A defaultCap under the maxCap must be preserved (min semantics, not a floor).
const cpuCap = Math.max(1, cpus().length - 1);
const workers = computeMaxWorkers({ defaultCap: 2, maxCap: 4 });
expect(workers).toBe(Math.min(2, cpuCap));
});
it("uses a CPU-aware default cap when no overrides are provided", () => {
delete process.env.VITEST_MAX_WORKERS;
delete process.env.FUSION_TEST_TOTAL_WORKERS;

40
packages/core/vitest.pg.config.ts Normal file
View File

@@ -0,0 +1,40 @@
import { defineConfig, mergeConfig } from "vitest/config";
import baseConfig from "./vitest.config";
import { computeMaxWorkers } from "./src/__test-utils__/vitest-workers";
/*
FNXC:PgTestWorkerCap 2026-07-18-18:00:
Dedicated vitest config for the PostgreSQL gate suite (`test:pg-gate`). The
pg-gate runs ONLY *.pg.test.ts files, each of which builds and copies a
per-file schema-template database (heavy CREATE/DROP DATABASE DDL that the one
shared Postgres server serializes). The bottleneck is that single DB, not CPU,
so fork fan-out must not scale with core count.
The base core config derives its worker count from CPUs: on a 28-core dev box
that is 6 forks. Six concurrent forks racing the DDL-heavy per-file setup
oversubscribe the one Postgres until every `beforeAll` exceeds the 15s
hookTimeout, and the whole gate reports 23/23 hook timeouts. CI's low-core
runners land near 2 forks and pass, so the failure only bites high-core local
machines. Measured on this 28-core box: 6 forks -> all time out; 4 forks -> 23
files / 126 tests pass in ~42s; 2 forks -> pass in ~78s.
`maxCap: 4` pins a DB-safe ceiling: high-core machines clamp to 4 forks while
low-core machines keep their smaller CPU-derived count (min(4, cpuCap)). This
is NOT timeout appeasement — the hookTimeout is unchanged; we right-size
concurrency to the actual constraint (a single shared Postgres) per the FN-5048
rule against oversubscribing worker/concurrency knobs. Only the pg-gate uses
this config; the interleaved full core suite is unaffected because pg files are
a small fraction of it and rarely run 6-at-once.
*/
const PG_MAX_WORKERS = 4;
export default mergeConfig(
baseConfig,
defineConfig({
test: {
maxWorkers: computeMaxWorkers({ maxCap: PG_MAX_WORKERS }),
minWorkers: 1,
},
}),
);