## Why
The AWS SDK defaults `requestTimeout` to `0`, which means *no timeout* —
see `DEFAULT_REQUEST_TIMEOUT` in `@smithy/node-http-handler`. None of
our clients overrode it, so a request that never completes holds its
socket forever. Once that happens to `maxSockets` requests (default 50),
the client's connection pool is permanently exhausted and every
subsequent call on that process queues indefinitely rather than failing.
To the caller it is indistinguishable from a hang.
This caused a production incident on 2026-08-06. A single
`twenty-server` pod reached:
```
@smithy/node-http-handler:WARN - socket usage at capacity=50 and 1004 additional requests are enqueued.
```
and never recovered — the queue grew monotonically and the pod completed
**zero** Lambda invocations over 12 hours while its six siblings
completed dozens each. Egress was fine (a direct HTTPS call to the
Lambda API returned in 53ms) and the pod was never OOM-killed or
restarted, so nothing surfaced as an error anywhere.
The user-visible effect was that workspace creation hung. Activation
reached `synchronizeTwentyStandardApplicationOrThrow`, blocked on a
Lambda call, and never returned or threw — so the `catch` in
`activateWorkspace` that resets a workspace to `PENDING_CREATION` never
ran, and the workspace was stranded in `ONGOING_CREATION`. Retrying
didn't help because the pod was still poisoned. With one bad pod out of
seven, roughly one signup in seven failed:
| pod | activations started | completed |
|---|---|---|
| healthy × 5 | 13 | 13 |
| poisoned | 3 | **0** |
Nothing appeared in Sentry, because nothing ever threw.
## What this changes
- **Every AWS SDK client now sets `connectionTimeout` and
`requestTimeout`** via a shared `buildAwsRequestHandlerOptions()`
helper. A saturated pool now surfaces as an ordinary error the caller
can catch and retry instead of hanging forever. This is the fix that
matters.
- **The Lambda client gets a higher ceiling and a larger socket pool.**
Synchronous invocations legitimately hold a socket for as long as the
function runs, so its `requestTimeout` clears
`EXECUTOR_LAMBDA_TIMEOUT_SECONDS` (900s) with a minute to spare, and
`maxSockets` goes to 200 so long invocations cannot starve the
control-plane calls (layer lookups, waiters) sharing that client.
- **`S3Client` and `STSClient` in `LambdaAwsClientService` are now
reused.** Both were constructed on every call and never destroyed, so
each leaked its own agent and socket pool. They are invalidated
alongside `lambdaClient` when assume-role credentials refresh.
No new dependency: `requestHandler` already accepts a plain
`NodeHttpHandlerOptions` object.
## Deliberately not in scope
- **A timeout around `activateWorkspace` itself.** A hung activation
still strands a workspace in `ONGOING_CREATION` until the 5-minute
stale-lock reclaim, and that only fires if the user happens to retry.
Worth fixing separately.
- **Alerting on `socket usage at capacity`.** That warning was the only
signal this was happening and nobody was watching it — an infra change
rather than a code one.
## Testing
- Unit tests for the helper, including that the timeout is always
non-zero.
- `tsc --noEmit` clean on the touched files; `oxlint` reports 0 warnings
and 0 errors.
- Not reproducible in a test environment — the leak needs a saturated
pool — so the mechanism above is evidenced from production logs and the
SDK's own defaults rather than from a regression test.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/23857?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
Charles Bochet
f5a42cdbae
