Repository navigation
fix(process): attribute slow memory Worker requests - #2457
Conversation
50c05ca to
af51486
Compare
There was a problem hiding this comment.
Changes requested: The new request diagnostics can mark a stalled response as complete.
Warning
Changes requested · head af51486 · 1 finding: 1 minor
| Severity | Finding | Where |
|---|---|---|
| minor | F1 Keep request diagnostics open through response-body consumption | deploy/cloudflare-memory/testFetch.ts:8 |
Full review
F1 (minor, high confidence): fetchMemoryTest marks a request passed as soon as SELF.fetch returns, but the test helpers read res.text() or res.json() afterward. If a response body stalls, the case can time out while the diagnostic reports that route as completed and shows no pending request. Keep the operation open through body consumption, and cover that timeout in the fixture test.
af51486 to
4bcc876
Compare
There was a problem hiding this comment.
LGTM: The prior response-body attribution gap is closed; no remaining findings in the complete PR diff.
Note
Approved · head 4bcc876 · no findings
Full review
The prior F1 is resolved: the operation stays pending through response-body consumption, and a stalled-body regression test covers it. No further findings.
Memory Worker tests now allow cold Durable Object startup and record which request is slow or still pending when a case fails. Related to #2119; this does not close it.
Why: After #2440, a multi-request transcript case exceeded the 5-second case limit on #2444 (5.53 seconds), then passed on the same head. A 50-run Linux probe measured one first /runs/transcript/owner request at 1.82 seconds while the following seven requests were fast. The old limit hid which request consumed the time. That probe supports a cold-start explanation but does not prove the failed request had the same cause.
Where to look
posthelpers.Feedback wanted: Check the 15-second case bound and whether the request markers identify the stalled operation if the flake returns.
Risk: A genuinely stalled case can take 10 seconds longer to fail. This changes the test harness only; the production Worker is unchanged. #2119 stays open for live CI evidence.
Verified: 50/50 fresh Linux suite runs on five Depot runners at the final head, 50/50 on each of two prior candidates, and 70/70 baseline runs. All exact-head PR checks passed. Local named Worker tests: 225 passed, one expected failure; four diagnostic tests passed. Worker and root typechecks, consistency, spec coverage and formatting passed. Current PR-head CI is pending.
Decisions (1)
Validation (3 criteria)
4bcc8765a.