Fighter90/career-ops-ui v1.227.5

on GitHub

[1.227.5] — 2026-08-29

Fixed — the regional scan ran the server out of memory, and a Playwright suite flaked on its own teardown.

Fixed

• The regional scan OOM-killed the server. ru-scanner accumulated every raw hit from every query and only deduplicated and filtered at the very end. The query list is deliberately full of near-synonyms (Golang, Go разработчик, Golang разработчик, Senior Go…), so the same vacancy comes back once per query and the array held several times the unique count — with descriptions attached, which are the bulk of a job object and are almost all discarded by the filters anyway. Measured at 742 MB of heap on a real 21-query run. The server caps Node's heap at 490 MB (V8 sizes it from the box's 956 MB of RAM), so the scan crashed the service with FATAL ERROR: Reached heap limit — four times in one day. Dedup and filtering now happen per query, which took the same run to 177 MB, a 76% reduction. Reported counts are unchanged: the unique total is carried by a Set of URL strings, so Total found still means what it always did.
• net::ERR_SOCKET_NOT_CONNECTED failed CI at random. Three Playwright suites call window.stop() to halt in-flight requests before asserting no console errors fired. Chromium reports that single deliberate act under three names depending on where the socket was: ERR_ABORTED when it cancels the request, ERR_EMPTY_RESPONSE mid-flight, and ERR_SOCKET_NOT_CONNECTED when the socket was already gone. The shared benign-console filter listed only the first two, so the same teardown failed a run at random. It surfaced on the locale sweep in tr — the suite with the most navigations, hence the most chances to lose the race — while playwright-forms and playwright-smoke, which also call window.stop(), were one unlucky run away from the identical failure.

Notes

• Why the scan only started failing now. The query list grew from 14 to 21 when product-manager searches were added. The old implementation's memory scaled with total hits, so +50% queries pushed it past the box's ceiling. Nothing was wrong with the new queries — they exposed an allocation pattern that had been one config change away from failing for a long time.
• The dedup change is behaviour-preserving, and that is tested rather than asserted. tests/ru-scanner-dedup-order.test.mjs pins the awkward direction — a later duplicate that FAILS the filters must evict an earlier one that passed, because the old end-of-run pass kept only the last copy and then filtered it — and runs a 200-trial randomised equivalence check against the legacy accumulate-then-filter implementation.
• The console filter stays narrow. Only the cancelled-request family is benign. A refused connection, DNS failure, connection reset, any 5xx and every uncaught JS exception still fail the assertion, pinned by new tests so the widening cannot creep.
• Two earlier hypotheses were tested and discarded rather than shipped: a Caddy timeout (its config is clean, flush_interval -1 is correct for SSE) and Node's 5-minute default requestTimeout (a scaled-down reproduction showed it does not terminate a long streaming response). The journal's Reached heap limit was the actual evidence. Test suite: 2860 → 2865.