mirror of
https://github.com/Monkopedia/rouse-context.git
synced 2026-11-01 20:20:25 +00:00
a74fb962c1
* test(tunnel): de-flake MultiClientConcurrencyTest (#501) The relay-backed multi-client test was flaky and hang-prone in CI: it failed with a SocketTimeoutException on one PR and wedged for ~35 min on another, burning the "Integration tests (relay)" job on unrelated changes. Two root causes, both test-timing (relay passthrough is fine in prod): 1. The class-level `@Timeout(180s)` runs in the default SAME_THREAD mode, which only checks elapsed time *after* the test method returns. A thread blocked in a `Socket` `read()` is never interrupted, so a wedged run hangs indefinitely instead of failing. Add a method-level `@Timeout(120s, threadMode = SEPARATE_THREAD)` so a stuck run is preempted and FAILS FAST (~2 min) instead of hanging the CI job. 2. The phase transitions waited on fixed `delay(...)`s ("let the stream register", "let handlers spin up", "let the CLOSE propagate") that race a wall clock and the 30s socket read timeout under load. Replace them with deterministic readiness signals: the device-side session collector now emits on per-stream open/close Channels, and the test awaits the actual lifecycle events (stream surfaced / handler returned) with bounded `withTimeout`. The test now tracks real events, so it can neither flake-timeout nor wedge. Test-only change; no production tunnel/relay behavior touched. Verified green and ~2.3s across 5 repeated runs, plus the full :core:tunnel integration suite (50 tests) with no regression. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * De-flake relay integration suite: shared IntegrationHttpSupport helper + SEPARATE_THREAD timeouts Extract the duplicated socket-HTTP request/response helper (the readLine()-racing-soTimeout source of the SocketTimeoutException flakes) into a shared core/testfixtures IntegrationHttpSupport, and convert ClaudeFullFlowEndToEndTest, OAuthEndToEndTest, TunnelMcpIntegrationTest, EndToEndSessionTest, and MultiClientConcurrencyTest to it. Robust read backed by a SEPARATE_THREAD hard @Timeout ceiling rather than racing per-read soTimeout. Test infra only. #501 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * De-flake: rely on generous socket read timeout, drop broad SEPARATE_THREAD The class-level @Timeout(threadMode = SEPARATE_THREAD) added across the end-to-end integration classes intermittently corrupted Gradle's per-test output store ("Could not write XML test results ... EOFException / Kryo buffer underflow"): these tests run long-lived background coroutines whose logging races the per-method separate-thread boundaries. Revert to the default SAME_THREAD class timeout and rely on the generous shared socket read timeout (IntegrationHttpSupport.SOCKET_READ_TIMEOUT_MS, 60s) as the fail-fast mechanism: every blocking read is bounded, so a stuck relay round trip raises SocketTimeoutException in ~60s instead of hanging, while normal CI slowness never trips it. MultiClientConcurrencyTest keeps its method-level SEPARATE_THREAD ceiling on its single test (#504), where the output-store interaction does not arise. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Jason Monk <monkopedia@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| bridge | ||
| mcp | ||
| testfixtures | ||
| tunnel | ||