The GitHub Actions job "Build" on rocketmq-flink.git/direct-connect-stability has succeeded. Run started by GitHub user lizhimins (triggered by lizhimins).
Head commit for run: 3d8b17c11ce120d3dab557f2b0581cea3be4ef98 / lizhimins <[email protected]> [ISSUE #151] fix(remoting): make split assignment and partition discovery survive direct-connect route races Two silent defects surface when the client talks to NameServers directly (open-source deployment, no proxy): 1. Starting offsets are ignored for a random subset of queues. RebalancePullImpl.computePullFromWhereWithException is hard-coded to 0, so assign() lets the client-internal rebalance thread dispatch a first pull from offset 0 while handleSplitsChanges is still seeking to the configured starting offset. No Flink-side lock scope can win that race. Enforce the starting offset deterministically at the fetch() outlet instead: records below max(startingOffset, 0) are dropped there, and seek() remains as a traffic optimization. 2. A NameServer outage restarts the job and rewinds offsets. The periodic partition discovery flattened per-topic route fetch failures into an empty queue set; diffed against the allocated set that revokes every split, and the queues returning after recovery are re-initialized from scratch (observed: +150 replayed records). A failed fetch says nothing about the actual route: exclude failed topics from the round, skip the round when every topic failed, and keep the current assignment when an empty result arrives while queues are allocated. Genuine deletions surface again on later rounds. Also document that EXACTLY_ONCE is not supported on proxy-fronted endpoints (transaction state lives in per-proxy-process memory and is silently dropped), while direct deployments do support it. Cluster verification of the outage scenario after the fix: job RUNNING throughout the 60s NameServer stop, zero split decreases, zero replay (sink grew exactly by the second wave), producer recovered within 150s of the NameServer coming back. Report URL: https://github.com/apache/rocketmq-flink/actions/runs/40826307202 With regards, GitHub Actions via GitBox
