michaellx1057 opened a new pull request, #18645: URL: https://github.com/apache/dolphinscheduler/pull/18645
## Was this PR generated or assisted by AI? YES — Codex assisted with source analysis, implementation, regression tests and this description. Human review of the final diff is pending. No production deployment of this patch has been performed. ## Purpose of the pull request Closes #18644. Fix the remaining shared HA ownership issue. A restarted server with the same address must not accept its predecessor's selector, and a server that loses an election must not retain ACTIVE. This follows the identity-reuse analysis in #18624. The serial-coordinator restart fix #18623 and etcd REMOVE-value fix #18641 address different layers and are not replaced by this change. ## Brief change log - Give each HA server instance a unique election identity. - Reconcile the current selector on removal and serialize local election/status application. - Demote on unsuccessful election and release only acquired election locks. - Respect terminal HA shutdown in Master and Alert; prevent a retry or queued notification from reactivating a closed consumer. - Add shared HA and Alert lifecycle regression tests and document selector-value compatibility. ## Verify this pull request Executed on JDK 11 against dev `9839c418c1a6d7f2a8c3395552edcd99c2de1b37`: ```bash mvn -B -pl dolphinscheduler-alert/dolphinscheduler-alert-server,dolphinscheduler-master -am -Dtest=AbstractHAServerTest,AlertServerHATest -Dsurefire.failIfNoSpecifiedTests=false -Djacoco.skip=true -Drat.skip=true -Dcheckstyle.skip -Dos.detected.classifier=osx-x86_64 spotless:apply test mvn -B -pl dolphinscheduler-registry/dolphinscheduler-registry-api,dolphinscheduler-alert/dolphinscheduler-alert-server,dolphinscheduler-master spotless:check ``` A local settings file selected Maven Central. The classifier override uses the available x86_64 protoc binary via Rosetta on Apple Silicon; no dependencies were changed. The full local invocation is recorded in the verification receipt. - Final build: 132-module reactor completed successfully, including Java production/test compilation; 16 shared HA tests and two Alert lifecycle cases passed. Other modules' complete test suites were not run by this focused command. - Baseline: before adding terminal-close cases, 13 shared tests against the original HA source produced 11 assertion failures and two passing controls, including failure of the startup/REMOVE ordering assertion. - Independent final Spotless checks and `git diff --check` passed. No Checkstyle, RAT or coverage result is claimed from the recorded skip flags. - Real-backend retained-predecessor comparison: etcd 3.5.21, ZooKeeper 3.8.3, JDBC/MySQL 8.0.46 and JDBC/PostgreSQL 16.0. All eight original/final cases passed: original ACTIVE without owning the selector, fixed STAND_BY, with backend owner/revision or row metadata unchanged. Both JDBC backends additionally passed real removal-event/re-election checks, creating a new selector owned by the successor client. - These fixtures deliberately retain a predecessor client and seed its legacy bare-address selector. They verify common ownership behavior, not complete multi-Master deployments or identical downstream consequences. The ZooKeeper/etcd probes use mixed cached supporting dependencies; JDBC uses the compiled dev classes and native schemas. - Separate ZooKeeper/etcd close-from-watch probes passed ephemeral cleanup checks. - The repository wrapper also passed root plus all three changed Java modules: `./mvnw -B -pl .,dolphinscheduler-registry/dolphinscheduler-registry-api,dolphinscheduler-alert/dolphinscheduler-alert-server,dolphinscheduler-master spotless:apply spotless:check` (using the same local Maven Central settings). - Independent source review found no remaining blocking findings. Human review and upstream CI remain pending. ## Compatibility and limits Custom `AbstractHAServer` subclasses overriding `close()` must delegate to `super.close()` to preserve terminal shutdown. Alert shutdown now closes its Registry resources, including when heartbeat shutdown fails. Selector values become opaque instance identities rather than bare addresses. Registry paths and SPI are unchanged. External scripts parsing addresses must be adjusted. Unpatched peers in a mixed-version cluster retain their defects. This is not a distributed fencing mechanism and does not guarantee zero overlap during delayed watches or interrupted connectivity. ## 中文摘要 在共享 HA 层区分每次启动的协调器身份，防止新进程继承旧临时节点；收到删除通知后根据当前所有权更新角色，并把选举与本地状态应用按顺序处理。补充通用回归测试与 Alert 关闭测试。真实 etcd、ZK、JDBC（MySQL/PostgreSQL）的原版/补丁对照均确认共同的旧身份误认已被阻止；这不代表每种后端都复现了完整双 Master 后果。现有 etcd 删除事件 PR 保持独立；本补丁未部署生产。 ## Pull Request Notice [Pull Request Notice](https://github.com/apache/dolphinscheduler/blob/dev/docs/docs/en/contribute/join/pull-request.md) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
