[fix](temp-table) not clean temp table temporary until fix mem leak (#59535) - #93

Requesting changes for one mixed-version correctness regression that is outside the PR's accepted scope.

The PR description explicitly accepts disabling orphan temporary-table cleanup after all relevant FEs run the new behavior. It does not accept deleting a table owned by a still-live session during a rolling upgrade. There is no deployment/traffic constraint in the description, and the session-report RPC carries only cluster/token (no capability/version handshake), so an old-master/new-follower window is possible and unguarded.

Part 1.3 review conclusions:

1. Goal / proof: The steady-state all-new-version path does stop aliveSessionSet / sessionReportTimeMap growth and matches upstream d172147f55e in the two production files. However, the added test only skips the orphan-cleanup case; no mixed-version test proves upgrade safety.
2. Scope: The patch is small and focused, but the transition semantics are not contained to the described orphan-cleanup behavior.
3. Concurrency: No new local lock/deadlock issue. The correctness failure is cross-FE daemon concurrency: an old master's 5s FESessionMgr polling and 30s TemporaryTableMgr cycle observe a new follower's empty report.
4. Lifecycle: New follower sessions remain live, but their IDs are never registered/reported; the old master therefore ages out their temporary tables. This lifecycle mismatch is the blocker.
5. Configuration: No new config. Existing defaults make the issue concrete: alive_session_update_interval_second=5, loss_conn_fe_temp_table_keep_second=60, and TemporaryTableMgr uses the 30s default daemon interval.
6. Compatibility: No storage/RPC schema change, but rolling-upgrade behavioral compatibility is broken in the old-master/new-follower direction. New-master/old-follower is safe because cleanup is disabled on the master.
7. Parallel paths: Master-local/all-new paths were checked; only the mixed direction above retains the destructive old cleanup while suppressing the new follower's liveness report.
8. Conditions: Old cleanup's containsKey / timestamp checks have no FE-version or tracking-enabled guard.
9. Tests: Existing normal-session tests remain, but the restart test is skipped and there is no mixed-version coverage.
10. Observability: The old master logs the drop, but logging cannot prevent or recover the active temporary-table deletion.
11. Transaction / persistence: The erroneous cleanup calls dropTableWithoutCheck, and that metadata deletion is then propagated; this is not a transient read-only inconsistency.
12. Data writes: Active temporary-table data can be removed while its client session is still using it.
13. FE/BE variables: None added; not applicable.
14. Performance: The intended all-new steady-state allocation reduction is valid; no additional hot-path regression found.
15. Other: Apart from the inline P1, I found no other P0–P3 issues.

Verification performed: reviewed the complete base-to-head diff, final merge result, relevant ConnectContext / forwarding / FESessionMgr / cleanup call chains, and the upstream merged commit; git diff --check passed. I did not run FE build/UT/regression locally because this isolated worktree lacks .worktree_initialized, the referenced setup hook, and thirdparty/installed/bin/protoc. The PR's FE compile claim is author-provided and was not independently reproduced; no CI checks were exposed for the current head.

LGTM on b05ef22f775dd0bb730e005a0c23723e6fd6adb7. The previous mixed-version P1 is fixed by code rather than by a documentation-only assertion.

Scope note: the current top-level PR description still explicitly accepts only disabling orphan temporary-table cleanup; it does not make active-table deletion during rolling upgrade accepted or out-of-scope. The previous P1 therefore remained in scope. The new commit resolves it by reporting real live connections while leaving the leaking bookkeeping and new-master cleanup disabled.

Part 1.3 review conclusions:

1. Goal / proof: The original all-new-version memory fix remains intact, and old-master/new-follower compatibility is restored. EnvTest#testGetAllAliveSessionIds covers registered session inclusion and removal; there is no full mixed-version cluster test.
2. Scope: The follow-up is focused: one production method plus one targeted FE unit test.
3. Concurrency: Both underlying connection maps are concurrent. MySQL and Arrow Flight share ConnectScheduler.nextConnectionId, so combining their maps does not overwrite live entries. Registration occurs before query processing and unregistration on close; a concurrent snapshot can at most transiently lag a connection transition and introduces no lock/deadlock issue.
4. Lifecycle: Real client contexts live in the scheduler registry for exactly their connection/session lifetime. Forwarded proxy and internal contexts are excluded, preventing the original random-UUID leak.
5. Configuration: No config is added or changed. Existing 5s reporting, 60s keep time, and 30s cleanup cadence are now safe in the mixed-version direction because the active session continues to be reported.
6. Compatibility: No RPC/storage format changes are required. The existing getAliveSessions response now preserves its old semantic contract for an old master, so rolling upgrade is compatible in both old-master/new-follower and new-master/old-follower directions.
7. Parallel paths: MySQL and Arrow Flight SQL pools are both included. Master-local, follower RPC, normal disconnect, forwarded proxy, and internal-context paths were checked.
8. Conditions: No new special condition is introduced; the old master's existing checkAndRefreshSession timestamp condition now receives the correct session ID.
9. Tests: The added unit test verifies the new registry-derived result and removal behavior. The skipped restart test still corresponds to the PR's explicitly accepted orphan-cleanup disablement. Mixed-version end-to-end coverage is absent but not blocking for this focused compatibility bridge.
10. Observability: Existing FE session/temporary-table logs remain sufficient; no new failure state or metric is required.
11. Transaction / persistence: No EditLog format or replay path changes. The fix prevents an old master from reaching the erroneous persistent drop path for a live session.
12. Data writes: The new method is read-only over the connection registry and does not alter transaction or table-write semantics.
13. FE/BE variables: None added; not applicable.
14. Performance: Reporting is bounded by configured real connections instead of the unbounded leaked set. The combined-map/list snapshot is proportional to live clients and runs on the existing reporting path.
15. Other: No new P0-P3 findings.

Verification performed personally: re-read the current full PR body and all commits/files/reviews/comments/checks, reviewed the old-to-new diff and the updated full base-to-merge diff, inspected final files plus MySQL/Arrow registration and cleanup call chains, and ran git diff --check successfully. I did not run FE build/UT/regression locally because this isolated worktree has no .worktree_initialized, no referenced setup hook, and no thirdparty/installed/bin/protoc. The new commit message reports EnvTest#testGetAllAliveSessionIds and ./build.sh --fe; those are author-provided claims, not independently reproduced here. The Clang Formatter and License Check failures occur at actions/checkout because the fork PR is blocked in pull_request_target, before project code runs; Check Large File passed.

当前 head 存在一个需要在合入稳定版前处理的 P1：通过 follower 创建临时表时，实际表名只登记在 master 的代理 ConnectContext；原始 follower ConnectContext 没有该记录。停用 TemporaryTableMgr 后，正常断开、连接超时或正常 kill 都可能永久遗留临时表，形成无界的元数据、tablet 和存储资源泄漏。详细调用链和修复建议见内联评论及补充复核评论。按要求未运行 FE UT，本结论基于静态调用链检查。

基于最新 head 5d84509 重新检查：上次 P1 的普通 CREATE TEMPORARY TABLE 路径已得到修复，但其他受支持的临时表创建方式和转发部分成功边界尚未覆盖；新增 timeout 回归也不能独立证明服务端超时路径。详细意见见内联评论。按此前约束，本轮仅做静态检查，未运行 FE UT。

基于最新 head 5d8450956681fb089572f8d5c17d7c2545fe906d 重新审查，结论：APPROVE，未发现临时表功能之外的 P0–P3 问题。

本轮范围已按当前决策调整：临时表功能语义、创建路径覆盖、会话退出清理以及相关回归测试充分性不再作为本 PR 的阻塞依据，将在其他 PR 单独处置。

关键检查结论：

1. 目标与实现：Env.registerSessionInfo() 不再向 aliveSessionSet 写入，直接阻断普通客户端、follower 转发代理和内部任务创建 ConnectContext 时随机 session UUID 的无界积累。该核心修改与上游 apache#59535 一致，且不依赖生产是否实际使用临时表语法。
2. 资源与并发：移除写入后集合保持为空；既有 unregisterSessionInfo() 对空并发集合执行 remove 不产生正确性或并发风险。没有新增锁、线程或资源所有权关系。
3. 兼容性：没有 EditLog、image、RPC/Thrift、FE-BE 协议、配置或持久化结构变化；滚动升级期间新旧 FE 的非临时表查询、连接和会话协议不受影响。
4. 性能：核心效果是消除随 ConnectContext 构造次数无界增长的 session 集合及其周期性复制成本；未发现新的非临时表热路径开销。
5. 错误处理与可观测性：本修改不改变客户端查询错误语义，也没有新增需要处理的异常边界。
6. 测试与静态验证：精确 merge-base 6b84416279527cab64114d098e3fdeddf891c391 到当前 head 的最小 diff 为 7 个文件，git diff --check 通过。本轮未运行 FE UT。当前 Clang Formatter/License Check 失败发生在 fork PR 的 pull_request_target checkout 阶段，项目代码未被检出执行，因此不是代码检查失败。

历史上由临时表语义产生的我方阻塞 review 已撤销，相关评论已删除；本 review 取代此前结论。