Repository navigation
Conversation
### What problem does this PR solve? Related PR: apache#63612 Problem Summary: Follower and observer FEs can reuse stale SQL results and metadata-derived caches after applying table changes that do not advance data visible versions. Backport the shared metadata-change notification, per-cache failure isolation, and common SQL cache admission/lookup validation to this release branch's native cache and constraint interfaces. Preserve table-wide data version checks and fence older asynchronous cache builds. ### Release note Fix stale cached query results on follower and observer FEs after table metadata changes. ### Check List (For Author) - Test: Unit Test: 232 focused cache/replay and affected existing FE tests; FE build and Checkstyle. Multi-FE/BE regression not run locally. - Behavior changed: Yes, reject cached results whose metadata dependencies changed. - Does this need documentation: No (cherry picked from commit f066153)
|
run buildall |
|
/review |
|
Codex automated review failed and did not complete. Error: Review setup stopped before context preparation ran; inspect the first failed or cancelled setup step. Please inspect the workflow logs and rerun the review after the underlying issue is resolved. |
|
run buildall |
|
/review |
There was a problem hiding this comment.
Static review of all 36 changed paths at head 4c2a2b3 found two P2 issues in the new MV rewrite invalidation path: canonical external table names can be normalized differently from registered dependencies, and concurrent dependency re-registration can hide an MV from invalidation. Both are reported inline. No P0/P1 finding or existing blocking inline was confirmed.
Critical checkpoint conclusions:
- Goal and proof: the change routes existing metadata journals into local derived-cache invalidation, adds an in-flight SQL revision fence, and adds OLAP schema identity to backend cache keys. The two inline issues leave gaps in MV invalidation. Added replay/unit/follower regressions were inspected; they were not executed.
- Scope and clarity: changes are confined to FE cache wiring, dependency validation, metadata mutation/replay hooks, and related tests/fixture setup. No unrelated source change was identified.
- Concurrency and locks: query/cache admission, DDL/replay, MV refresh, and external metadata refresh were traced. SQL revision cells use atomics; table read locks retain ID ordering. MV dependency resolution and cache construction release MV locks before external lookup. No new direct deadlock was substantiated; the dependency-registration visibility race is reported inline.
- Lifecycle and initialization: snapshots retain weak-cache revision cells strongly, so active dependencies survive collection without permanent tombstones. Admission validation closes its temporary StatementContext and restores the executing context. The static ID-generator fallback no longer retains a connection. MV generation fencing is present, but only helps when invalidation discovers the MV, as described inline.
- Configuration: no new configuration was added. Existing dynamic SQL-cache replacement preserves the revision owner. Per-catalog versus global case configuration is the first inline finding.
- Compatibility: existing journals and persisted objects are reused; no storage or wire format change was introduced. The schema addition changes backend cache hashes, resulting in cold cache entries across differing FE versions rather than protocol incompatibility.
- Thrift/protobuf: no gensrc definitions were changed; cross-branch field-ID checks are not applicable.
- Parallel paths: reviewed internal/external catalogs, live/replay mutations, cloud/shared-nothing restore, FE/BE SQL admission, views/tables/namespaces, schema/rollup completion, constraints, and rename/swap/drop/recovery. Canonical identity and concurrent MV registration gaps are reported; no additional verified defect was found.
- Conditional behavior: checkpoint invalidation exclusion, temporary-partition behavior, and cloud restore's remote version ownership were checked against callers. Manual REFRESH TABLE's wider database invalidation masks the case issue; the inline example uses post-commit refresh and its full replay.
- Test coverage: inspected positive/negative dependency snapshots, schema keys, late puts, data versions, namespace isolation, failure fanout, serialized replay, prepared-context ownership, and latch-based MV race/lock tests. The existing tests do not cover the two inline scenarios. The follower regression skips single-FE deployments; its view/truncate checks do not explicitly establish a warm cache immediately beforehand.
- Expected results: the new regression output matches the scripted schema and row transitions on inspection. Its generation provenance and runtime results were not independently verified.
- Observability and errors: cache invalidation failures are isolated and logged; MV failures include IDs and source names, and admission-validation failures include the query ID. No new critical RPC/transaction path or missing metric was identified.
- Persistence and failover: no new durable state was added. Metadata write/replay hooks and checkpoint behavior were traced; DML continues to rely on visible-version comparisons without scanning SQL caches during transaction replay.
- Writes and atomicity: no data-write implementation, delete-bitmap, or rowset visibility logic changed. Metadata invalidation follows mutation, cloud restore retains remote version ownership, and SQL admission/lookup independently check data and schema versions. The MV discovery race remains as reported.
- FE/BE variable propagation: no new transmitted variable was added. The updated TableVersion cache-key component is shared by the FE and CacheAnalyzer backend-key computation.
- Performance: metadata invalidation scans resident SQL/partition entries and relevant MV dependencies; DML replay avoids those scans. Admission performs fresh dependency validation; lock/resource lifetimes were traced. No additional actionable performance regression was substantiated statically.
- Further checks: independently challenged and deduplicated all candidates, including the existing equal-version Hive backend-cache limitation, which was present in the base and is excluded from new findings. No additional user-provided focus was supplied.
Validation is static only: no builds, tests, benchmarks, source modifications, or submodule initialization. Dynamic race behavior and generated-output provenance remain outside this review's execution scope.
Review converged after 2 rounds: all three latest-round reviewers returned NO_NEW_VALUABLE_FINDINGS, full static coverage is complete, and every candidate and risk focus has an explicit disposition. No unresolved concrete question remains.
Review converged after 2 round(s).
### What problem does this PR solve? Related PR: apache#70397 Related PR: apache#70398 Problem Summary: External refresh can miss dependent MV rewrite caches when SQL name normalization changes a canonical mixed-case table name. A concurrent MV refresh can also temporarily remove the MV from its dependency indexes, allowing invalidation to miss a newly published pre-refresh plan. Build lookup identities from canonical metadata names and coordinate index replacement/unregistration with invalidation discovery. Release the index monitor before table resolution or MV cache invalidation to preserve lock order. ### Release note Fix missed MV rewrite-cache invalidation after external metadata refreshes. ### Check List (For Author) - Test: Unit Test: 72 focused FE tests passed. The new 9-case regression suite reproduced 7 failures before the fix; all cases pass after the fix. FE build and Checkstyle passed; multi-FE/BE regression was not run locally. - Behavior changed: Yes, preserve canonical dependency identity and avoid missing registered MVs during concurrent dependency replacement. - Does this need documentation: No
### What problem does this PR solve? Related PR: apache#70397 Related PR: apache#70398 Problem Summary: External refresh can miss dependent MV rewrite caches when SQL name normalization changes a canonical mixed-case table name. A concurrent MV refresh can also temporarily remove the MV from its dependency indexes, allowing invalidation to miss a newly published pre-refresh plan. Build lookup identities from canonical metadata names and coordinate index replacement/unregistration with invalidation discovery. Release the index monitor before table resolution or MV cache invalidation to preserve lock order. ### Release note Fix missed MV rewrite-cache invalidation after external metadata refreshes. ### Check List (For Author) - Test: Unit Test: 72 focused FE tests passed. The new 9-case regression suite reproduced 7 failures before the fix; all cases pass after the fix. FE build and Checkstyle passed; multi-FE/BE regression was not run locally. - Behavior changed: Yes, preserve canonical dependency identity and avoid missing registered MVs during concurrent dependency replacement. - Does this need documentation: No
|
run buildall |
|
/review |
There was a problem hiding this comment.
Static review of all 37 changed paths at head 44e6c2e found no new substantiated introduced defect. The two existing P2 findings are repaired: canonical external names now bypass SQL-name normalization, and dependency replacement/discovery share a monitor that is released before MV locking. Neither is duplicated. No existing P0/P1 finding was independently confirmed to remain applicable. No additional user review focus was supplied.
- Goal and evidence: metadata mutation/replay now invalidates derived caches, fences in-flight SQL contexts, and checks OLAP schema identity. Reviewed the complete mutation-to-admission/lookup flow and associated unit/replay/follower test assertions.
- Scope and clarity: shared Env callbacks and external invalidation wrappers keep changes focused on cache freshness and supporting test fixtures.
- Concurrency: traced query admission, journal replay, DDL, external refresh and MV refresh. Atomic revisions and existing sorted table locking remain intact; relation/catalog monitors are released before MV invalidation. No introduced deadlock was substantiated.
- Lifecycle and initialization: snapshots strongly retain their weak-cache revision cells. Fresh admission validation closes its own StatementContext and restores the executing context. The allocation-only ID fallback no longer retains a connection. MV generation checks reject overlapping discovered invalidations.
- Configuration: no configuration item was added; SQL-cache container replacement retains the shared revision owner.
- Compatibility: existing journal and persisted formats remain unchanged. OLAP schema identity changes opaque backend cache hashes, safely separating entries across differing FE versions; constructor overloads remain available.
- Thrift/protobuf: no definitions changed, so cross-branch numeric-ID checks are inapplicable.
- Parallel paths: inspected live/replay schema and rollup completion, constraints, rename/swap/drop/recovery, partition changes, cloud/shared-nothing restore, FE/BE SQL admission, and external name/typed/engine/full/partial/cold invalidation routes.
- Conditional behavior: checkpoint exclusion, stable-ID fallback for missing objects, legacy CREATE payloads, cloud version ownership and failed dependency validation agree with their callers. External closing row-count fences remain in finally blocks.
- Test coverage: inspected revision capture, late puts, data/schema changes, namespace isolation, failure fanout, serialized replay, prepared contexts, canonical names and latch-controlled MV races. The follower suite skips single-FE setups; its view/truncate stages do not explicitly establish a warm cache immediately beforehand.
- Expected results: new regression rows match the ordered SQL/schema transitions on inspection. Generated-output provenance and runtime outcomes were not independently verified.
- Observability and errors: independent cache failures are isolated and logged; MV failures include source names and IDs, and admission failures include query IDs. No new critical distributed operation requires additional metrics.
- Persistence/failover: no durable state was added. Notifications follow existing applied mutations/journals and skip checkpoint cache work; legacy replay handling remains covered.
- Writes/atomicity: transaction publication, rowset visibility and delete-bitmap logic are unchanged. Data versions remain authoritative, and cloud restore avoids local version increments.
- FE/BE variables: no transmitted variable was added; the new schema component is shared cache-key input rather than a protocol change.
- Performance: metadata invalidation scans resident caches/indexes; DML still uses lazy version checks without walking SQL caches during transaction publication. No actionable introduced hot-path regression was substantiated.
- Further checks: all candidates and risk focuses have explicit dispositions. ADMIN partition-version rollback backend-cache behavior and the task-local MV new-dependency publication ordering were confirmed unchanged from the supplied base and excluded from new findings; no distinct introduced failure was established for those hypotheses.
Review converged after 2 rounds: all three latest-round reviewers returned NO_NEW_VALUABLE_FINDINGS, full static coverage is complete, and no unresolved introduced-defect candidate remains. Both known P2 fixes were independently checked against current code and their targeted test assertions.
Validation was static only: no builds, tests, benchmarks, source modifications or submodule initialization. Runtime race behavior and generated-output provenance remain outside this review's execution scope.
Review converged after 2 round(s).
### What problem does this PR solve? Related PR: apache#63612 Problem Summary: With a one-entry external database object cache, constructing an MV plan that references two databases evicts a database object. The eviction was reported as a metadata change, invalidating the MV generation and restarting every successful build indefinitely. Separate automatic local-object retirement from explicit metadata invalidation while retaining typed engine cleanup and row-count fencing. Bound rewrite-cache construction to four attempts so sustained metadata changes fall back through the callers' existing optional-optimization exception handling. ### Release note Avoid repeated query planning caused by external database object cache eviction during materialized view rewriting. ### Check List (For Author) - Test: Unit Test - 192 focused FE tests passed with zero failures, errors or skips; FE build and integrated Checkstyle passed. The unfixed 4.1 reproduction had 6 failures among 11 tests. - Behavior changed: Yes - automatic database object eviction no longer reports a semantic metadata change, and unstable optional MV cache construction terminates after four attempts. - Does this need documentation: No
Local follow-up review — ✅ PASSCommit: When an MV plan resolves two external databases with a one-entry database object cache, eviction used to report a semantic database change. That invalidated the rewrite generation of every successful build and caused unlimited reconstruction.
The original 4.1 reproduction failed six of eleven tests before the production fix. This is a focused repair receipt; remote CI and the newly triggered review remain separate gates. |
|
run buildall |
|
/review |
There was a problem hiding this comment.
Static review of all 41 changed paths at head 4dd904a found no new substantiated introduced defect. Both existing P2 findings are fixed: dependency lookup preserves canonical external names, and dependency replacement/discovery share a monitor released before MV locking. Neither is repeated. No existing P0/P1 inline finding remains to carry forward. No additional user review focus was supplied.
- Goal and evidence: traced applied metadata changes through derived-cache invalidation, revision capture, fresh admission validation and lookup. Focused FE/replay/race tests and the follower regression exercise the intended behavior on inspection.
- Scope and clarity: shared Env notifications and external invalidation wrappers keep freshness handling centralized; supporting planner/lifecycle changes prevent invalidation from disrupting cache construction.
- Concurrency: inspected query planning/admission, DDL, journal replay, external refresh/retirement and MV refresh. Revision checks reject late stale publication. Dependency discovery and metadata retirement release their monitors before MV callbacks; table locks protect OLAP partition-cache construction. No new lock cycle was substantiated.
- Lifecycle and initialization: snapshots strongly retain weak-cache revision cells; fresh validation closes its own StatementContext and restores the executing context. The ID-only fallback avoids retaining a connection. Automatic DB-object eviction retires local metadata without reporting a schema change; repeated MV construction is bounded and all callers handle its optional failure.
- Configuration: no new setting; replacing the SQL-cache container preserves its revision owner. Existing dynamic settings remain in use.
- Compatibility: no journal/storage format change. Added OLAP schema identity changes opaque backend cache hashes and prevents cross-schema reuse; existing constructor overloads remain available.
- Thrift/protobuf: no definitions changed; cross-branch numeric-ID checks are inapplicable.
- Parallel paths: reviewed live/replay schema and rollup completion, constraints, drop/recover/rename/replace, partition operations, cloud/shared-nothing restore, FE/BE SQL admission, and external name/typed/engine/namespace/cold/partial routes.
- Special conditions: legacy CREATE names come from the registered table; rename/replace preserve original identities; checkpoint callbacks skip local cache work; cloud restore keeps version ownership in the meta service. Nested finally blocks retain closing row-count fences on invalidation failure.
- Test coverage: inspected positive/negative dependency tests, in-flight publication and late puts, namespace isolation, error fanout, serialized journals, prepared contexts, canonical names, capacity eviction and latch-controlled locking races. The follower suite skips single-FE setups; its view/truncate stages do not explicitly establish a warm cache immediately beforehand.
- Expected results: added ordered regression rows match the schema/data/view transitions on inspection. Generated-output provenance and runtime outcomes were not verified.
- Observability and errors: failures are isolated per cache and logged; MV failures include source scope and MV ID, admission failures include query ID. No additional critical metrics requirement was identified.
- Persistence/failover: no durable state added. Existing mutation journals remain authoritative; replay notifications cover derived local state, including legacy identities, without persisting cache revisions.
- Writes/atomicity: transaction publication, rowset visibility and delete-bitmap logic are unchanged. Data-version validation remains authoritative; DML VISIBLE does not scan cached queries under publication locks.
- FE/BE variables: no transmitted variable added. Schema identity participates in the shared cache hash without a protocol field change.
- Performance: metadata invalidation scans resident caches/indexes; ordinary transaction publication retains lazy validation. No actionable introduced hot-path regression was established. The bounded MV retry prevents indefinite planning under repeated invalidation.
- Further checks: all candidates and risk items have explicit dispositions. Backend identity and newly discovered MV-dependency publication hypotheses were compared with the supplied base; no introduced failure was established. Selective Hive replay updates the freshness timestamp and current MV partition checks remain in place.
Review converged after 1 round: all three reviewers' latest results were NO_NEW_VALUABLE_FINDINGS, all changed paths were covered, and every candidate/risk was verified as fixed, deduplicated or dismissed with evidence. No unresolved introduced-defect candidate or static coverage gap remains.
Validation was static inspection only: no builds, tests, benchmarks, source changes or submodule initialization. Runtime behavior and generated-output provenance remain unverified.
Review converged after 1 round(s).
CI triage — infrastructure failureHead: The environment preflight failed before the PR checkout/merge, compilation or test execution. The log reports The agent was This is an infrastructure blocker. CI maintenance needs to reconcile the prior-build cleanup state and restore the agent before rerunning FEUT. The PR head is unchanged. The previously published local test/build results and successful current-head static review remain separate evidence; this remote FEUT gate is failed, with no tests executed. |
|
run buildall |
|
/review |
There was a problem hiding this comment.
Static review of all 41 changed paths at head 4dd904a found no new substantiated introduced defect. Both existing P2 findings are fixed: external dependency keys preserve canonical names, and dependency replacement/discovery share a monitor released before MV locking. Neither is duplicated. No existing P0/P1 inline finding remains to carry forward. No additional user review focus was supplied.
- Goal and evidence: traced metadata mutation/replay through derived-cache invalidation, revision capture, fresh admission validation and lookup. The inspected FE/replay/race tests and follower regression exercise these mechanisms.
- Scope and clarity: shared Env callbacks and external invalidation wrappers centralize freshness handling; planner/lifecycle changes support cache construction during invalidation.
- Concurrency: inspected query binding/admission, DDL, journal replay, external refresh/retirement and MV refresh. Atomic revisions reject stale publication. Relation discovery releases its monitor before catalog/MV work, and MV cache construction and external dependency resolution run outside the MV lock. No introduced lock cycle was substantiated.
- Lifecycle and initialization: live snapshots strongly retain weak-cache revision cells. Temporary admission validation closes its own resources and restores the executing StatementContext. The ID-only fallback avoids retaining a connection. Automatic DB-object eviction still retires metadata and typed engine entries without reporting a metadata change. MV rebuild attempts are bounded; all eight current callers catch optional construction failure.
- Configuration: no new setting; dynamic SQL-cache replacement retains its revision owner. Existing refresh flags retain their meaning.
- Compatibility: no journal/storage format change. OLAP schema identity feeds existing opaque backend cache hashes, separating entries across schema changes and differing FE versions. Existing constructor overloads remain available.
- Thrift/protobuf: no definitions changed, so cross-branch numeric-ID comparison is inapplicable.
- Parallel paths: reviewed live/replay rollup and schema-change completion, constraints, drop/recover/rename/replace, partition changes, cloud/shared-nothing restore, FE/BE cache publication, views/tables/namespaces, guarded/unguarded MV caches, and external name/typed/engine/partial/cold refresh routes.
- Special conditions: rename/replace preserve original identities after mapping mutation; legacy CREATE journals use table-object identity; checkpoint callbacks skip local cache work; cloud restore keeps remote version ownership. Engine failures still reach derived-cache and final row-count fences.
- Test coverage: inspected revision scopes, late capture/put, data/schema changes, transaction COMMITTED versus VISIBLE, live/replay constraints, legacy names, swapped identities, failure isolation, canonical names, dependency replacement races, catalog retirement, automatic eviction, bounded retries and prepared context reset. Latches and bounded cleanup checks support deterministic race tests. Tests were not run.
- Expected results: replay golden values match the insert/drop/rename/view/truncate SQL on inspection; ordered output and the expected Unknown column check are consistent. Output-generation provenance was not verified.
- Observability: invalidation failures identify cache category; per-MV failures include target/source identity; admission failures include query ID. No additional distributed operation requiring new metrics was identified.
- Persistence and failover: callbacks add process-local derived state and run after existing replay mutations. Existing journal writes/formats, restore finish ordering and transaction persistence remain unchanged; no new unjournaled durable state was identified.
- Writes and atomicity: no rowset, bitmap or commit-protocol change. Existing table/partition visible-version checks remain the data-freshness authority; schema/revision checks supplement them. Derived-cache failures are isolated from already-applied metadata operations.
- FE/BE variable transfer: no new wire variable. Both cache population and validated lookup use the existing hash/protocol path with the added schema identity.
- Performance: revision checks scale with query dependencies; metadata changes scan resident caches/indexes, while transaction VISIBLE replay retains lazy validation. Automatic eviction avoids invalidation-driven MV rebuild churn; bounded MV retries prevent indefinite construction. No actionable introduced performance regression was established.
- Further checks: all initial risks and investigated candidates were independently reviewed and dismissed with evidence or identified as fixed existing findings. Backend-key concerns were checked against the actual consumer gate: normal replanning does not directly reuse BE packets. Nondefault cold catalog-refresh and namespace-registration hypotheses were compared with unchanged base code and do not establish an introduced failure.
The review converged after 1 round: all three reviewers returned NO_NEW_VALUABLE_FINDINGS, all changed paths were covered, and no unresolved introduced-defect candidate or assigned static coverage gap remains.
Validation was static inspection only: no builds, tests, benchmarks, source modifications or submodule initialization. Runtime behavior, stress/cloud integration and generated-output provenance remain unverified.
Review converged after 1 round(s).
CI triage — external test infrastructure prerequisiteHead: The artifact identity check and package verification passed. The external preflight then failed in
The source-payload check requires Doris deployment, third-party dependency deployment and external test execution were explicitly skipped. Final cleanup reports a clean accepted receipt with no cleanup errors; there is no failing external test case from this run. This is a branch/pipeline prerequisite blocker. CI needs a branch-compatible runner or separately reviewed rootless-Docker support on the target branch. The cache-invalidation PR remains focused and its source head is unchanged. Current-head compile build 3249 and vault_p0 build 3346 passed. Other pending/failed checks retain their own status; this preflight failure is not reported as an executed regression failure. |
### What problem does this PR solve? Related PR: apache#63612 Problem Summary: The synchronous-MV cache regression warmed base-table queries before publishing rollups and expected their old SQL cache entries to survive. Rollup publication now notifies table metadata changes and correctly invalidates those entries. Share one check across the rename, insert and drop scenarios to verify cache misses after publication, then rerun each query and verify cache hits. Preserve the later rename, insert and drop assertions. ### Release note None ### Check List (For Author) - Test: 40 focused FE unit tests passed; regression framework and modified Groovy suite compiled; FE build with integrated Checkstyle passed. End-to-end regression was not run locally because this worktree has no BE binary. - Behavior changed: No production behavior change; update regression expectations for metadata invalidation. - Does this need documentation: No
Local CI repair review — validated locallyCommit: The regression warms three base-table SQL caches before creating synchronous MVs. The repair uses one shared assertion in all three create-MV scenarios: require cache misses after publication, execute the same queries again, then require cache hits. Later rename, insert and drop checks are preserved. Only this regression script changes; no production or golden-result changes. Validation:
Other failures remain separate and are not claimed fixed:
Remote gates will be triggered for this commit; no all-green or merge-readiness claim. |
|
run buildall |
|
/review |
There was a problem hiding this comment.
Static review of all 42 changed paths at head 742a21e found no new substantiated introduced defect. Both existing P2 findings are fixed: canonical external dependency names bypass SQL-name normalization, and dependency replacement/discovery share a monitor released before MV locking. Neither is duplicated. No existing P0/P1 finding remains to carry forward. No additional user review focus was supplied.
- Goal and evidence: traced metadata mutation/replay through SQL, sorted-partition, MV rewrite and aggregate-cache invalidation, plus SQL revision capture, admission and lookup. Inspected focused FE tests and the follower regression; no tests were executed.
- Scope and clarity: shared Env callbacks and external invalidation wrappers centralize cache freshness. Supporting lock/lifecycle changes and test fixture stubs are related to that behavior.
- Concurrency: inspected query binding, cache admission, DDL, replay, external retirement and MV refresh. Atomic revisions and MV generations reject stale publication. Dependency discovery releases its monitor before catalog/MV work; cache construction and external dependency resolution run outside MV locks. Catalog subtype monitor wrappers were also considered; no introduced lock cycle was substantiated.
- Lifecycle and initialization: active snapshots strongly retain weak-cache revision cells. Temporary admission validation owns and closes its separate resources and restores the executing StatementContext in finally. The ID-generator fallback no longer retains a connection or enters Env creation. MV reconstruction retries are bounded and optional rewrite/trait callers catch construction failures.
- Configuration: no new setting; SQL-cache container replacement retains the revision owner. Automatic DB-object eviction retires typed connector metadata without reporting a logical metadata mutation.
- Compatibility: no journal/storage format changes. Base-index/schema identity feeds the existing opaque backend cache hash, separating schema generations and cache entries across differing FE versions. Existing constructor overloads remain available.
- Thrift/protobuf: no definitions changed; cross-branch numeric-ID comparison is inapplicable.
- Parallel paths: inspected live/replay schema and rollup completion, constraints, drop/recover/rename/replace, partition changes, cloud/shared-nothing restore, FE/BE SQL admission, views/tables/namespaces, both MV cache variants, and external typed/name/engine/cold invalidation routes.
- Special conditions: rename/replace retain original identities after mapping changes; legacy CREATE journals use table-object identity. Checkpoint callbacks skip local cache work, cloud restore preserves remote version ownership, and finally blocks retain derived-cache and closing row-count fences on engine failure.
- Test coverage: inspected revision scopes, late capture/put, data/schema changes, COMMITTED versus VISIBLE transactions, live/replay constraints, legacy journals, swapped identities, failure isolation, canonical names, dependency replacement races, catalog retirement, automatic eviction, bounded retries and prepared contexts. Not every journal opcode has a dedicated test. The follower suite skips single-FE setups; its view/truncate stages do not explicitly establish a warm cache immediately beforehand.
- Expected results: new/changed assertions were inspected. Ordered golden values and output shapes match the scripted insert/drop/rename/view/truncate transitions; generated-output provenance and runtime results were not verified.
- Observability and errors: invalidation failures are isolated per cache and logged; MV failures include target/source identity and admission failures include query ID. No additional distributed operation requiring new metrics was identified.
- Persistence and failover: no new durable state. Callbacks derive process-local invalidation after existing replay mutations; existing journal writes, payload formats and finish ordering remain intact.
- Writes and atomicity: no BE rowset, delete-bitmap or commit-protocol change. Published table/partition versions remain the data-freshness authority; metadata/schema fences supplement them. Ordinary DML VISIBLE replay avoids cache scans under publication locks.
- FE/BE propagation: no new wire variable. CacheAnalyzer and validated lookup share the existing hash/protocol path with the added schema identity.
- Performance: revision checks scale with query dependencies; metadata mutations scan resident caches/indexes. Weak revision cells avoid permanent tombstones, eviction avoids unnecessary MV reconstruction, and bounded retries prevent indefinite construction. No actionable introduced performance regression was established statically.
- Further checks: all initial risks and investigated hypotheses were independently verified and dismissed with evidence or identified as fixed existing issues. No unresolved introduced-defect candidate or changed-file static coverage gap remains.
The review converged after 1 round: both complete reviewers and the dedicated risk reviewer returned NO_NEW_VALUABLE_FINDINGS, coverage is complete, and every risk has an explicit disposition.
Validation was static inspection only: no builds, tests, benchmarks, source modifications or submodule initialization. Runtime behavior, stress/cloud integration and generated-output provenance remain unverified.
Review converged after 1 round(s).
cherry pick from #63612