perf(disttae): eliminate exact catalog lookup probe allocations

[VioletQwQ-0]

What type of PR is this?

• API-change
• BUG
• Improvement
• Documentation
• Feature
• Test and CI
• Code Refactoring

Which issue(s) this PR fixes:

issue #27235

What this PR does / why we need it:

This Draft PR preserves the existing exact-BTree catalog invalidation semantics while reusing the temporary lookup probes used by CatalogCache.HasNewerVersion.

• Named database and named table lookups obtain a per-cache sync.Pool probe, fully overwrite its fields, run the existing Ascend and identity/timestamp checks, clear the probe through a shared release helper, and return it to the pool.
• The public method signature, account-wide and database-only branches, prepared-plan/RC-cache query construction, mutation/replay/GC paths, and all SQL/wire/protobuf APIs are unchanged.
• The probe is returned only after Ascend completes; a panic drops it instead of returning a potentially dirty object. Pool eviction by GC only causes a cold allocation and cannot change the result.

The first commit adds a legacy exact oracle, deterministic differential coverage, concurrent-reader/lifecycle checks, and a benchmark matrix for table/database histories 1/16/256/4096 in warmed-negative and changed states. The follow-up commits add automatic testing.AllocsPerRun gates, deterministic release/reset coverage, a frontend CheckTableDefChange smoke, and an RC getCachedTableByKey replay/version/GC/identity smoke.

This PR deliberately does not include the rejected account-bucket named-dependency scheme (147 measured false positives), the deferred precise shadow watermark, or any instrumentation/report API. #27447 remains Draft and is not rebased or merged.

Historical profile numbers (17,060,022,442 B and 52,996,163 objects) are owner attribution only. The A+C TPCC result (+0.7026%) is directional combination evidence; this PR makes no independent TPS, P95/P99, or +20% claim.

Validation

• Exact differential tests, replay/alter/truncate/drop/recreate/GC sequence, direct dirty-probe reset, and concurrent readers: PASS.
• Table and database warmed testing.AllocsPerRun gates require 0 allocations while asserting both negative and changed boolean results; cold and post-GC measurements are logged separately.
• pkg/vm/engine/disttae/cache focused/full tests and full race: PASS through the repository CGo wrapper.
• pkg/vm/engine/disttae RC consumer focused/full tests and full race: PASS through the repository CGo wrapper.
• pkg/frontend CheckTableDefChange focused/full tests: PASS through the repository CGo wrapper.
• Adaptive focused race stress for the new tests and owning-package race runs: PASS.
• make build-with-prebuilt-native: PASS on a8efb132191943b7b1a40a00f023dd9e276ad7ae.
• Escape analysis on the final head: no hot-path TableItem/DatabaseItem probe escape; only cold sync.Pool.New allocation remains.
• Five-run benchmark matrix and benchstat: pooled warmed paths report 0 B/op, 0 allocs/op; final-head longer database-history control shows no stable regression.
• Exact-head semantic preflight: PASS at a8efb132191943b7b1a40a00f023dd9e276ad7ae, diff hash f06d8c32c25b5f0c4b1eb04ab5a7c44b6e500a1184245994150639d6b4b93c37.
• Plain GOWORK=off go build and go vet remain environment-blocked: a clean base reproduces the existing pkg/common/docfilter CGo declaration errors (C.mo_cbitmap_*/C.mo_croaring_*). They are not claimed as passes.

BVT: N/A — the changed contract is an internal lookup-probe allocation optimization with unchanged SQL semantics; SQL cannot assert probe allocation. Exact invalidation behavior is covered by the cache differential tests plus the frontend and RC consumer smoke tests. Full multi-CN SQL DDL validation remains a QA task.

QA required: yes. The direct consumer tests close the local wrapper/eviction path; prepared-plan and multi-CN RC DDL smoke remain QA validation. This PR is ready for human review only after the required remote checks and approval; it makes no claim that local consumer smoke replaces that QA.

[VioletQwQ-0]

Addressed the review findings on head a8efb132191943b7b1a40a00f023dd9e276ad7ae:

• Added table/database testing.AllocsPerRun gates with warmed negative/changed boolean assertions; cold and post-GC observations are logged separately.
• Added shared exact-object release helpers and a deterministic dirty-probe reset test; removed the nondeterministic sync.Pool.Get reset assertion.
• Added frontend CheckTableDefChange and RC getCachedTableByKey consumer smoke coverage, including replay/version changes, GC, and identity recreation.
• Reworked the local semantic preflight so direct cache cases no longer claim same_statement or consumer terminal coverage; it is PASS on the exact final head.

The PR remains Draft for now because the new-head MatrixOne ALL CI is pending and there is no approval/review decision yet. No SQL/multi-CN QA result is being inferred from the local consumer smoke tests.

[XuPeng-SH]

Deep review completed on exact head a8efb132191943b7b1a40a00f023dd9e276ad7ae.

The production change preserves the existing exact-BTree comparator, pivot fields, callback checks, and consumer semantics. The pooled pivot is not returned until synchronous Ascend completes, is fully overwritten on acquisition and zeroed before reuse; panic drops the active probe to GC, while sync.Pool GC eviction only restores the cold allocation path. CatalogCache remains pointer-owned, and no persistent/wire/public contract changes. Q1-Q3 and reuse-generation review found no leak, wait, stale-state, or unbounded-retention blocker.

Fresh reviewer evidence: all nine new focused cache/frontend/RC tests passed; the concurrent-reader test passed under -race -count=100; the cache owning package passed in full under -race; five-run table/database benchmarks at histories 1/16/4096 showed the pooled paths at 0 B/op, 0 allocs/op with consistently lower latency than the legacy 1 alloc/op controls. Exact-head CI is green. BVT is correctly N/A for this internal allocation optimization.

[qodo-code-review]

Qodo reviews are paused for this user.

Troubleshooting steps vary by plan Learn more →

On a Teams plan?
Reviews resume once this user has a paid seat and their Git account is linked in Qodo.
Link Git account →

Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center?
These require an Enterprise plan - Contact us
Contact us →