Repository navigation
feat(bench): the Bulk and Jobs benches run on Oracle behind a provider switch (G7 Oracle 8/12 Part 1/2, #556) - #560
Conversation
…r switch (G7 Oracle 8/12 Part 1/2, #556) - BulkBenchTests, BulkIntakeBenchTests and JobsBenchTests are written once on SuiteStore: PostgreSQL in the class itself (output unchanged), Oracle in BulkBenchOracleTests, BulkIntakeBenchOracleTests and JobsBenchOracleTests on the Oracle collection's pinned container, each carrying Category=Bench AND Category=Oracle so neither the integration lane (Category!=Bench&Category!=Oracle) nor the oracle lane (Category=Oracle&Category!=Bench) runs them. Oracle bench lines end in `provider=oracle`; the intake's size on Oracle is the schema's segments. - scripts/bench-bulk.sh and scripts/bench-jobs.sh take GOLDPATH_BENCH_PROVIDER (postgres|oracle, default postgres), and refuse an Oracle run that printed no `provider=oracle` line. bench-jobs.sh on PostgreSQL now filters Category!=Oracle, so it runs the set it ran before. - bench.yml: a `provider` dispatch input; the monthly schedule stays PostgreSQL-only (Category=Bench&Category!=Oracle). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…bind PostgreSQL only (#556) - jobs-benchmarks.md: Oracle Free 23.26.3 beside PostgreSQL 17 on the same sandbox (3 Oracle runs; checkpoint 8.3-9.2 ms/chunk vs 7.6 ms; plan->complete 1.7-1.8 s vs 1.5 s). The interactive proof misses its own budget on PostgreSQL too on this 4-CPU profile. - bulk-benchmarks.md: no Oracle number - every upload over 2,000 bytes fails with ORA-01460 (GoldpathBulkFileChunks.Data is RAW(2000) on Oracle, chunks are 256 KB); #559 filed. - Bench facts assert their budgets on PostgreSQL only (oracle RFC §6: measured, not promised); correctness asserts run on both. - bench.yml's Oracle step runs both scripts and fails if either did; the summary also keeps `BENCH items=` lines. - CHANGELOG, ga-gate-status G7, open-threads T68. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…yml runs them (#556) CI gates on 17eecab: merge-tier selftest red, "gate script not Tier C" for both scripts, because bench.yml's Oracle step now calls them. Listed beside bench-campaign.sh; selftest green, docs-freshness and ledger-check rc 0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
CI on Fixed in Generated by Claude Code |
Verifier verdict (run 2026-11-11)Head verified: Verdict: no blocking findings. 0 BLOCKER, 0 MAJOR, 0 MINOR, 2 NIT. The switch, the refusal, the lane isolation and the Jobs-on-Oracle numbers all reproduce here. #556's Bulk criterion is unmet, which the PR says plainly (#559, T68), but it still What I checked
Deep review
Edge-case probes (this run)
NITs (optional)
Question for the owner (
|
CHANGELOG.md: both sides' [Unreleased] entries kept. open-threads.md: main's T67 and this branch's T68 both kept, in number order. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018AFdNZbL5UCo9Kkefph8f3
|
Brought up to date with main@02128c5. The head is now Conflicts resolved in:
Build, tests and gates:
The two The PR's behaviour and recorded decisions are unchanged by the merge. I removed Generated by Claude Code |
Verifier verdict (run 2026-11-11, daytime)Head verified: This morning's scope question is resolved by the owner. The PR body now says Verdict: no findings in the PR's content. Evidence
Finding
Edge-case probes (the provider switch's refusal; read-only runs of the PR's scripts)
Deep review (delta-scoped: merge-only; the full review is in my verdict of 03:53 today)
Generated by Claude Code |
# Conflicts: # CHANGELOG.md # docs/strategy/ga-gate-status.md
|
Synced with Conflicts resolved in:
In These ran locally on the merged tree:
The Owner's session (Claude Code). |
|
Synced with On The merge brings only #573's 9 files, and this PR touches none of them. Its own lines in Owner's session (Claude Code). |
Verifier verdict (run 2026-11-12)Head verified: Verdict: verified. No blocking findings, 3 optional NITs. Delta since the last verdict (99a51e3)
Steps 1–6
The red
Deep review
Edge-case probes (throwaway, not pushed)
NITs (optional, none blocks)
Could not run: a Labels: Owner: safe to merge — summary: Generated by Claude Code |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What and why
Part 1/2 of #411 (G7 Oracle 8/12): the Bulk and Jobs benches can run on Oracle. Tests, scripts and
bench.ymlonly; no package changes.BulkBenchTests,BulkIntakeBenchTestsandJobsBenchTestsnow take aSuiteStore, the store the Campaign, Bulk and Archival suites already use for both providers.BulkBenchOracleTests,BulkIntakeBenchOracleTests,JobsBenchOracleTests) run on a fresh schema of the Oracle collection's pinned container (OracleFixture.Image).GOLDPATH_BENCH_PROVIDER=postgres|oracleonscripts/bench-bulk.shandscripts/bench-jobs.sh, withpostgresas the default.provider=oracle.bench.ymlgains aproviderdispatch input.oracleruns Bulk (intake at 1,000,000 rows) and Jobs on Oracle, and the step fails if either script did. The monthly schedule stays PostgreSQL-only (Category=Bench&Category!=Oracle).Acceptance criteria of #556, each with its evidence
Both scripts take a provider choice, default PostgreSQL; with Oracle the same facts run on
OracleFixture's image; PostgreSQL output is unchanged.BENCH-BULK ingest10k upload=0.71s total=1.89s (budget 300s)from tonight's run.bench-jobs.shon PostgreSQL now filtersCategory!=Oracle, so it runs exactly the set it ran before.No test that is both Bench and Oracle runs in the per-PR lanes. Both lanes' real filters were run against the built assembly, narrowed to the twins (
&FullyQualifiedName~BenchOracleTests), and each answered "No test matches":Category=Oracle&Category!=Bench(the oracle lane) andCategory!=Bench&Category!=Oracle(the integration lane). So didCategory=Bench&Category!=Oracle(bench.yml's PostgreSQL filter).Bulk: the 1M intake and the 10k/100k facts on Oracle, with numbers: NOT MET. Every Bulk fact fails on Oracle on its first save with
ORA-01460. That is a product defect, filed as Bulk on Oracle: any upload over 2,000 bytes fails with ORA-01460, because the file chunk column is RAW(2000) and the store writes 256 KB chunks #559:GoldpathBulkFileChunks.DataisRAW(2000)on Oracle (pinned inOracle/composed-model.golden.txt), while the store writes 256 KB chunks. So any upload over 2,000 bytes fails.BulkOracleTestsuploads files of a few hundred bytes, which is why it is green.bulk-benchmarks.mdsays this, and names what "Reading the numbers" must cover once Bulk on Oracle: any upload over 2,000 bytes fails with ORA-01460, because the file chunk column is RAW(2000) and the store writes 256 KB chunks #559 lands (EF array binding, RFC §2.6).Jobs: every
JobsBenchTestsfact has an Oracle column injobs-benchmarks.md. These are three Oracle runs and one PostgreSQL run on the same sandbox, the same night:Nothing is an order of magnitude apart.
Every Oracle number states its profile (machine, image tag, build configuration) and date. No budget is promised.
bench.ymldispatch input. It is Tier C (.github/**).Ledgers: CHANGELOG
[Unreleased];ga-gate-status.mdG7 names the two pages; open-threads T68.coverage-matrix.mdhas no Bench column, so it is unchanged.docs/rfc/goldpath-ga-readiness.mdis not edited.Reproduce
Gates run
dotnet build tests/Goldpath.IntegrationTests -c Releasedotnet format whitespace --verify-no-changeson the four touched test filesdocs-freshness,ledger-check,schema-honesty,skills-parity,template-pinsBench_100k…green.Bench_interactive…misses its own budget (1.5 → 3.4 ms against 1.25× + 1 ms) on this 4-CPU co-located profile. Its PostgreSQL code is unchanged by this PR; the page says so.Pre-PR self-check
///blocks, never between a block and its declaration. Re-read.provider=oraclelines.bench.yml's filters: the PostgreSQL step's filter answered "No test matches" for every Oracle twin.choiceinput, interpolated only intoif:expressions, never into a shell line. No permission, secret or action change.dotnet test's exit code (set +earound it) and fails on it before the Oracle-line check. An Oracle run with zero bench lines is refused, and so is a run where any line lacks the tag.BulkTests/BulkOracleTests' existing pattern, andBenchProvideris the one place the tag is spelled.Not verified
bench.ymldispatch run: the workflow is dispatch or monthly only.Part of #411.
Refs #556 — not closed by this PR: Bulk's Oracle numbers are still open under T68 / #559, and #556 closes with them (owner decision, 2026-11-11).
Generated by Claude Code