Skip to content

Fail KV workloads on final errors and wrong data - #69

Draft
KirillKurdyukov wants to merge 5 commits into
masterfrom
fix/slo-kv-checks-20261014
Draft

KirillKurdyukov wants to merge 5 commits into
masterfrom
fix/slo-kv-checks-20261014

Conversation

@KirillKurdyukov

@KirillKurdyukov KirillKurdyukov commented Oct 15, 2026 •

Copy link
Copy Markdown
Collaborator

Changes

  • Fail all four existing Query, JDBC, Spring Data JDBC and Spring Data JPA
    workloads on an unrecovered final logical operation error, preserving its
    SDK cause through supervision and cleanup.
  • Keep one control owner with an independently retained expected generation
    and full payload. Check successful reads before advancing and after each
    acknowledged write, including during ordinary load.
  • Reserve the control key outside random traffic and use a process-private
    table for the complete create/run/drop lifecycle.
  • Normalize outgoing timestamps to storage precision before writing.
    Preserve existing schema, CLI, retry policies and load metrics.

Verification

  • Five core tests pass and all four workload modules package with JDK17.
    The runner regression exercises stale-data detection while load is active.
  • Eight original SDK source-image builds pass, without fallback images.
  • Eight real ordinary processes pass across all four adapters. Concurrent
    same-label peers keep separate tables and the longer peer survives the
    shorter peer's cleanup.
  • Sixteen real negative processes fail as expected: missing row, stale
    generation, wrong payload and a final SDK failure for each adapter.
    Data injection follows actual successful SDK read/decoding; SDK failures
    come from an independently dropped acknowledged private table.
  • Independent source-publication gate: APPROVE, no blockers.

Both SDK roles currently use the same SDK library revision, so this is
adapter/ownership verification rather than a two-version benchmark.
Recovered successful outcomes are covered at the core seam; native forced
retry recovery and hosted SLO remain unverified.

Current workload metric contract

Use only existing workload sdk.* counters and successful-operation latency Gauges. workload_type selects table (default), topic, or topic_tx; it does not select a metric format. No SDK convention instruments or native/legacy schema split are supported by this change.

Init pins Action PR81 at immutable d66bd1e10f893e5f70d542c4430a3cffb6d6f9f8. Table retains independently built current/baseline SDKs and requires successful read AND write observations for both refs. Topic requires read/write and Topic-tx transaction observations, including any supplied baseline. Existing actual scenario names, SDK source, metrics YAML, retries, deadlines, load, data checks and fatal process criteria are unchanged.

The selector/pin changes pass actionlint across all seven SDK workflows. Action has135/135 Linux tests, clean static checks/bundle, public exact-SHA verify, independent gate approval and real SDK/OTLP PASS/FAIL/INVALID evidence. These are not a claim that historical native SDK runtime failures were fixed or that new SDK CI is green.

All PRs remain draft; no merge, ready transition or v3 tag move. Default-branch publisher changes are not active merely because this branch was pushed.

Related migration PRs

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant