Description
PR38 failed on the OpenTelemetry upgrade. Track a compatible release-set upgrade with API adaptation and exact-source validation rather than pinning back incompatible log APIs. Root owns post-auto-RC telemetry proof and finalization.
Acceptance Criteria
- #1 Read the PR38 failing log and upstream changelog through Firecrawl before implementation; record the compatibility evidence.
- #2 All go.opentelemetry.io/otel* modules move to the v1.47 release set: otel, metric, trace, sdk, sdk/metric and OTLP metric/trace exporters v1.47.0; otlplog exporters v0.23.0; prometheus v0.69.0; log and sdk/log use the versions required by otlplog v0.23.0, with API adaptation, not pinback.
- #3 No other version changes except named MVS requirements; go mod tidy is clean; internal/semconv diff is empty; no tests are weakened.
- #4 just check and exact-SHA ci-success are green.
- #5 Root verifies post-auto-RC telemetry build_commit matches the accepted SHA, every collector last-success age is <1800s within 30 minutes, failed-attempt delta is zero over the observation window, and fresh Loki logs are present. Long-interval collectors may be reread once at 45 minutes. Root owns telemetry proof and finalization.
Definition of Done
- #1 just check (fmt-check, lint, vet, test, tidy-check, build, vuln)
- #2 just ci before a change that touches the Dockerfile, goreleaser or the image (adds snapshot + image)
- #3 Every new signal or attribute name declared in internal/semconv and listed in docs/signals.md
Implementation Plan
OT lane: read PR38 job and PR37 module diff plus upstream release/API docs; upgrade the complete OTel release set using exporter module requirements; adapt only proven API incompatibilities while preserving the wire contract; tidy, exercise local telemetry export, run just check and CodeRabbit, then return uncommitted for root security review.
Implementation Notes
Compatibility evidence before implementation: PR38 build/vet/test run 37127010446 job 111520175636 compiled and passed race tests; it failed at go mod tidy -diff because go.sum retained old sums and metric/x needed v0.69.0. PR38 mixed v1.47.0 API/SDK with previous exporters; PR37 lists the full release set. Firecrawl read https://github.com/open-telemetry/opentelemetry-go/releases/tag/v1.47.0 and https://pkg.go.dev/go.opentelemetry.io/otel/log@v1.47.0. Logs API/SDK stabilized at v1.47.0; both otlplog v0.23.0 go.mod files require log and sdk/log v1.47.0, not v0.23.0. Current call sites already use attribute.Value/attribute.KeyValue as required by the stable API, so no call-site adaptation or pinback is justified. prometheus v0.69.0 requires procfs v0.22.0; this is the sole non-OTel MVS version increase expected.
Implemented the full OTel release set, including log/sdk/log v1.47.0, indirect otlptrace v1.47.0, metric/x v0.69.0 and upstream-test-only sdk/log/logtest v0.23.0. Sole non-OTel MVS increase is github.com/prometheus/procfs v0.21.1 -> v0.22.0, required by prometheus v0.69.0; selected-module before/after comparison confirms no other changes. go mod tidy completed; cmd/ and internal/ diffs, including semconv and all tests, are empty. Existing local race-enabled telemetry boundary tests passed uncached: OTLP HTTP metrics/logs/traces decoded at an httptest receiver, log bodies/attributes and span context, denied/inferred identity cases, Prometheus alongside OTLP, and rejected-datapoint handling. No production call-site changes were needed because baseline already uses the stable attribute APIs. Required-version check failed on unchanged base and passed after upgrade. Root must still review, commit/push, observe exact-SHA ci-success and post-auto-RC telemetry; criteria/status left untouched.
Loop26: 1 implementation attempt; upgrade landed at 97698a0de490d3b42dd6d545ce42cac84cca1f14 with local/composed just check and exact-SHA CI run 37451033089 ci-success green. Security review and CodeRabbit complete, zero findings. auto-RC run 37452249712 green. Post-RC conservative epoch 2026-10-06T10:57:10Z: exact build, 49 collector census, single current instance, zero resets and fresh Loki verified. At the sole authorized 45-minute reread, certs.packs, httpreq.threats and httpreq.transfer had zero attempts and ages 2798.461s, 2838.461s and 2828.461s. Their source defaults are hourly, so absence of an attempt is not an observed failure; all-collector age acceptance remains unmet. Raw attempt-minus-success counters and sampled window extrema are zero for all 49; tiny PromQL extrapolation differences are not integer failed attempts. Parked needs owner: approve interval-aware acceptance or additional bounded observation; interval changes and live-hour proof were excluded. Resume telemetry acceptance only, not implementation. AC5 deliberately unchecked.
Loop30 consolidation: loop27 zero implementation attempts; GET-only interval-aware proof after allowlist infrastructure block could not establish unique epoch deltas (38 duplicate collectors); overwritten duplicate values rejected. Loop28 both helper attempts used; independent review found Decimal rounding could hide an increase of 1 and count-only census allowed arbitrary replacement; no P1. Loop29 2/2 helper cycles including raw-sample correction; zero network calls: nonexistent hash-test filename then sole infrastructure retry shell parse failure; parked at ceiling. Park commits and loop29 file were read only, not copied. Owner accepts interval-aware limits and all six checks for 49 collectors selected by T full label set in restart mode at E_prime=T-3900; no P1 rerun/helper repair, no dependency/code changes or attempts consumed. Evidence /Users/rob/repos/cf2otel/codex/live-loop30/P1-manual-summary.json: candidate 97698a0de490d3b42dd6d545ce42cac84cca1f14, running version 0.18.0-rc.21, restart mode; E=1791284230, T=1791304066, E_prime=1791300166. Helper exit 1, check 1 passed then Ambiguous E series for access.login_metrics. In-place restart roughly 112s before E retained stale old-process samples in five-minute lookback, same instance identity/different version (97698a0 vs 0.18.0-rc.21); 11/49 had no new-version E sample. Helper rejects the duplicate identity but only switches to restart mode when no E sample shares it, leaving this shape uncovered. No instance identifier or hostname recorded. Manual evaluation uses exact Decimal integers, T full label set, checks 2-4 at E_prime, raw samples over process lifetime. check 1: pass: exactly one cf2otel_build_info_ratio series at T, build commit 97698a0de490d3b42dd6d545ce42cac84cca1f14; check census: pass: 49/49 collectors at T and E, none missing or unexpected; check 2: pass: zero counter decreases in raw attempts and success samples for all 49 collectors from process start (about E-112 s) to T; T instance’s earliest sample is about 15448 s before T-4500; check 3: pass: attempts at T exceed attempts at E_prime for all 49; check 4: pass: attempts minus success identical at E_prime and T for all 49 (zero for all 49); also identical E to T for the 38 collectors with an E series; check 5: pass: last-success age under limit for all 49; check 6: pass: Loki record about 19601 s after E. Verdict: data passes under restart mode; reviewed helper exits 1 on an uncovered in-place-restart shape. Owner explicitly accepts this evidence for AC5. Prior upgrade exact-SHA ci-success run 37451033089 and auto-RC run 37452249712 remain accepted.
loop30, attempts 0, owner-accepted evidence closure
Final Summary
Complete OTel v1.47 release-set upgrade at 97698a0de490d3b42dd6d545ce42cac84cca1f14; only named procfs MVS increase. Local receiver regressions, just check and exact-SHA ci-success run 37451033089 passed. Owner accepts all six manual post-auto-RC checks on 49 collectors by full label set at restart epoch E_prime=T-3900 despite helper exit 1 on stale old-process series. No rerun or repair. Process-start observability is separate follow-up. Conditional DoD image/new-signal items are not applicable to this dependency-only upgrade.