regression-e2e

name

regression-e2e

description

Triage frame for end-to-end (E2E) test, smoke test, regression, and fault-injection validation failures. Use when the user reports an E2E/smoke/regression run is broken, a fault-injection flow fails mid-pipeline, a previously-green test is now flaky, or asks to "triage", "debug the E2E", "why is the regression failing", "help me figure out what broke the smoke test". Produces a short hypothesis list and next inspection surfaces rather than a full runbook. Trigger words — e2e, end-to-end, regression, smoke test, fault injection validation, flaky test, test broke.

Regression / E2E Triage

See also: the aegisctl skill for general CLI composition (NDJSON streaming, name-not-id filters). Specific aegisctl … invocations below are illustrative — aegisctl <noun> [verb] --help is the source of truth and supersedes anything that drifts here.

Provide a short triage frame for E2E failures. Focus on likely problem classes and next inspection directions rather than detailed runbooks.

Triage Order

1. Confirm the test contract still matches product behavior.
   • Stale expected terminal events, outdated fixtures, renamed statuses, old snapshots.
2. Confirm the environment is actually ready.
   • Service health, cluster dependencies, seeded config, storage, queues, required ports.
3. Confirm version alignment across the stack.
   • Local CLI / test harness, deployed backend, worker images, schemas, container tags — look for drift.
4. Confirm auth and config wiring.
   • Tokens, permissions, endpoints, namespaces, feature flags, project/tenant resolution.
5. Confirm rerun semantics.
   • Dedupe, idempotency, cleanup, leftover state that blocks identical reruns.
6. Confirm the data path completes.
   • Fault creation, task execution, artifact generation, uploads, persistence, downstream reads.
7. Confirm observability is sufficient.
   • Follow one trace / job / execution end-to-end across logs, events, and DB state.

Common Problem Buckets

• Stale contract — product is healthy, but the test asserts an old event, payload, or sequence.
• Environment drift — docs, fixtures, or assumptions no longer match the real local or CI environment.
• Version drift — wrong binary/image is running, or a remote pull overrides the intended local build.
• Auth or config mismatch — the flow starts, but later steps fail because required creds, flags, or routing are missing.
• Timing / race conditions — async consumers, startup latency, or eventual consistency make the test flaky.
• Replay / dedupe behavior — a new run is suppressed because an older valid run or cached state still exists.
• Missing result data — workflow appears complete, but artifacts, uploads, or DB rows are incomplete/absent.
• Weak observability — user-facing error is generic; the real blocker is only visible in trace or runtime logs.
• Selector/label mismatch from operators — app-label-based selectors return a partial set of services because operator-generated pods (Coherence, Kafka, Postgres controllers) carry different labels than plain Deployments. Pattern to watch: "everything installs healthy, but the selector only sees N of M services."
• Bootstrap-edge gaps — seed path populates some stores (DB, configmap, etc.) but not the one the runtime actually reads on first boot. Symptom: the data is "there" somewhere, but the process that needs it gets not found the first time and works after a manual reconciliation.
• Duplicate-submission suppression — aegisctl regression silently refuses re-submitting the same (app, namespace, chaos_type, duration) within a cooldown; symptom is duplicate submission suppressed (batches [0]). Vary spec[].app or duration to force a fresh hash.
• App name filtered by serviceendpoints — backend pre-filters live pod labels against the serviceendpoints metadata store before returning them to guided-resolve. A pod WITH the label (e.g. currency in otel-demo) still errors app ... not found; available apps: accounting, ad, cart, ... if it isn't in the endpoint map. Pick an app that IS in the "available apps" list the error prints.
• Empty parquet with .parquet suffix required — RCABENCH_OPTIONAL_EMPTY_PARQUETS values match on the full filename INCLUDING .parquet. Leaving the extension off silently fails to opt in.
• Namespace landed on the wrong slot — submit asked for <sys>14 but the actual fault ran in <sys>0. On a current backend this should fail loudly instead; if it doesn't, the running binary is too old. Rebuild and redeploy. As a workaround, re-submit via aegisctl inject guided --apply --auto and let the server pick a slot.
• Single HTTP injection rejected as a duplicate — submit returns service 'X' at positions 0 and 0 warning and zero items. Self-loops (e.g. GET / against the same app) used to trip this. On a current backend it's gone; if it appears, the backend predates the fix or the same regression has resurfaced. Workaround: change the route or method to break the self-loop.
• Auto-allocate fails with "pool exhausted" — every namespace in the system's pool is either locked or empty. Two clean fixes: pass --allow-bootstrap so the server extends the pool with a fresh slot, or pre-deploy a workload via aegisctl inject guided --install --namespace <sys>N to give an existing slot something to inject against.
• Asked for a specific namespace, got "not found in current configuration" — that namespace isn't registered in the system's pool yet. Easiest fix is to re-submit through aegisctl inject guided --apply --namespace <sys>N (current backends register it on submit). If that doesn't help, the namespace was created out-of-band; expand the system's count via the systems admin API and retry.
• Silent trace drop, one-system-only — one benchmark (typically ts* or tea*) shows zero spans in ClickHouse for the last hour while peer systems flow normally. Root cause is almost always the OTel deployment-collector HPA pegged at maxReplicas with memory_limiter returning UNAVAILABLE: data refused due to high memory usage. Java agents (TT, teastore) don't retry on UNAVAILABLE — spans dropped on the floor. Go / Node SDKs use BatchSpanProcessor + retry, so they squeeze through. Check kubectl -n monitoring get hpa opentelemetry-kube-stack-deployment-collector: if REPLICAS = MAXPODS and memory > 100% of target, raise autoscaler.maxReplicas and resources.limits.memory in aegislab/manifests/byte-cluster/otel-kube-stack.values.yaml.
• Worker rollout collides with submission — restarting runtime-worker / api-gateway while traces are mid-pipeline silently orphans them (no error log, redis queues empty, trace just stops at whatever stage the worker was handling). Pattern: traces hang at the same stage cluster-wide right after a rollout. Rule of thumb: only redeploy aegis between rounds, after terminals_round<N>.tsv fully reaped.

Output Shape

• Name the most likely blocker classes first.
• Separate contract problems from product bugs, environment bugs, and tooling bugs.
• Point to the next inspection surface instead of writing a full playbook.
• Prefer concise hypotheses like "check X because Y" over exhaustive detail.

Project-specific inspection surfaces (aegis)

When triaging in this repo, common concrete surfaces worth naming early:

• aegis/docs/troubleshooting/{datapack-schema,app-label-key,benchmark-integration-playbook}.md — consolidated E2E pitfalls.
• aegis/docs/deployment/kind/otel-collector-{cfg,rbac,externalname}.yaml — kind-profile collector manifests (k8sattributes + upsert + 3-signal pipelines).
• pkg/guidedcli — guided-first inject pipeline; /translate and GET /metadata are 410.
• etcd injection.system.* — runtime source of truth for injection config (not YAML).
• ClickHouse OTLP traces + Redis task keys — the usual "did the data path complete" inspection.
• Per-stage failure triage: datapack.build.failed → inspect the job pod logs (kubectl logs -n exp <task-uuid>-xxxx) for UNKNOWN_TABLE (collector missing pipeline), Parquet file has no data rows: X.parquet (wrong env-var filename or missing resource enrichment), or No such file or directory: /data/drain_template/*.bin (initDrainTemplate disabled but detector algo still wired in).
• --app-label-key must match the workload's actual label key (otel-demo uses app.kubernetes.io/name, not the default app). Mismatch surfaces as "app not found; available apps: " at backend submit.