| metadata | {"api_version":"azuresre.ai/v2","kind":"Skill"} |
| name | performance-investigation |
| description | Use for Zava Learning incidents where the platform is reachable but a compute tier is degraded — elevated 5xx from the APIs, slow quiz responses, exceptions, an API with no healthy instances, or a back-office batch job failing on the reporting-worker VM (e.g. nightly grade exports). Diagnoses from Application Insights / Log Analytics / Syslog and remediates the application or worker tier. |
| tools | ["RunAzCliReadCommands","RunAzCliWriteCommands","GetAzCliHelp","SearchMemory","ExecutePythonCode","microsoft-learn_microsoft_docs_search","microsoft-learn_microsoft_docs_fetch"] |
Zava Learning — Application Incident Runbook
Resource Group: @@RG@@. Services (Container Apps): learner-portal, course-api, assessment-api.
Quiz launch path: portal -> assessment-api -> course-api.
Confirm the network/edge is healthy first (App Gateway backend healthy, NSG clean) so you don't
misattribute an app fault. Then:
- Query App Insights failures/exceptions filtered by the failing
cloud_RoleName.
- Check Container App revision health and replica counts — an API with zero replicas serves no
requests and surfaces as quiz launch / 502 failures with a clean network path. A revision can hit
zero replicas two ways: scaled to zero, OR the active revision was deactivated (an inactive
revision runs zero replicas).
az containerapp revision list shows only ACTIVE revisions, so a
deactivated lane looks like it has "no revisions" — list with --all and inspect
properties.active / properties.latestRevisionName. The latent cause is often a scale floor of
zero in IaC (appLaneMinReplicas: 0); the durable fix restores a non-zero minimum.
- Inspect recent revisions/config changes.
- For LATENCY regressions (slow quiz responses / the latency alert) on a clean network path: pull assessment-api request durations from Log Analytics console logs (each request logs
ms=<duration>), find the step-change in latency and align it to the most recent assessment-api revision/image deployment, then inspect src/assessment-api/server.js for expensive synchronous work added on the request path (e.g. a synchronous KDF/crypto call). This is a code regression — mitigate live by rolling back to the prior revision, then fix durably with a code PR.
Reporting-worker grade-export failures
If the symptom is the nightly grade-export job failing (no quiz/portal impact — this is a back-office
batch job on a dedicated VM vm-zava-reporting-*):
- Query Log Analytics
Syslog for ProcessName == "zava-export" — the worker logs a heartbeat on
success and a FAILED line plus the raw OS error (export write error: ...) on failure. Read the
raw error to identify the failure mode.
- Correlate with disk telemetry: the
Perf table carries Logical Disk % Used Space and
Free Megabytes for the worker. A data volume (/data) at ~100% used / near-zero free is the
signal that the export writes are failing for lack of space.
- Confirm on the VM with a read command (
az vm run-command invoke ... --scripts "df -h /data").
- Mitigate live: free space on the worker's data disk (remove the accumulated backlog file under
/data/exports) via az vm run-command. Re-check that the next export run logs a success heartbeat.
- Durable fix: the worker's export retention/rotation lives in
src/reporting-worker/cloud-init.yaml
— a PR that enforces retention (or grows the data disk in infra/modules/vm.bicep) is the lasting fix.
Database-backed quiz lanes (slow loading / errors under load / auth failures)
The quiz lanes read PostgreSQL on the request path, so a DB fault produces app-shaped symptoms on a
clean edge + healthy replicas. Do NOT stop at "the DB looks slow" and report root cause unknown —
identify WHICH database surface — log-first, confirm from the lane's error detail, then optionally
verify in the database:
- Read the quiz service's console logs (
ContainerAppConsoleLogs_CL for the failing quiz-* app /
assessment-api) for the request shape: a slow ms=<duration> on /quiz/* → index/latency;
a burst of status=500 on /quiz/* (and any pool error: lines) → a backend DB failure.
- Get the EXACT backend error straight from the lane — the service returns the underlying Postgres
error in the response
detail field (it is NOT written to the logs): curl the lane's /health
(returns 503 {"status":"degraded","detail":"<pg error>"} when the DB path is broken) or
/quiz/<courseId> (500 {"error":"...","detail":"<pg error>"}). The detail names the cause
directly — e.g. password authentication failed (28P01) → bad DB secret; too many connections
/ remaining connection slots / connection-limit → pool exhaustion; a slow-but-200 /quiz with
climbing ms= → missing index. This needs no DB credentials and is the primary confirmation.
- Optionally confirm against the database itself with a read-only query (connection recipe in the
knowledge base — server/login from
az postgres flexible-server list, admin password from Key
Vault db-password, then az postgres flexible-server execute --querytext "<SQL>"):
- Slow quiz loading (query lane, db
zava_query): check the question_bank indexes —
SELECT indexname FROM pg_indexes WHERE tablename='question_bank'; and
SELECT relname,seq_scan,idx_scan FROM pg_stat_user_tables WHERE relname='question_bank';.
A missing index on question_bank with seq-scans climbing is the root cause; the durable fix is
to (re)build it (REINDEX/CREATE INDEX) — delivered as a PR.
- Errors under load (pool lane, db
zava): check the connecting role's connection limit —
SELECT rolname,rolconnlimit FROM pg_roles WHERE rolname='app_pool'; and live connections in
pg_stat_activity. A rolconnlimit of 1 (or far below the pool size) is the root cause; restore
a sane limit (ALTER ROLE app_pool CONNECTION LIMIT <n>).
Permitted autonomous actions
- Restore replica counts / restart a Container Apps revision.
- Reactivate a deactivated revision (
az containerapp revision activate) to bring back zero-replica lanes.
- Roll back to the last-known-good revision.
- Free space on the reporting-worker VM's data disk (remove an accumulated backlog file under
/data).
Code fix
If the root cause is in application source, the app code lives under src/ in @@REPO@@. After
the live mitigation (revision rollback/restart), the durable code fix is delivered as a GitHub PR
by the pr-delivery skill and recorded as a Change Request by servicenow-change-management.
Incident communication
PagerDuty acknowledgement, status/summary notes, and resolution are owned by the
pagerduty-incident-update skill.
Verification
Confirm /api/quiz/* returns 200 and the API success rate is back to baseline.