vaquarkhan Incident Triage And Pipeline Recovery
Overview
Use this skill when production behavior is already wrong and the team needs controlled recovery. It helps agents contain blast radius, diagnose quickly, restore trust, and avoid making the incident worse.
When to Use
- failed production runs
- bad or partial publishes
- corrupted incremental state
- freshness SLA breaches
- emergency rollback or replay decisions
Do not jump to fixes before stabilizing the system and understanding impact.
Workflow
Contain the incident. Decide whether to:
- pause schedules
- block downstream publish
- isolate bad partitions
- notify owners and consumers
Pull live signals before mutation when possible. Load
mcp-data-observability-integrationto inspect lag, run state, or Spark stage metrics before reruns or replays.Classify impact. Identify:
- affected datasets
- time window
- downstream consumers
- whether data is late, missing, duplicated, or wrong
Determine the safest recovery path. Options include:
- rerun
- replay
- rollback
- partial correction
- full backfill
When replay or backfill is chosen, load
safe-backfill-and-replay-orchestrationand drafttemplates/backfill-plan.yamlbefore execution.Validate recovery before reopening publish paths.
Record the incident and add a guardrail. A fixed incident with no prevention work is unfinished.
Turn high-value incidents into repeatable resilience tests. Load
skills/data-resiliency-testing-and-failure-injection/SKILL.mdorreferences/data-resiliency-testing-patterns.mdwhen the same failure mode must be prevented from surprising the team again.
Common Rationalizations
| Rationalization | Reality |
|---|---|
| "Let us rerun everything now." | Blind reruns can duplicate data or destroy evidence. |
| "We can clean up downstream later." | Downstream trust loss is often harder to recover than the pipeline itself. |
| "The root cause can wait." | Without a guardrail, the same incident often returns quickly. |
Red Flags
- no attempt to contain blast radius
- reruns happen before impact is understood
- downstream consumers are not informed
- recovery succeeds but no new guardrail is added
Verification
- Impact scope and affected consumers are identified
- The chosen recovery path is safer than the alternatives
- Publish is reopened only after validation
- Post-incident guardrails or checks are added