Finishing a feature or fixing an incident does not necessarily improve the next delivery. The learning principle asks whether feedback changes the way work is done. If the same failure pattern returns on the next feature, the team solved a local problem but did not improve its delivery system. section 1.2.2
Learning runs through the delivery loop, not only after production incidents. Test failures, review delays, missing staging data, runtime behavior, and user feedback can all expose assumptions worth revisiting. Look for recurring gaps: unclear requirements, frequently violated standards, absent production characteristics in staging, or defects that repeatedly escape checks. section 1.2 section 1.2.5
Turn an observation into a change the team can use next time:
- Trace what happened and where an earlier check could have caught it. Use a blameless incident review when the signal comes from production. section 1.2.4 section 15.6
- Make a specific change to standard work: update a checklist, runbook, monitoring signal, test, or automated check. Track prevention work so it is completed rather than merely proposed. section 15.6 section 1.2.3
- Make the reason visible. A decision log can connect a current rule to the incident and reasoning that produced it. section 15.6
From one crash to a review standard
A background job exhausts memory after loading an entire result set, causing a 22-minute outage. The immediate fix changes that job to use a cursor. The learning step is broader: the team finds and fixes four other jobs with the same pattern, then adds a code-review question about bounding results or using a cursor. One debugging discovery becomes a shared check for future work. section 15.6
Table 15.4 Incident review output template
|
Field
|
Content
|
|
Incident
|
Background job OOM crash, 22-min outage
|
|
Root cause
|
Unbounded SELECT loading the full result set into memory
|
|
Immediate fix
|
Switch to a server-side cursor for the daily report job
|
|
Systemic fix
|
Audit and fix 4 other jobs with the same pattern
|
|
Standard work update
|
Add cursor/pagination check to code review checklist
|
|
Owner
|
Assigned developer, due by next Thursday
|
The incident-review template separates the immediate fix from the systemic fix and standard-work update, and assigns an owner and due date. section 15.6
A useful closing question is: “What could have caught this earlier, and how do we make that check automatic?” For a performance problem found only after deployment, that might mean load testing database changes against production-like data before staging. The aim is not more activity; it is fewer repeat failures across the delivery loop. section 1.2.4 section 1.2.3
