A restart clears a queue; a retry makes a flaky test pass. Neither explains why the failure happened. Cause analysis investigates recurring or costly failures so the next change addresses the condition that produced them, rather than repeatedly treating the symptom. Before investigating, weigh how often the failure occurs against its cost in time, trust, and risk. Then ask: What allowed it? What would prevent it from recurring? section 8.0 section 8.1
Combine four tools rather than expecting one to tell the whole story:
- Is/Is Not compares where and when the failure occurs with where and when it does not, narrowing what changed. section 8.2.5
- 5 Whys traces a chain of explanations, testing each answer with another question. section 8.9
- A fishbone diagram checks categories of possible causes that the first chain may have missed. section 8.9
- Pareto analysis ranks failures by frequency so effort goes to the dominant problem first. section 8.9
A notification failure
Suppose emails stop reaching Pro-plan users in the EU after Tuesday’s deploy, while Free-plan and US-East deliveries succeed. That boundary points toward the deploy and the affected onboarding path. A 5 Whys chain reaches a missing field mapping in a migration specification generated from the US schema. A fishbone check then finds two more incorrectly mapped fields and a staging environment that lacks the EU schema. Restarting the service may clear the queue, but it cannot correct those conditions. section 8.2.5
Finish the investigation with countermeasures. Fix the immediate defect, prevent the same class of defect through safer defaults or automated checks, and detect conditions that still escape. For an undersized monetary database column, that could mean a schema migration, a CI schema rule, and a boundary alert. An alert alone is not prevention if someone must keep intervening. Evaluate proposed patches—including AI-generated ones—as hypotheses: does the change remove the cause, or merely hide the failure? Record the evidence and actions in a concise A3 report, and review incidents without blame so people can describe what actually happened. section 8.1 section 8.4 section 8.6 section 13.2
Table 8.3 Countermeasure versus quick-fix evaluation
|
Change
|
Type
|
How to tell
|
|
Schema migration prevents invalid data
|
Countermeasure
|
The invalid state cannot exist — the database enforces it
|
|
Rate limiter caps request volume
|
Countermeasure
|
Overload is structurally blocked before it reaches the
service
|
|
Retry loop hides intermittent failures
|
Quick fix
|
The failure still happens; the retry just hides it
|
|
Input sanitization works around bad upstream data
|
Quick fix
|
The upstream service still sends bad data every time
|
|
Updated deployment template with correct defaults
|
Countermeasure
|
The standard work changed, future engineers get the right
config without thinking about it
|
|
Circuit breaker with fallback mode
|
Countermeasure
|
The system handles the failure by design, not by luck
|
Compare proposed changes by whether they block or contain the failure, or leave it happening behind a workaround. section 8.5



