Key stats
- 3 matching fields: root cause, failed control, and service scope
- 2 action types: containment and prevention
- 2 rates: recurrence among all incidents and recurrence among severe incidents
- 1 rule: a completed action is not a prevention action until a check proves it
Key takeaways
- Match incidents on cause and failed control, not on service name alone.
- Report recurrence as a share of incidents with a recorded cause and show the excluded records.
- Separate containment from prevention, because a faster alert does not stop the failure.
- Inspect the original action when a cause returns, and strengthen it with a test or control.
- Review recurrence by service and cause, never as an individual score.
Question and scope
This report examines incident recurrence as a support signal in outsourced application maintenance. The unit is one incident that shares a root cause with an earlier incident and could have been prevented by the earlier action. The scope covers production incidents and severe support escalations with a recorded cause. It does not count a new incident that merely affects the same service.
The question is whether recurrence can show whether post-incident actions actually changed the system. A team can close actions quickly and still repeat the failure. The report treats recurrence as a test of action quality, not as a count of mistakes. The measure is useful only when the team is willing to link the second incident to the first.
What recurrence means
Recurrence requires a link between two incidents. The link can be the same root cause, the same failed control, or the same missing monitor. A shared service name is not enough, because many unrelated failures can affect one component. A shared component is a place to look, not a cause by itself.
Some incidents recur because the underlying cause is accepted risk. In that case the recurrence is expected and should be labeled. The useful signal is an unplanned recurrence after an action was marked complete. That gap is where the learning failed.
Matching incidents
Use a short cause taxonomy and a structured field in the incident record. Match on the cause and the failed control, then review the pair. A text search alone will miss differently worded incidents and create false matches. The link also gives the next responder a starting point, because it turns a label into a short history that can be read in a minute. That history is often enough to choose the next action without reopening the whole review.
Keep a small set of examples for each cause so reviewers apply the taxonomy consistently. Review a sample of matches each quarter. If two reviewers disagree, refine the definition rather than forcing a match. A false match is worse than a missed match because it sends the review toward the wrong action.
A short cause list is easier to apply than a detailed taxonomy. Start with five or six causes that match the team's actual incidents. Add a free-text note for the specific evidence and review the list each quarter. If a new cause appears repeatedly, add it; if one is never used, remove it. The taxonomy should follow the work, not an ideal process.
The denominator
A raw recurrence count grows with incident volume. Report recurrence as a share of incidents with a recorded cause, and separately for severe incidents. The second view is often more useful because severe incidents receive a deeper review. Severe incidents deserve a separate view because their review depth is usually different.
Exclude incidents with no cause record from the denominator, but report how many were excluded. A high exclusion rate is a data-quality finding. It may mean the review process is skipping the cause step or that the record is hard to complete. A high exclusion rate is a reason to simplify the record, not to pressure responders to guess.
The match should be recorded with a link to the earlier incident, not only a label. A link lets the next reviewer read the original action and its evidence. It also makes the count auditable without a separate spreadsheet. If the link is missing, the match is a guess. Treat the link as part of the definition, not an optional note.
Action quality
An action can be specific, owned, and testable, or it can be a general instruction to be more careful. Recurrence is more likely after the second kind. Review the action text and ask whether it names a change, an owner, and a check. An action that cannot be checked is a note, not a control.
Separate containment actions from prevention actions. A monitor that detects the failure faster is useful, but it does not prevent recurrence. Record both types and measure them separately. A team can be good at detection and weak at prevention.
A prevention action should name the check that will fail if the cause returns. That check can be a test, a monitor, or a review question. It should be owned by someone who can maintain it. An action without a check is a hope that the team remembers. The review should ask for the check before marking the action complete.
Attribution and blame
Recurrence is easy to turn into a blame metric. A team that fears blame will avoid linking incidents or will label a repeat as a new cause. The result is a cleaner chart and a worse system. The team should hear that the review is about the system, not about who was on call.
Report recurrence at the service and cause level, not by individual. Discuss the action quality, not the person who wrote it. The goal is to find the gap between a completed action and a changed system. A blame-free review still needs a named owner for each action.
A practical protocol
Add a cause and an action-quality field to the incident record. Match new incidents against the last two quarters. Review the matches monthly with the support lead and the buyer's service owner. Report the share and the excluded records.
Pick one recurring cause and inspect the original action. Strengthen it with a test, a monitor, or a code change. Measure recurrence for that cause for a quarter. The protocol is working when the same cause does not return after a completed prevention action.
The review should also protect the responder who reports a repeat. A person who says the same failure has returned is providing useful evidence, not admitting a mistake. Thank the reporter and inspect the action. A team that punishes the messenger will lose the signal and keep the incident.
A worked example
A service has four incidents in three months, all involving a queue that backs up after a deploy. Each incident has an action to add a dashboard note. The note is added every time, and the queue still backs up. The first review should agree on the cause list, not on a target rate.
The team replaces the note with a deploy check that measures queue depth and a rollback trigger. The next deploy pauses automatically when the depth crosses the threshold. No incident occurs, and the action is now a control rather than a reminder. If the cause returns, reopen the original action rather than writing a second identical one.
Data to collect first
Record the incident date, the service, the cause, the failed control, the action, the action owner, and the completion date. Add whether the action was containment or prevention. Keep the fields short enough that a responder can complete them during the review. Each note looked reasonable in isolation, which is why the pattern was easy to miss.
Collect two quarters if possible, because recurrence needs history. Check a sample of old incidents to confirm the cause field is usable. If the old records are thin, start the taxonomy now and mark the earlier period as incomplete. The new check turned a repeated reminder into a stopping condition.
Counterevidence and limits
Some recurrence is acceptable when the cost of prevention exceeds the risk. A rare, low-impact incident may not justify a large engineering change. Record that decision explicitly rather than leaving the action open forever. Two quarters of history is usually enough to see a repeat without waiting a year.
A low recurrence rate can also mean the matching is too narrow or the team is not reporting. Sample the incidents and the links each quarter. The metric is a prompt to inspect action quality, not a proof that the system is safe. An incomplete early period should be labeled rather than silently mixed with the new data. The label should be applied at the time of the match, not reconstructed later.
Evidence table
| Signal | What to inspect | Owner |
|---|---|---|
| Outcome | Acceptance evidence for the bounded task | Task reviewer |
| Control | Access, test, and approval boundary | Internal owner |
| Handoff | Open risks and next decision | Next owner |
Good distributed work is observable at the handoff: the result, evidence, limitations, and next owner are explicit.
Frequently asked questions
Is every repeat incident a failure?
Not always. Some repeats come from accepted risk or a genuinely new cause in the same service. The useful case is an unplanned repeat after a prevention action was marked complete.
How far back should matching go?
Two quarters is a practical start. It covers the life of most post-incident actions without forcing the team to review years of records.
Who should own the recurrence review?
The support lead with the buyer's service owner. Keep the discussion at the service and cause level, not the individual level.
