0695 — Partition alerts calculate retained duration
Edit this page
Superseded by 0699 for configuration shape; the retained-history behavior below remains in effect.
Context
The shared evaluation result [0694] separates healthy from pending and missing evidence. The next implementation chunk connects partition count to the retained history supplied by metrics [0693], using the storage-time contract [0690].
Decision
- Every Evaluator accepts the consumed JobPayload. PartitionCountAlertConfig adds Pending: Duration and MaximumGap default to two minutes; Disabled permits immediate activation with fresh evidence. Older payloads resolve these defaults.
- Partition-count Evaluate reads one database-time window through metrics. GetMeasurementHistory returns that evaluation time and the existing compaction CreatedAt-window read, ordered by CreatedAt/id descending without a row cap. Retention must exceed the requested window or be disabled.
- Read Duration + 2*MaximumGap, or MaximumGap when disabled. Reject nonpositive durations and overflowing window arithmetic before reading. Highest id wins timestamp ties. Healthy/unusable samples and excessive gaps break the span; activation requires the newest-to-earliest sample span to reach Duration.
- Fresh healthy evidence resolves through Record. Missing, stale, or unusable evidence returns insufficient_evidence, leaves the alert unchanged, and counts as a failed-topic check. Read/decode errors keep their existing error path.
- Registration warnings use the same Evaluate with pending disabled and require fresh evidence. They log insufficient evidence and never produce alerts.
- Other conditions keep live reads until metrics owns their required evidence. Alert cadence stays hourly; no measurement production moves into alert code.
Consequences
An overnight gap starts a new unhealthy span. Waiting alone cannot turn one sample into an active alert. Delayed writes count as fresh storage-time evidence. The observed span is calculated again on each check; no timer, cursor, or pending state table exists. Metrics’ unused absolute-time wrapper is replaced by the database-time history read; the underlying compaction method is unchanged. Race/unit checks cover the calculation and payload policy. Database integration and alert-lab checks remain deferred to the review checkpoint.