compaction_read_cost [SQL0095]
Edit this pageAs it appears on __system.alerts:
compaction_read_cost
A compacted stream holds enough partitions that replaying a never-superseded key is a long scan.
A consumer replaying a key nothing has superseded scans from that key’s
partition to the current tail. The cost grows with the partition count,
about 10µs per partition, and never amortizes. The hourly check publishes
an active alert after retained observations establish the configured pending
duration above the threshold (two minutes by default);
the alert’s data carries partition_count and threshold.
Compact more aggressively or lower retention so old partitions drop, which bounds the replay. The next run that finds the count under the threshold produces the resolution under the same key. Count and compaction applicability come from the same collected measurement. Missing or stale evidence cannot resolve the alert; see pending rules.
sqlstreams alert listLooking for a different code? Search every thread.