Repository navigation
Run prod's canary and indexer offset from staging's - #62
Merged
Merged
Conversation
Both namespaces took the base's schedules, so their canaries started on the same minute every 15 minutes, on a node whose two vCPUs are one physical core. Measured 2026-09-21 from Prometheus on the node: - 240 canary searches per namespace in six hours, the only search traffic; 235 of prod's and 213 of staging's took over 3 s. - Mean 3.7 to 4.0 s a search, 3.55 to 3.82 s of it in the reranker, p95 about 4.9 s. No CPU throttling; each API pod peaked near one core, and the node averaged 10 percent busy. - The same ten canary questions run alone, between scheduled runs: 1.88 s a search, 1.79 s of it reranking. So the pods were only ever slow together, and SearchLatencyHigh opened and closed fourteen GitHub issues (#48 to #61) that afternoon measuring exactly that. The fix is the cause, not the alert: prod's canary moves to seven past the quarter hour. Prod's indexer moves from 06:00 to 06:30 by the same mechanism, which the k8s README already listed as a follow-up and which has not itself been measured.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
SearchLatencyHighopened and closed 14 GitHub issues (#48–#61) between14:10 and 21:25 UTC on 2026-09-21 while the node sat idle between game-day
scenarios. The cause was not load and not the alert. It was the schedule.
What Prometheus showed
The canary is the only search traffic between deploys, and both namespaces
took the base's
*/15schedule, so twenty reranked searches started on thesame minute on a node whose two vCPUs are one physical core. There was no CPU
throttling, each API pod peaked near one core, and the node averaged 10% busy:
the two pods were only ever slow together. The "alone" column is a one-off
canary Job run between scheduled runs, with prod idle in the same window.
The change
deploy/k8s/overlays/prod/schedules.yamlpatches two schedules in prod only:7-59/15, seven minutes after staging's; each run takes about half a minutedeploy/k8s/README.mdalready listed as a follow-up; that one is not measuredStaging is unchanged. The alert rule is unchanged too, because it was
reporting something real.
Checks
Both overlays render with
kubectl kustomizeand passkubeconform -strict(20 resources). The rendered prod CronJobs carry the new schedules and
staging's carry the old ones. Merging this deploys it; the evidence it worked
is the next few hours of canary latency, which I'll read back from Prometheus.