High CPU load for unknown reasons
Closed this issue · 2 comments
Occasionally we see sudden load spikes, typically resulting in the instance being killed because it doesn't respond to the health check anymore. The CPU load will raise to 100% without obvious reasons; in particular, we don't see unusual patterns in the number of requests. We might be locked in some inefficient algorithm or some sort of blocking operation.
First attempt at an explanation: The algorithm for filtering out modified comments is quadratic, so it might cause high load in some cases. However, the table of frozen comments (zappr_data.frozen_comments) is flushed after a PR is processed, so we can't reconstruct the state at the time of the incident. At present it only contains some 70 entries...
Kinda fixed with #511.