Delayed scheduling of BuildHooksWorker
While rolling out #326095 (closed) we noticed an increase in the error rate ratio for jobs.
It seems that the actual scheduling_latency_s
is 0.003s
, and that the replication process couldn't be finished this fast.
We saw that more than 50% of the scheduled BuildHooksWorker
jobs were retried because the replica was not up to date.
In the case of workers with data_consistency
set to delayed
and sticky
, we could delay their execution a few seconds in the future in order to give the replication process more time to complete.
In order for this to work, we also need to prevent overwriting the DB WAL location if it's already provided
Edited by Nikola Milojevic