|
I verified this is related to cache hit/miss ratio. It appears the steadiness before is because we were getting 100% hit rates consistently, and then at some point we were unable to get 100% consistently, generally get some were between 95% and <100%, most likely due to increased prunning activity. In the case of per-variant caches this is expected as we prune a little bit a lot more.
However per variant caches were not introduced until feb 22. So that is not likely the original cause. I suspect earlier than that we were simply still prunning more as we were having size and management problems with the cache (which instigated the move to per-variant cache), and I assume size pressure was causing increased prunning activities.
|