Closed nedtwigg closed 3 years ago
This PR kinda works, but we'll need to do more. Example memory trace after this PR was merged:
Timewise, the problem seems to come when searches are performed. Which is a bit weird, because we try to eagerly warmup the Lucene index:
Seems like maybe Lucene is doing some caching of its own as it gets used for novel searches. That is bad, because we aggressively cache search results at the HTTP level anyway, so internal caching (if that is what's happneing) is totally redundant.
We're running at about 15% swap, which is no good. Here's a few things to bring that down.