ltamasi
ltamasi
Just to confirm: you are sharing a single `WriteBufferManager` object among all 1000 database instances, right?
Chances are flushes can't keep up with the incoming writes. You could try increasing the number of flush threads/jobs; or, alternatively, you could set the `allow_stall` option to make `WriteBufferManager`...
> I don't think we should introduce small _passive_ performance regressions (affecting performance when new feature is not enabled) for features that haven't been proven necessary. > > One possibility...
> This is among the hottest of hot code, and this change introduces strictly more tracking to be done regardless of enabling the feature. Just look at LRU_Remove code, which...
> The low pri pool was added as "mid-point" insertion. The purpose is to make sure occasional full scans (logical backups, shard movements, administrative scans, etc) that touch a large...
> The answer to my question is yes, or no? Sorry I'm not sure I understand your question. If you have all three pools enabled (i.e. high ratio > 0,...
> Metablock in high pri pool is just a paranoid optimization, rather than the main purpose of the two-pool approach. One data point regarding that: technically the root cause of...
@siying @mdcallag Please let us know if you have any concerns about landing this considering the performance numbers above. Thanks!
Did a test with BlobDB to quantify the effects of the bottommost cache priority. After populating a smallish (fits in the OS page cache) DB with 1KB-sized values using `fillrandom`,...