- Lobsters doesn't use Puma's
preload_app!
because it relies on phased restarts. This seem to cause massive GVL contention after a deploy. - Would be worth trying to exclude that data, but aside from the very begining it's not clear where deploys are in the dataset. Need to check with @pushcx.
- The IO heavy requests are in the higher percentile.
- IO heavy requests are the one suffering the most from GVL contention.
- Which makes sense, the more IO you do, the more you need to re-acquire the GVL, so the most you suffer if it is contented.