A poorly-behaved client could use keepalive requests to monopolize Puma's reactor and create a denial of service attack.
If more keepalive connections to Puma are opened than there are threads available, additional connections will wait permanently if the attacker sends requests frequently enough.
This vulnerability is patched in Puma 4.3.1 and 3.12.2.
Reverse proxies in front of Puma could be configured to always allow less than X keepalive connections to a Puma cluster or process, where X is the number of threads configured in Puma's thread pool.
If you have any questions or comments about this advisory: