The fix for CVE-2019-16770 was incomplete. The original fix only protected
existing connections that had already been accepted from having their
requests starved by greedy persistent-connections saturating all threads in
the same process. However, new connections may still be starved by greedy
persistent-connections saturating all threads in all processes in the
cluster.
A puma server which received more concurrent keep-alive connections than the
server had threads in its threadpool would service only a subset of
connections, denying service to the unserved connections.
This problem has been fixed in puma 4.3.8 and 5.3.1.
Setting queue_requests false also fixes the issue. This is not advised when
using puma without a reverse proxy, such as nginx or apache, because you will
open yourself to slow client attacks (e.g. slowloris).
The fix is very small. A git patch is available here for those using
unsupported versions of Puma.