User set suspend_timeout to 20s after which endpoint frequently restarted and at the same time became unavailable for the deployed app for 15-30 minutes. Seems like proxy -> compute is one more place where we miss timeouts. However there is no KAs in main postgres protocol, so probably we just need to configure TCP keepalives/timeouts on the kernel level.
User set suspend_timeout to 20s after which endpoint frequently restarted and at the same time became unavailable for the deployed app for 15-30 minutes. Seems like proxy -> compute is one more place where we miss timeouts. However there is no KAs in main postgres protocol, so probably we just need to configure TCP keepalives/timeouts on the kernel level.
copied from https://neondb.slack.com/archives/C04DGM6SMTM/p1684338654994679