Sporadic 503 errors when behind proxy and high traffic under AWS ECS

[steve-chavez]

Problem

On a PostgREST configured under ECS undergoing high traffic, sometimes the proxy (envoy) logs a 503 coming from the PostgREST upstream, while no log is produced on PostgREST side.

This case has turned to be real hard to debug so far, the new logs added on #4646 don't match the 503 happening.

(envoy shows a upstream_reset_before_response_started{connection_termination} in this case)

It's also very hard to reproduce, my attempts in PostgREST/postgrest-benchmark#4 haven't worked so far.

Notes

• This is the same environment as the one mentioned on Thundering herd problem in PostgREST cluster on AWS ECS #4642

[steve-chavez]

To help with this case we could try logging whenever a connection is closed: settingsOnClose.