Bug: no coredumps generated if Postgres segfaults

[bayandin]

Environment

Initially spotted on Staging and Production, haven't tried to reproduce it locally.

Steps to reproduce

• Make Postgres segfault. You can find an example in [0]

Expected result

We have a logic in compute_ctl that checks core dumps and prints stack traces: https://github.com/neondatabase/neon/blob/350865392cd5bb38eb2a7ff6f45b36f11ac8a911/compute_tools/src/compute.rs#L1000-L1010

I expect, segfaulted postgres to produce a core dump, and compute_ctl prints the stack trace from it.

Actual result

• Postgres doesn't produce core dumps
• The reason for it might be in neonVM setup, we expect (as mentioned in the compute_ctl comment, and how it's configured for k8s pods):

  /proc/sys/kernel/core_pattern -> core
  /proc/sys/kernel/core_uses_pid -> 1
  ulimit -c -> unlimited

But in neonVM, we have:

/proc/sys/kernel/core_pattern -> core
/proc/sys/kernel/core_uses_pid -> 0
ulimit -c -> 0

Other logs, links

• [0] https://neondb.slack.com/archives/C03QLRH7PPD/p1706797854242749?thread_ts=1706797255.300329&cid=C03QLRH7PPD
• [1] https://neondb.slack.com/archives/C03TN5G758R/p1706866472714979?thread_ts=1706801092.664529&cid=C03TN5G758R
• [2] https://neonprod.grafana.net/goto/Flr52spSg?orgId=1

[stradig]

• Testing that dumping backtraces works
• Kill switch
• Use sysctl setting on the VM spec for enabling core dumps
• Implement ulimit