[valentynkit]

Quick one on the benchmark: was the 2.8GB peak virtual or resident? Go reserves a large virtual arena it mostly never faults in, so RSS tends to be a fraction of the virtual peak, and if Postgres headroom was getting squeezed off the virtual number you were sizing against memory the kernel never actually charges for.

[__s]

Correct. We tune overcommit so postgres reliably returns out of memory. It becomes complicated to accurately tune overcommit for every AWS instance type. We configure GOMEMLIMIT/cgroups but those are about RSS. Outliers come together: instances running queries out of memory on our service tend to also be pushing other resource limits, causing wal-g & prometheus exporters to start having more erratic memory usage at the worst time

This helps on both ends of the cost spectrum. Large 64 core instances are where our heuristics fall off the most as variance increases, & tiny instances with 8GB of memory can use every 100MB of RSS we can get

[nasretdinov]

You probably could limit the bloating of Go programs by setting GOMAXPROCS to something like 1 or 2 on smaller machines, but then again you wouldn't get the best performance. So IMO good call here to rewrite it in a language without GC.

[pjmlp]

Or make better use of the language capabilities to use value types.

[nasretdinov]

Value types do indeed reduce the GC pressure, however large Virtual Memory usage (in my understanding) is mostly due to large thread count by default (one per core), and some unavoidable GC overhead in general.