This benchmark started failing recently (this weeks microbenchmark run).
It only fails when conns=x exceeds a certain threshold hold, I suspect around ~1000.
This snippet below from the benchmark is how conns is determined.
numConns := runtime.GOMAXPROCS(0) * parallelism
Upping the cpu count --test-args='-test.cpu 24' simulates the environment the microbenchmarks run on, and increases the conns significantly that reproduces this issue. This could possibly just be a system resources issue, but needs further investigation to confirm, and decide if a maximum / cap should be put on this value.
I ran into this, too, when creating the benchmark (see this comment). Sadly changing net.ipv4.ip_local_port_range didn't seem to help. I'll add an upper limit to the number of connections.
Notes:
This benchmark started failing recently (this weeks microbenchmark run). It only fails when
conns=x
exceeds a certain threshold hold, I suspect around ~1000.This snippet below from the benchmark is how
conns
is determined.Upping the cpu count --test-args='-test.cpu 24' simulates the environment the microbenchmarks run on, and increases the
conns
significantly that reproduces this issue. This could possibly just be a system resources issue, but needs further investigation to confirm, and decide if a maximum / cap should be put on this value.To reproduce:
Output:
Jira issue: CRDB-38666