Closed let4be closed 3 years ago
Most likely related to sysctl settings as it appears only during high load(60k+ concurrent tcp connections)
Today I almost pinned it with(net.ipv4.tcp_mem = 383865 1535460 2303190
) but failed to reproduce, needs to more debugging.
Sometimes when testing on aws c5.metal I see that writers "hang" which leads to pile of unprocessed messages in buffers(particularly
metrics_task
as the heaviest clickhouse hitter)