With backend serving many clients with widely varying performance profile of individual requests when latency spikes happen there is no particular hot loop. Just many go routines each doing reasonable thing but with a particular request pattern hitting pathological case of GC.