Keeping same search rate on one index spikes CPU in data nodes vs distributing the same load in two indices

We tested out two scenarios:

  • 100 TPS search traffic was pointed to one index
    20% CPU usage in data nodes and Threadpool search queue spikes observed. Application side there was increase in TP90 by 3x

  • 80-20 TPS search traffic pointed to two indices
    CPU was at 4% with no spikes and no Threadpool search queue spikes observed

Both indices have same number of shards and replicas.

Can anyone help understand why CPU spike behaviour is seen (same load) with to one index vs two indices?

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.