_msearch does not distirbute queries across the cluster

[alepuccetti]

I am running kibana dashboards and kibana uses _msearch to bundle all the queries of a dashboard in one request. I have 3 nodes and my dataset is one index with 1 primary shard and 2 replicas.

I would expect elasticsearch to distribute every query across the nodes but it does to load balance the queries. If I run all the queries separately using _search on the command line they are distributed across all nodes as expected and return much faster because it uses all the nodes instead of a single one.

I do not know if this is an expected behavior or a bug, but I believe that would be much better if elasticsearch would be able to multiplex queries submitted with _msearch across all the nodes to exploit the parallelism.

[jimczi]

This is certainly due to the fact that Kibana adds a preference to the _msearch request that is sent to es. This preference is set to the session id of the Kibana user and ensures that all search inside a dashboard use the same shards/replicas:
https://www.elastic.co/guide/en/elasticsearch/reference/current/search-request-preference.html
Without preference, _msearch just send each individual query independently so they could all potentially hit a different replica.

[alepuccetti]

@jimczi Thank you very much for the explanation. I was not aware of preferences. I was pretty confused about the fact that _msearch would not distribute queries, it seems very backward.

I will raise this issue on at the kibana repository.

[songlong]

msearch performance low concurrence request?