lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Toke Eskildsen>
Subject Re: &fq degrades qtime in a 20million doc collection
Date Fri, 15 Jan 2016 22:43:14 GMT
Anria B. <> wrote:
> Thanks Toke for this.  It gave us a ton to think about, and it really helps
> supporting the notion of several smaller indexes over one very large one,>
> where we can rather distribute a few JVM processes with less size each, than
> have one massive one that is according to this, less efficient.

There are not many clear-cut answers in Solr land...

There is a fixed overhead to running a Solr instance and you need to have some wriggle room
in the heap for temporary peaks, such as index updates. This calls for a few or only a single
Solr instance handling multiple collections.

On the other hand, large Java heaps are prone to long stop-the-World garbage collections and
there is the memory overhead when exceeding 32GB.

Locally we run 50 Solr instances with 8GB heap each, each holding a single shard. At some
point I would like to try changing this to 25 instances with 15GB and 2 shards or maybe 12
instances with 28GB and 4 shards. I will not exceed 31GB in a single JVM unless forced to.

- Toke Eskildsen

View raw message