Having a few issues with search performance.  I was hoping there would be 
some way to get a detailed output on what phases of the search took all my 
time.  I'm planning on passing the slowlog along to Kibana to set some idea 
of what is going on on a shard level, but thought I'd put this out there 
incase there was any way to get a breakdown along with the query response. 
 i.e. I had to wait on shard X due to Y.  Plus for highlighting any flaws 
in what I'm doing.

Setup: 6 data nodes, 256GB, 8xSSD, 60GB heap, G1GC

Main search index: ~8TB + 1 replica.  ~600mil docs (news articles)

Other indices: 1TB + replicas, ~1bil docs

Indexing: 150 docs/sec

Querying: Full text, date histogram, faceting.

Problems:

   - Some of our queries are quite large.  Large near queries, up to 100 
   terms or'd together, neared with 5 terms, including wildcards.  That type 
   of thing.  They take a long time, sometimes upwards of 3 mins.  We want to 
   be able to support this.  1 second responses maybe not, but I'd say 30 
   seconds at a max..
   - These slow queries tend to prevent other queries from completing in 
   sensible times, even for simple queries.
   - There's a large variance in query times.  Sometimes 1 sec, sometimes 
   much higher.
   
I've tried to correlate to GCs and merges without success.  Ideally I'd 
like to know what were queries waiting on, from which I can make 
adjustments.  Obviously I could add more nodes to the cluster, but I'd like 
to know why I'm doing that rather than just taking a punt.  Advise/abuse 
welcome.

Harry

-- 
You received this message because you are subscribed to the Google Groups 
"elasticsearch" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/elasticsearch/1a668bd4-86a8-42d4-820c-a29d1be68762%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Reply via email to