In our cluster, we have observed that the use of percentile aggregations increases the execution time of our query by multiple times (almost like 10 times! depending upon the number of documents).
I know that percentile aggregation is in experimental phase but is this a known issue or am I doing something wrong?
The percentile aggregation is expressed as:
"handshakeTime":{
"percentiles":{
"field":"handshakeTime",
"percents":[ 85 ]
}
}
Setup:
ES version : 1.7.2
Cluster : 3 c3.large master nodes and 6 m3.xlarge data nodes
Index: 3 shards, 1 replica, 24h ttl on documents
Load: 5000 docs/sec for about 24 hrs
Query: Lot of aggregations including terms, histograms and filtered aggregations