Elasticsearch version:
1.7.3
JVM version:
1.7
OS version:
CentOS6.5
Description of the problem including expected versus actual behavior:
Steps to reproduce:
1.ingest more than 3 billon event into the index.
2.send a percentiles aggregation for those event
3.retry the query for severy times
Provide logs (if relevant):
the right result:

and the error result:
