Mark Payne created NIFI-2681:
--------------------------------
Summary: Avoid caching Provenance Index Searchers
Key: NIFI-2681
URL: https://issues.apache.org/jira/browse/NIFI-2681
Project: Apache NiFi
Issue Type: Bug
Components: Core Framework
Reporter: Mark Payne
Assignee: Mark Payne
Priority: Critical
Fix For: 1.1.0
In NIFI-2600 and NIFI-2452, we addressed two bugs where the Provenance
Repository closes a cached IndexSearcher too soon. The IndexManager keeps the
searchers cached in an effort to offer better performance when performing a
Provenance Query. This was done because it was recommended in the Lucene
documentation. However, we occasionally still see nodes crashing with segfaults
due to the Lucene Searching. We should update the Persistent Provenance
Repository to stop caching Index Searchers in order to trade a slight
performance improvement for significantly better reliability.
Playing around with the idea in order to test it out shows very favorable
results. On a system where I could cause a seg fault almost every time that I
ran a large provenance query, I updated the code to no longer cache the readers
and saw perfect stability with no noticeable performance degradation.
I will cleanup the code and submit a PR for these changes.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
