JVM Heap usage spikes


I have a cluster 2.2.1 with 3 nodes. 4 CPU / 14 Gb RAM / SSD. Java(TM) SE 1.8.0_66-b17
ES_HEAP is 7 gb.
RefreshInterval set to 30s.
mlockall: true

I'm bulk indexing documents. There is no search query others than for Marvel monitoring.
I see a lot of spikes on the JVM Heap usage - https://infinit.io/_/3jvRpGs

What could explain the behaviour. Is it normal ? Any parameters to check ?

I have another cluster, with the same configuration and i don't see theses spikes.

Thanks for your help!


Here is a new screenshot: https://infinit.io/_/p3h5YDG

Between 16h40 and 17h40 - Bulk index - And after the cluster is idle. Why the heap is increasing and then dropped ?

If you're indexing a ton with a 30s refresh, that's not terribly surprising.

Despite the transaction log, Lucene is building up data in memory to be flushed to a new segment. The point of the transaction log is to act as a backup to this process, if Elasticsearch goes down, you can replay the index updates.

So yes I'd expect quite a bit of JVM churn when writing, and probably a lot of in-memory objects with a 30s refresh interval.

So it could be interested to reduce this time to reduce the quantity of data in memory but. A lot of refresh interval is a costly operation and will have an impact on performance right ?