Indexing performance of V2.3.5

Hi all, we are evaluating indexing performance of V2.3.5, comparing with V1.7.5, the cluster nodes of the two versions are the same as:

o 1 client node
o 8 data nodes
o 3 masters nodes
    o standard analyzer
    o the same source documents on the same index
    o same mappings and settings including shard number and replicas

we are using scroll API on a jump box node inside the cluster to do indexing test and getting below data:

in V1.7.5 indexing time of tests by locales:


ar-eg, 01:49:51

ru-ru, 01:00:08

en-gb, 00:54:31

tr-tr, 00:45:51

in V2.3.5 indexing time of tests by locales:

en-us, 03:13:30

ar-eg, 02:37:54

ru-ru, 01:27:39

en-gb, 01:15:35

tr-tr, 01:05:26

as can be seen indexing time of all tests grows in V2.3.5 test.

I saw there's change about doc value default change write on disk in V2.3.5 which will bring up to 25% slower performance:

but above performance data in our testing shows it's much more than 25%. (like en-us one it's 50%), any one can help explain why this happened in our testing?

or is there any default changes in V2.X we need to care in this comparison?


In the interests of improved durability there was a change to the settings around fsync frequency. If you want to optimise for performance you can relax these settings but be aware of the implications - see