Github user joewitt commented on the issue:
https://github.com/apache/nifi/pull/2416
@markap14 my observations of running a build just before this PR and with
this PR is that for a bencmharking/ff repo stressing flow their is effectively
no change in performance.
Basic GenerateFF to UpdateAttr with 25 ms batching, 4 generation thread, 6
update thread, 6000 as queue limiter
The rate was steady at 25M/5mins or roughly 83K events/sec end to end.
WALI then was dealing with at least 170Kevents/sec. After switching to the new
implementation it leveled off slightly higher . GC, system load, io
utilization in both cases was very consistent and very similar.
The write rate to the now single partition was around 20MB/sec. I had a 15
sec checkpoint window.
This is heading in a nice direction in terms of simplicity and closing the
window over which different partition checkpoints could be susceptible to
different OS flush timing. Encouraging stuff.
---
