Benedict created CASSANDRA-8763:
-----------------------------------
Summary: Trickle fsync should not be optional, but should have a
large default interval
Key: CASSANDRA-8763
URL: https://issues.apache.org/jira/browse/CASSANDRA-8763
Project: Cassandra
Issue Type: Improvement
Components: Core
Reporter: Benedict
Priority: Minor
Fix For: 2.1.4
The reason to avoid "trickle fsync" is that it permits more efficient flushing;
however once we get above a few hundred MBs, it really doesn't make a great
deal of difference. Contrarily, it can cause runaway utilisation of memory by
dirty pages in the page cache, damaging performance for other components and
potentially invoking the kernel OOM killer. I suggest we pick an amount of
memory proportional to the size of the heap (or check the actual amount of
memory in the system), and divide this by the number of flush threads we have,
and use this as the default trickle fsync interval, and we _always_ "trickle"
fsync, ignoring the ability to disable it. We only permit on override of the
size of the interval (if you want to disable it, you can set the interval
absurdly large).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
