Description
glog is configured by default to call posix_fadvise with DONTNEED every time it flushes, and also configured such that any WARN or above causes such a flush. Every such fadvise call generates new IO to the log directory, journal commits, etc, and I think this may be affecting latency and stability adversely.
Attachments
Issue Links
- relates to
-
KUDU-695 contention on glog disk can block reactor threads
- Resolved