Content-Description: message body text
content-type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
X-RT-Original-Encoding: us-ascii
Content-Length: 1200

Hi

I've been using permanent query logging to plain files on moderately
busy systems (1000-2000 qps, authoritative-only server) for a long
time to have a short backlog of incoming queries.  In a recent change
of our setup, we started writing the query log to an NFS file system.
The unbuffered output mode of named for logging to a file together
with a (rather weird) behaviour of the standard Linux nfs Client led
to massive iowait times which created significant jitter in query
response times.

I found that I could get rid of this jitter completely by using
buffered I/O for the query log.  The attached simple-minded patch adds
the boolean option "buffered" to the "channel" configuration phrase,
e.g.

  channel my_queries {
    file "/var/log/bind/queries" versions 15 size 2m;
    severity info;
    print-time yes;
    buffered yes;
  };

The patch simply skips the fflush() call in lib/isc/log.c and should
have no effect on channels that do not log to a file.  This works fine
on a busy system but needs some polishing to make it robust (like
flushing regularly).  I also don't know if it has any side-effects.

Can you please consider to add this functionality to BIND9?

Regards,
Alex