[Spread-users] mod_log_spread and RELIABLE_MESS

[Theo Schlossnagle]

Daniel Rall wrote:
> John David Duncan <jdd at greatschools.net> writes:
>>So here's a simple question:  is there any advantage in having
>>mod_log_spread send messages as UNRELIABLE_MESS rather than
>>RELIABLE_MESS?
> 
> 
> Indeed.  Without flow control built in to the Spread server, the
> "reliable" messages aren't nearly so reliable.

I guess there are a few ways to look a this.  This is a perfect example 
of why flow control shouldn't be built into Spread.  The flow control 
for this sort of application (if really needed) could be tailored 
specifically to logging and be done very effeciently on the application 
level.

However, even on large clusters, if your logging machine can't keep up, 
you have much bigger problems that flow control over Spread.  Either you 
are running a really big site (2000+ req/sec) in which case you should 
talk to a professional about mod_log_spread (like the author).  Or your 
logger is far underpowered.  Or our Spread ring is malfunctioning.

I have mod_log_spread logging fine in a decent size production cluster 
(4 million hits/day) and see almost no message loss on the Spread level, 
so even if I did change the RELIABLE_MESS to UNRELIABLE_MESS, I would 
have only lost a handful (<100) messages over the past week.

If you are seeing message loss on the link layer or are overflowing your 
recv buffers, then you need to investigate the way you have Spread set up.

--
Theo