Gluster users - Oct 2016 - How gluster parallelize reads

2016-10-03 20:13 GMT+02:00 Alastair Neil <ajneil.tech at
gmail.com>:
> I think this might give you something like  the behaviour you are looking
> for, it will not balance blocks across different servers but will
distribute
> reads from clients across all the servers.
>
> cluster.read-hash-mode 2
>
> 0 means use the first server to respond I think - at least that's my
guess
> of what "first up server" means
> 1 hashed by GFID,  so clients will use the same server for a given file but
> different files may be accessed from different nodes.
I think that 1 is better.
Why "0" is the default ?

> > 0 means use the first server to respond I think - at least that's
my guess
> > of what "first up server" means
> > 1 hashed by GFID,  so clients will use the same server for a given
file but
> > different files may be accessed from different nodes.
> 
> I think that 1 is better.
> Why "0" is the default ?
Basic storage-developer conservatism.  Zero was the behavior before
read-hash-mode was implemented.  As strongly as some of us might believe
that such tweaks lead to better behavior - as I did with this one in
2012[1] - we've kind of learned the hard way that existing users often
disagree with our estimations.  Thus, new behavior is often kept as a
"special" for particular known environments or use cases, and the
default is left unchanged until there's clear feedback indicating it
should be otherwise.

[1] http://review.gluster.org/#/c/2926/