[Gluster-users] Question on file reads from a distributed replicated volume

[Jeff Darcy]

> Do, "gluster volume set help". There is a pretty good explanation of
> the read subvolume preferences and options.

Specifically, you'll want to look at the cluster.read-hash-mode option,
which has one of three values:

(0) Each client will determine which brick seems fastest, and use that
    for all files unless a brick *failure* causes it to re-evaluate.  If
    the once-fastest brick becomes slower this will *not* be noticed by
    clients unless there's a failure.  Unfortunately, this mode is
    likely to *create* such a condition by overloading one server.

(1) The "read child" for each file will be found using a hash of its
    GFID, to ensure even distribution.  Note that if some servers are
    faster than others, the distribution will be *even* but not
    *optimal*.  This mode is the default.

(2) Similar to (1) except that each client used the hash of the file's
    GFID *plus its own PID*, so that different clients will be spread
    across different bricks and avoid file-level hot spots.

All of these modes might be overridden if one of the bricks is local to
the client.  In that case, the client will always read the local copy
and this option is effectively ignored.