[Gluster-devel] semi-sync replication

[Jeff Darcy]

> Do we have plans to support "semi-synchronous" type replication in the
> future? By semi-sync I mean writing to one leg the replica, securing the
> write on a faster stable storage (capacitor backed SSD or NVRAM) and then
> acknowledge the client. The write on other replica leg may happen at later
> point in time.

This is possible, but introduces a lot of consistency/ordering concerns.
It has always been part of the plan for NSR, with leader election to
help with ordering and issue/completion counts to give the user some
control over consistency.  The same basic idea can be implemented in
AFR, but without the mechanisms mentioned above the consistency might
not be noticeably better than we already have with geo-replication.  I
also wonder whether deferring unlocks or pending-count updates further
might introduce performance glitches, or exacerbate the split-brain
problems we've been battling all these years.