On 2025-04-30 17:07, Doug Robinson wrote:
On Wed, Apr 30, 2025 at 10:25 AM Doug Robinson
<doug.robin...@cirata.com> wrote:
On Wed, Apr 30, 2025 at 4:17 AM Andreas Stieger
<andreas.stie...@gmx.de> wrote:
WANdisco is not a typical Subversion server implementation. For
synchronous multi-site replication (as opposed to svn sync
which is
asynchronous), it provides a proxy layer with a consensus
protocol
(think PAXOS, wsrep, raft).
The WANdisco implementation is still asynchronous even though we try
our hardest to make the replication occur absolutely as fast as possible.
The PAXOS implementation is used to schedule the order of the repository
updates. The updates then occur in that order at every site as soon
as the
data is available and the prior update has been applied.
While a synchronous implementation might be great within a single data
center, it comes with way too many problems for a world-wide WAN
environment.
Certainly, I should have been more clear with that I mean: In a
multi-master replication setup you need consensus at a point when you
want a candidate transaction to become a revision. If there is only such
candidate ready, you can commit to both your global sequence and to the
committing the client rather quickly. Not every other node may see it
instantaneously, but would not be able to commit to another until they
are caught up. I briefly played with the wsrep replication API between
svn_fs layers to make a FLOSS version of that but never got beyond a
prototype.
Andreas
