Re:Jepsen: Distributed Systems Safety Analysis
by biokoda on Wednesday October 21, 2015 @03:39AM
(#10424130)
> Yeah, basically with PostgreSQL it boils down to the fact that even if a client receives an error on commit, its writes may in fact have succeeded -- and it cannot assume that they haven't.
This is true for almost every database. Otherwise they would require a distributed transaction between the client and server.
Re:Jepsen: Distributed Systems Safety Analysis
by lomnakkus on Friday October 23, 2015 @04:09PM
(#10441014)
Indeed, but given that we were talking about $DISTSYS vs. $CONCRETE_THING, I just mentioned the $CONCRETE_THING by name.
You are of course technically right... which is the best kind of right! :)
As an example: This means that idempotency is really important if you're doing any kind of retries on commit failure.
EDIT: Another example: If you're doing application-level optimistic concurrency control (using e.g. an incrementing version number) you can end up in a situation where an update appears to conflict with itself. Usually, though, this isn't a big concern since client<->server paritions are usually catastrophic anyway.
(Off the top of my head that's what I can recall -- there may be other things to be gleaned from the original article.)
[1] At least for all practical purposes.