Re:Strict-serializability, but at what cost, for what purpose?
by _benedict on Thursday August 04, 2022 @07:18PM
(#32349520)
> you lose strict serializability across the shards
Nope. That's precisely what Accord manages to maintain. It is not the first such protocol to do so, but so far none of the others have left the lab.
Re:Strict-serializability, but at what cost, for what purpose?
by kbr- on Thursday August 04, 2022 @07:33PM
(#32349647)
Ok. From skimming through the paper, it seems that for a given transaction, only the replica sets of the shards touched by this transaction need to participate. Thus for two transactions touching disjoint sets of shards, they can run completely in parallel. Hence the scalability. Is that right?
I understand scalability by being able to handle more traffic through adding more nodes to your cluster. But if you want to handle more traffic, adding more replicas to a single Accord replica set won't help. Any two (fast path) quorums must intersect. And that intersection point for two queries is where you need to spend resources to handle both queries.
So - unless I'm missing something here - in order to scale out to be able to handle more traffic, you must shard the data, there's no way around it. Even assuming 0 failures won't help.
And once you shard your data and use multiple Paxos/Raft/EPaxos/Accord/whatever groups, you lose strict serializability across the shards.