noria
noria copied to clipboard
mit-pdos
Static reader node replication
Place replicas in separate domains, including when there is only 1 replica. Place these domains round robin on separate workers. (migrate/mod.rs)
Support for removing all reader replicas along with a view. (controller/inner.rs)
Obtain views to replicas round robin. CLI options to set replicas and reader threads.
Benchmarks: 16 client threads and cores, 4 Noria servers with 1 core per server, 1 write every 100 on the vote benchmark. As number of replicas increases, Noria can handle a greater number of ops/s before falling behind. Tests: Same as before, with updated node counts to account for additional replicas. One new test to check that view actually returns different replicas, and that writes are reflected in these replicas.
Going to simplify commit() and clarify the language surrounding replicas to address the remaining comments. It seems like sharding followed the convention where None meant 1 shard. For replication, i think it makes sense to think of the replication parameter as the "replication factor" rather than the number of replicas in addition to the original reader --> so replication factor of 1 = 1 reader.
Given 232453f9bbd94383af312d432f5a41766f10f725, I'd particularly like @fintelia to take a look at this.
Where did we end up with this? I vaguely remember discussing the outstanding issues with @fintelia before the holidays.
Last I recall query removal was being corrected because it didn't properly handle deletion of non-reader nodes. I'm not sure what the current status is
@fintelia I don't have any more changes to make. Is there a case you're concerned about in particular? It looks like the old code removes non-reader nodes in reverse topological order before it reaches my code for removing leaf/reader nodes, so it shouldn't be affected.