It's a bit unfortunate that NTS does not maintain the ability to lose a rack without loss of quorum for RF > #racks > 2, since this can be easily achieved by evenly placing replicas across all racks.
Since RackAwareTopologyStrategy is a superset of NetworkTopologyStrategy, can't we just use the new correct placement logic for newly created keyspaces instead of having a new strategy? The placement logic would be backwards-compatible for RF <= #racks. On upgrade, we could mark existing keyspaces with RF > #racks with use_legacy_replica_placement=true to maintain backwards compatibility and log a warning that the rack loss guarantee is not maintained for keyspaces created before the fix. Old keyspaces with RF <=#racks would still work with the new replica placement. The downside is that we would need to keep the old NTS logic around, or we could eventually deprecate it and require users to migrate keyspaces using the legacy placement strategy. Alternatively we could have RackAwareTopologyStrategy and fail NTS keyspace creation for RF > #racks and indicate users to use RackAwareTopologyStrategy to maintain the quorum guarantee on rack loss or set an override flag "support_quorum_on_rack_loss=false". This feels a bit iffy though since it could potentially confuse users about when to use each strategy. On Mon, Mar 6, 2023 at 5:51 AM Miklosovic, Stefan < stefan.mikloso...@netapp.com> wrote: > Hi all, > > some time ago we identified an issue with NetworkTopologyStrategy. The > problem is that when RF > number of racks, it may happen that NTS places > replicas in such a way that when whole rack is lost, we lose QUORUM and > data are not available anymore if QUORUM CL is used. > > To illustrate this problem, lets have this setup: > > 9 nodes in 1 DC, 3 racks, 3 nodes per rack. RF = 5. Then, NTS could place > replicas like this: 3 replicas in rack1, 1 replica in rack2, 1 replica in > rack3. Hence, when rack1 is lost, we do not have QUORUM. > > It seems to us that there is already some logic around this scenario (1) > but the implementation is not entirely correct. This solution is not > computing the replica placement correctly so the above problem would be > addressed. > > We created a draft here (2, 3) which fixes it. > > There is also a test which simulates this scenario. When I assign 256 > tokens to each node randomly (by same mean as generatetokens command uses) > and I try to compute natural replicas for 1 billion random tokens and I > compute how many cases there will be when 3 replicas out of 5 are inserted > in the same rack (so by losing it we would lose quorum), for above setup I > get around 6%. > > For 12 nodes, 3 racks, 4 nodes per rack, rf = 5, this happens in 10% cases. > > To interpret this number, it basically means that with such topology, RF > and CL, when a random rack fails completely, when doing a random read, > there is 6% chance that data will not be available (or 10%, respectively). > > One caveat here is that NTS is not compatible with this new strategy > anymore because it will place replicas differently. So I guess that fixing > this in NTS will not be possible because of upgrades. I think people would > need to setup completely new keyspace and somehow migrate data if they wish > or they just start from scratch with this strategy. > > Questions: > > 1) do you think this is meaningful to fix and it might end up in trunk? > > 2) should not we just ban this scenario entirely? It might be possible to > check the configuration upon keyspace creation (rf > num of racks) and if > we see this is problematic we would just fail that query? Guardrail maybe? > > 3) people in the ticket mention writing "CEP" for this but I do not see > any reason to do so. It is just a strategy as any other. What would that > CEP would even be about? Is this necessary? > > Regards > > (1) > https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/locator/NetworkTopologyStrategy.java#L126-L128 > (2) https://github.com/apache/cassandra/pull/2191 > (3) https://issues.apache.org/jira/browse/CASSANDRA-16203
