On Thu, Jan 5, 2017 at 3:07 AM, Greenhorn Techie <greenhorntec...@gmail.com> wrote:
> Hi, > > We are planning to setup MirrorMaker based Kafka replication for DR > purposes. The base requirement is to have a DR replication from primary > (site1) to DR site (site2)using MirrorMaker, > > However, we need the solution to work in case of failover as well i.e. > where in the event of the site1 kafka cluster failing, site2 kafka cluster > would be made primary. Later when site1 cluster eventually comes back-up > online, direction of replication would be from site2 to site1. > > But as I understand, the offsets on each of the clusters are different, so > wondering how to design the solution given this constraint and > requirements. > It turns out this is tricky. And once you start digging in you'll find it's way more complicated than you might originally think. Before going down the rabbit hole, I'd suggest taking a look at this great talk by Jun Rao (one of the original authors of Kafka) about multi-DC Kafka setups: https://www.youtube.com/watch?v=Dvk0cwqGgws Additionally, I want to mention that while it is tempting to want to treat multi-DC DR cases in a way that we get really convenient, strongly consistent, highly available behavior because that makes it easier to reason about and avoids pushing much of the burden down to applications, that's not realistic or practical. And honestly, it's rarely even necessary. DR cases really are DR. Usually it is possible to make some tradeoffs you might not make under normal circumstances (the most important one being the tradeoff between possibly seeing duplicates vs exactly once). The tension here is often that one team is responsible for maintain the infrastructure and handling this DR failover scenario, and others are responsible for the behavior of the applications. The infrastructure team is responsible for figuring out the DR failover story but if they don't solve it at the infrastructure layer then they get stuck having to understand all the current (and future) applications built on that infrastructure. That said, here are the details I think you're looking for: The short answer right now is that doing DR failover like that is not going to be easy with MM. Confluent is building additional tools to deal with multi-DC setups because of a bunch of these challenges: https://www.confluent.io/product/multi-datacenter/ For your specific concern about reversing the direction of replication, you'd need to build additional tooling to support this. The basic list of steps would be something like this (assuming non-compacted topics): 1. Use MM normally to replicate your data. Be *very* sure you construct your setup to ensure *everything* is mirrored (proper # of partitions, replication factor, topic level configs, etc). (Note that this is something the Confluent replication solution is addressing that's a significant gap in MM.) 2. During replication, be sure to record offset deltas for every topic partition. These are needed to reverse the direction of replication correctly. Make sure to store them in the backup DC and somewhere very reliable. 3. Observe DC failure. 4. Decide to do failover. Ensure replication has actually stopped (via your own tooling, or probably better, by using ACLs to ensure no new data can be produced from original DC to backup DC) 5. Record all the high watermarks for every topic partition so you know which data was replicated from the original DC (vs which is new after failover). 6. Allow failover to proceed. Make the backup DC primary. 7. Once the original DC is back alive, you want to reverse replication and make it the backup. Lookup the offset deltas, use them to initialize offsets for the consumer group you'll use to do replication. 8. Go back to the original DC and make sure there isn't any "extra" data, i.e. stuff that didn't get replicated but was successfully written to the original DC's cluster. For topic partitions where there is data beyond the expected offsets, you currently would need to just delete the entire set of data, or at least to before the offset we expect to start at. (A truncate operation might be a nice way to avoid having to dump *all* the data, but doesn't currently exist.) 9. Once you've got the two clusters back in a reasonably synced state with appropriate starting offsets committed, start up MM again in the reverse direction. If this sounds tricky, it turns out that when you add compacted topics, things get quite a bit messier.... -Ewen > > Thanks >
