Sure, the update queue could just as well replicate problems, but the queue would be a lot simpler than cassandra and it would not modify already acknowledged data like like for instance compaction or read-repair/hint deliveries may. There is a fair bit of re-writing/re-assemblying of data even though it is actually never updated since the original data was acknowledged. From a statistically viewpoint, there is clearly a noticeable risk in data getting messed up internally in Cassandra during these operations.
Of course, in a perfectly replicated case, the same error is likely to occur even in two isolated systems if they have the exact same data in the same order, but lets add more fun things that replicate well such as operator mistakes. Yes, isolated systems increases maintenance work, and increase in work increases risk of mistakes, but it reduces risk that you do a mistake that brings down everything. Bottom line is, I am far from convinced about the benefits of one big magical system across all datacenters vs. more isolated setups. It is not a Cassandra specific concern though. It applies to any system. Backups needs to be there anyway of course, but if you have a system that is somewhat working ok, but compactions have stopped due to a bad sstable, how do you recover that from backup without taking down the service or going through the interesting task of splitting up the cassandra setup, recover on a limited set of nodes and then re-joining again? (maybe easier to do with datacenter replication, haven't actually thought about how to do it there). Yes, I realize that I could do 2 queries (or more) to get the wanted behavior, but doubling the number of queries when the system is already in trouble is rarely a good idea. Terje On Thu, Apr 21, 2011 at 10:49 AM, Adrian Cockcroft < adrian.cockcr...@gmail.com> wrote: > Queues replicate bad data just as well as anything else. The biggest > source of bad data is broken app code... You will still need to > implement a reconciliation/repair checker, as queues have their own > failure modes when they get backed up. We have also looked at using > queues to bounce data between cassandra clusters for other reasons, > and they have their place. However it is a lot more work to implement > than using existing well tested Cassandra functionality to do it for > us. > > I think your code needs to retry a failed local-quorum read with a > read-one to get the behavior you are asking for. > > Our approach to bad data and corruption issues is backups, wind back > to the last good snapshot. We have figured out incremental backups as > well as full. Our code has some local dependencies, but could be the > basis for a generic solution. > > Adrian > > On Wed, Apr 20, 2011 at 6:08 PM, Terje Marthinussen > <tmarthinus...@gmail.com> wrote: > > Assuming that you generally put an API on top of this, delivering to two > or > > more systems then boils down to a message queue issue or some similar > > mechanism which handles secure delivery of messages. Maybe not trivial, > but > > there are many products that can help you with this, and it is a lot > easier > > to implement than a fully distributed storage system. > > Yes, ideally Cassandra will not distribute corruption, but the reason you > > pay up to have 2 fully redundant setups in 2 different datacenters is > > because we do not live in an ideal world. Anyone having tested Cassandra > > since 0.7.0 with any real data will be able to testify how well it can > mess > > things up. > > This is not specific to Cassandra, in fact, I would argue thats this is > in > > the blood of any distributed system. You want them to distribute after > all > > and the tighter the coupling is between nodes, the better they distribute > > bad stuff as well as good stuff. > > There is a bigger risk for a complete failure with 2 tightly coupled > > redundant systems than with 2 almost completely isolated ones. The logic > > here is so simple it is really somewhat beyond discussion. > > There are a few other advantages of isolating the systems. Especially in > > terms of operation, 2 isolated systems would be much easier as you could > > relatively risk fee try out a new cassandra in one datacenter or upgrade > one > > datacenter at a time if you needed major operational changes such as > schema > > changes or other large changes to the data. > > I see the 2 copies in one datacenters + 1(or maybe 2) in another as a > "low > > cost" middleway between 2 full N+2 (RF=3) systems in both data centers. > > That is, in a traditional design where you need 1 node for normal > service, > > you would have 1 extra replicate for redundancy and one replica more (N+2 > > redundancy) so you can do maintenance and still be redundant. > > If I have redundancy across datacenters, I would probably still want 2 > > replicas to avoid network traffic between DCs in case of a node recovery, > > but N+2 may not be needed as my risk policy may find it acceptable to run > > one datacenters without redundancy for a time limited period for > > maintenance. > > That is, if my original requirement is 1 node, I could do with 3x the HW > > which is not all that much more than the 3x I need for one DC and a lot > less > > than the 6x I need for 2 full N+2 systems. > > However, all of the above is really beyond the point of my original > > suggestion. > > Regardless of datacenters, redundancy and distribution of bad or good > stuff, > > it would be good to have a way to return whatever data is there, but with > a > > flag or similar stating that the consistency level was not met. > > Again, for a lot of services, it is fully acceptable, and a lot better, > to > > return an almost complete (or maybe even complete, but no verified by > > quorum) result than no result at all. > > As far as I remember from the code, this just boils down to returning > > whatever you collected from the cluster and setting the proper flag or > > similar on the resultset rather than returning an error. > > Terje > > On Thu, Apr 21, 2011 at 5:01 AM, Adrian Cockcroft > > <adrian.cockcr...@gmail.com> wrote: > >> > >> Hi Terje, > >> > >> If you feed data to two rings, you will get inconsistency drift as an > >> update to one succeeds and to the other fails from time to time. You > >> would have to build your own read repair. This all starts to look like > >> "I don't trust Cassandra code to work, so I will write my own buggy > >> one off versions of Cassandra functionality". I lean towards using > >> Cassandra features rather than rolling my own because there is a large > >> community testing, fixing and extending Cassandra, and making sure > >> that the algorithms are robust. Distributed systems are very hard to > >> get right, I trust lots of users and eyeballs on the code more than > >> even the best engineer working alone. > >> > >> Cassandra doesn't "replicate sstable corruptions". It detects corrupt > >> data and only replicates good data. Also data isn't replicated to > >> three identical nodes in the way you imply, it's replicated around the > >> ring. If you lose three nodes, you don't lose a whole node's worth of > >> data. We configure each replica to be in a different availability > >> zone so that we can lose a third of our nodes (a whole zone) and still > >> work. On a 300 node system with RF=3 and no zones, losing one or two > >> nodes you still have all your data, and can repair the loss quickly. > >> With three nodes dead at once you don't lose 1% of the data (3/300) I > >> think you lose 1/(300*300*300) of the data (someone check my math?). > >> > >> If you want to always get a result, then you use "read one", if you > >> want to get a highly available better quality result use local quorum. > >> That is a per-query option. > >> > >> Adrian > >> > >> On Tue, Apr 19, 2011 at 6:46 PM, Terje Marthinussen > >> <tmarthinus...@gmail.com> wrote: > >> > If you have RF=3 in both datacenters, it could be discussed if there > is > >> > a > >> > point to use the built in replication in Cassandra at all vs. feeding > >> > the > >> > data to both datacenters and get 2 100% isolated cassandra instances > >> > that > >> > cannot replicate sstable corruptions between each others.... > >> > My point is really a bit more general though. > >> > For a lot services (especially Internet based ones) 100% accuracy in > >> > terms > >> > of results is not needed (or maybe even expected) > >> > While you want to serve a 100% correct result if you can (using > quorum), > >> > it > >> > is still much better to serve a partial result than no result at all. > >> > Lets say you have 300 nodes in your ring, one document manages to > >> > trigger a > >> > bug in cassandra that brings down a node with all its replicas (3 > nodes > >> > down) > >> > For many use cases, it would be much better to return the remaining > 99% > >> > of > >> > the data coming from the 297 working nodes than having a service which > >> > returns nothing at all. > >> > I would however like the frontend to realize that this is an > incomplete > >> > result so it is possible for it to react accordingly as well as be > part > >> > of > >> > monitoring of the cassandra ring. > >> > Regards, > >> > Terje > >> > > >> > On Tue, Apr 19, 2011 at 6:06 PM, Adrian Cockcroft > >> > <adrian.cockcr...@gmail.com> wrote: > >> >> > >> >> If you want to use local quorum for a distributed setup, it doesn't > >> >> make sense to have less than RF=3 local and remote. Three copies at > >> >> both ends will give you high availability. Only one copy of the data > >> >> is sent over the wide area link (with recent versions). > >> >> > >> >> There is no need to use mirrored or RAID5 disk in each node in this > >> >> case, since you are using RAIN (N for nodes) to protect your data. So > >> >> the extra disk space to hold three copies at each end shouldn't be a > >> >> big deal. Netflix is using striped internal disks on EC2 nodes for > >> >> this. > >> >> > >> >> Adrian > >> >> > >> >> On Mon, Apr 18, 2011 at 11:16 PM, Terje Marthinussen > >> >> <tmarthinus...@gmail.com> wrote: > >> >> > Hum... > >> >> > Seems like it could be an idea in a case like this with a mode > where > >> >> > result > >> >> > is always returned (if possible), but where a flay saying if the > >> >> > consistency > >> >> > level was met, or to what level it was met (number of nodes > answering > >> >> > for > >> >> > instance).? > >> >> > Terje > >> >> > > >> >> > On Tue, Apr 19, 2011 at 1:13 AM, Jonathan Ellis <jbel...@gmail.com > > > >> >> > wrote: > >> >> >> > >> >> >> They will timeout until failure detector realizes the DC1 nodes > are > >> >> >> down (~10 seconds). After that they will immediately return > >> >> >> UnavailableException until DC1 comes back up. > >> >> >> > >> >> >> On Mon, Apr 18, 2011 at 10:43 AM, Baskar Duraikannu > >> >> >> <baskar.duraikannu...@gmail.com> wrote: > >> >> >> > We are planning to deploy Cassandra on two data centers. Let > us > >> >> >> > say > >> >> >> > that > >> >> >> > we went with three replicas with 2 being in one data center and > >> >> >> > last > >> >> >> > replica > >> >> >> > in 2nd Data center. > >> >> >> > > >> >> >> > What will happen to Quorum Reads and Writes when DC1 goes down > (2 > >> >> >> > of > >> >> >> > 3 > >> >> >> > replicas are unreachable)? Will they timeout? > >> >> >> > > >> >> >> > > >> >> >> > Regards, > >> >> >> > Baskar > >> >> >> > >> >> >> > >> >> >> > >> >> >> -- > >> >> >> Jonathan Ellis > >> >> >> Project Chair, Apache Cassandra > >> >> >> co-founder of DataStax, the source for professional Cassandra > >> >> >> support > >> >> >> http://www.datastax.com > >> >> > > >> >> > > >> > > >> > > > > > >
