Of course two systems can only be consistent if they can communicate, so you have to either sacrifice availability until the partition is resolved, or give up on consistency.
I'm sorry, but I can't resist: Isn't the cap theorem irrelevant, because true network partitions never happen in the real world? If a link fails, an administrator will fix it eventually. With any system implementing ACK packets (tcp is one example) a link that fails but is then fixed is the same as a very slow link.
I don't think you should be sorry. I was wondering the same thing. If messages can't get lost (because of good network monitoring), can't you have perfect consistency and availability?
That said, practically speaking your system may have to wait for the partition to be fixed, which would either make the system practically unavailable or practically inconsistent. But not theoretically, at least, not as this site describes it.
We can turn the link catch into something practical and real by having redundant linking hardware and a really good sysadmin team. When your network going down is just as unlikely as all your servers going down at the same time you can be said to have won.
> With any system implementing ACK packets (tcp is one example) a link that fails but is then fixed is the same as a very slow link.
This is only even theoretically true if a) you have no bound on memory and b) the link will always come back up. In practice, neither of these is true.
Further, if you're relying on never being partitioned then any network break requires delaying availability until the partition is resolved. CAP is then relevant if you're not willing to do this (and virtually no one is going to be so inclined).
"A partition is when the network fails to deliver some messages to one or more nodes by losing them (not by delaying them - eventual delivery is not a partition)."
That part is confusing to me. Doesn't the term partition have another meaning in distributed system design? For instance, consistent hashing "partitions" keys to multiple nodes. I haven't heard partition as a term describing dataloss.
consider the network of nodes separating (partitioning) into two groups - each internally. connected, but unconnected from each other. then messages from one group never reach the other group.
so partitioning still has the sense of "splitting" - it's just that the explanation focuses on messages rather than the network.
interesting FAQ...i like the idea of bringing this info together.
i've found there are lots of more-common things that cause partitions in practice than equipment-in-the-middle failures. human errors are probably the biggest: network configuration changes, fresh bugs in your own software - or in your dependencies, etc.
also, while a network might be asynchronous, there's usually a limit to how long a message can be delayed in practice. ...the limit might be how much memory you have to queue up messages...or perhaps how long your client-side software (or your end-user) is willing to wait for a message when a dialog is more complex than request/response.
when designing distributed software, i've found that it's helpful to ask: when (not if) X process/server/cluster/data-center fails or becomes unreachable - temporarily or forever - how should the rest of my system respond?
so, perhaps the most important take-away from the FAQ for designers is #13: that C and A are "spectrums" that you tune to meet your own requirements when the various failure scenarios happen.
Sort of. CAP talks about behaviour during potentially ambiguous failures (network partitions), but most of the systems that call themselves "eventually consistent" also sacrifice consistency under normal operation. Examples are Cassandra, Riak and the original Dynamo.
The main tradeoff is that after writing the values 1, 2, 3 in order, reads could see anything from no value or any one of those three values until the nodes converge.
In a Consistent system, if a read happens after writing and you see a 3, you will never see a 2, 1 or no value on subsequent reads. In the case of a network partition, the system will prefer to not be available than to return reads older than reads that have already been returned.
That's not true for Cassandra. Cassandra allows for stronger consistency, if you only wish to trade some performance for it. If R + W > RF, you are guaranteed to get at least the latest successfully written value. There may be only a temporary inconsistency: during the write some nodes may see the old value, but some other may already see the new value.
Sure, if you're willing to trade off latency. As I understand it, if all reads and writes are QUORUM, Cassandra is Consistent (and not Available).
However, I was under the impression that this is not the "default" and I am likely to get performance poorer than distributed systems designed to be Consistent (like HBase).
Not so: first, writes are always sent to all replicas, so throughput for ONE and QUORUM is identical. The primary difference is that ONE is allowed to complete when more nodes are down.
It's true that QUORUM reads will have half (in a 3-replica system) the throughput as ONE, but with Cassandra reads up to 8x as fast as HBase [1], it still wins handily.
Finally, the reason this is worth making configurable in the first place is that almost all applications do just fine with most operations at ONE.
> There is another way. You can't avoid the CAP theorem, but you can isolate its complexity and prevent it from sabotaging your ability to reason about your systems.
I don't think he really debunked it there, he just came up with a way around it. If you accept what the theorem assumes about your system, then the cap theorem applies to you. Any system that squirrels away from it's assumptions can be said to "beat" it, but only in a practical sense.
"16. Have I 'got around' or 'beaten' the CAP theorem?
No. You might have designed a system that is not heavily affected by it. That's good."
Our thoughts on CAP and how we've dealt with it while building a distributed truly ACID database might also be interesting to some: http://foundationdb.com/white-papers/the-cap-theorem/