FWIW, the original Dynamo concept supported safety by returning not just the latest version, but all conflicting versions: the client then had the opportunity to make a merged version of the data that was newer than either of the inputs, and store that as a replacement.
It should therefore be remembered that many/most implementations of "eventual consistency" have these issues, it is not a requirement of the mechanism, and some implementations realize this and either have merge implementations or have plans to provide them.
(I am not certain where Cassandra is on this axis, but last I paid attention they were actively trying to decide whether to modify the client protocol to match Dynamo, or provide server-assisted merge operators more similar to their existing server-assisted comparison operators.)
> FWIW, the original Dynamo concept supported safety by returning not just the latest version, but all conflicting versions: the client then had the opportunity to make a merged version of the data that was newer than either of the inputs, and store that as a replacement.
Riak can do this as well; tracking siblings and inheritance with vector clocks, allowing a writer to say that a particular version represents the merge of two previous versions.
You're right--I definitely simplified the discussion of handling concurrent writes. Keeping around concurrent versions and using user-specified merge functions are two ways of deciding which versions to store. Riak indeed supports keeping concurrent versions or using an automatic last-writer-wins policy [1]
The question of which versions will be returned depends on the safety properties of the consistency model. My main point isn't that Dynamo or Riak don't provide safety properties, it's that "eventual consistency" isn't a safety property on its own.
[1] Eventual convergence is likely the strongest convergence property we can guarantee given unbounded partition durations.
I don't think this is true. Consider the property I call "eventually known consistency", wherein the system can be asked "are all operations performed before time T visible everywhere?", with a "yes"/"maybe" response, where "yes" is guaranteed to eventually be returned after some bounded period of non-partition.
Eventually known consistency can be used to get AP (just ask for the data), CP (let T be the current time; spin until ConsistentUpTo(T) returns true; then perform the read), or CA (in the sense that if as long as a partition does not occur, the algorithm for CP provides a response within a bounded time), and is thus strictly stronger than other properties.
I'm not sure I entirely follow your argument. I agree that you can use some sort of heartbeat protocol or negative acknowledgments to determine whether writes occurred within a given time window (e.g., I haven't heard from the cluster, so maybe I missed an update in the time since the last heartbeat). However, in general, I don't believe it's possible to guarantee non-trivial convergence for fixed T and unlimited partition durations.
For any given T, if I partition each of your nodes for T+1 seconds, you won't be able to guarantee convergence--your nodes won't communicate. Am I missing something?
If you can bound partition durations, you can definitely make stronger guarantees.
If you can model your network delays, you can use some modeling like our work on PBS (Probabilistically Bounded Staleness) to predict staleness:
http://pbs.cs.berkeley.edu/#demo
As Dylan16807 said, the key property in "eventually known consistency" is that when partitions don't happen you'll eventually get a "yes, that write has propagated everywhere" answer. You can do this with gossiping updates and two clock vectors (vector #1 is "I have seen all the writes node X did up to this time", and vector #2 is "node X says it has seen everything done by any node prior to this time").
While you have partitions, you'll get a "maybe" back, because you can't distinguish between the cases of "a write hasn't propagated everywhere" and "it has propagated everywhere but I haven't received an ACK for it"; that's unavoidable, but doesn't mean that this is not useful anyway.
tl;dr. "Eventual consistency" is a b.s. marketing term that actually has nearly no meaning.
The easiest way to see just how empty the definition is, is by negating it and seing what undesirable property a system would have to have in order to fail to be eventually consistent. Take the definition from wikipedia:
"Given a sufficiently long period of time over which no changes are sent, all updates can be expected to propagate eventually through the system and all the replicas will be consistent."
Negation:
"No matter how long you wait without sending changes, updates may not propagate through the system, and disagreement may continue to exist between replicas indefinitely."
So basically, saying that something is "eventually consistent" just means "we won't completely ignore you forever". Great.
Eventual consistency is definitely weak, but it's useful: if we couple eventual convergence with safety properties, then we can describe non-trivial properties.
The problem, which I alluded to in my first footnote, is that it's hard to guarantee anything stronger than eventual in the presence of partitions. If you want to make sure your system behaves "correctly" under arbitrary partitioning, you necessarily have to admit the trivial cases. What you also want to guarantee is that, in the absence of partitions, the system still "does the right thing"; this often gets lost in the definition of eventually consistent systems, which is why you should consider both safety and liveness.
If you completely understand the weaknesses of eventual consistency you will use is only where it is viable and the risk is justified. There is a reason the relational databases haven't gone away completely.
As a shameless plug, I might add that if you can model your network delays, you can use some modeling like our work on PBS (Probabilistically Bounded Staleness) to predict staleness: http://pbs.cs.berkeley.edu/#demo
I'd also add that the consistency related to ACID semantics from relational databases refers to transactional consistency, not replica consistency. Indeed, distributed RDBMSs often opt for strong (replica) consistency models, but there is no reason a distributed relational database can't be weakly (replica) consistent while maintaining ACID semantics on a single machine. Moreover, if the RDBMS must needs to be available in the presence of partitions, it must be weakly (replica) consistent.
The problem isn't really the term "eventual coincistency" but rather neglecting to define "sufficient time", which needs to be done case-by-case. Something like "ram size divided by the slowest connection between servers, plus 40% margin". That gives you a useful and predictable model that even handles partitions (treat as speed 0).
13 comments
[ 5.7 ms ] story [ 42.5 ms ] threadIt should therefore be remembered that many/most implementations of "eventual consistency" have these issues, it is not a requirement of the mechanism, and some implementations realize this and either have merge implementations or have plans to provide them.
(I am not certain where Cassandra is on this axis, but last I paid attention they were actively trying to decide whether to modify the client protocol to match Dynamo, or provide server-assisted merge operators more similar to their existing server-assisted comparison operators.)
Riak can do this as well; tracking siblings and inheritance with vector clocks, allowing a writer to say that a particular version represents the merge of two previous versions.
The question of which versions will be returned depends on the safety properties of the consistency model. My main point isn't that Dynamo or Riak don't provide safety properties, it's that "eventual consistency" isn't a safety property on its own.
[1] http://wiki.basho.com/Vector-Clocks.html
I don't think this is true. Consider the property I call "eventually known consistency", wherein the system can be asked "are all operations performed before time T visible everywhere?", with a "yes"/"maybe" response, where "yes" is guaranteed to eventually be returned after some bounded period of non-partition.
Eventually known consistency can be used to get AP (just ask for the data), CP (let T be the current time; spin until ConsistentUpTo(T) returns true; then perform the read), or CA (in the sense that if as long as a partition does not occur, the algorithm for CP provides a response within a bounded time), and is thus strictly stronger than other properties.
For any given T, if I partition each of your nodes for T+1 seconds, you won't be able to guarantee convergence--your nodes won't communicate. Am I missing something?
> "yes" is guaranteed to eventually be returned after some bounded period of non-partition
If you can model your network delays, you can use some modeling like our work on PBS (Probabilistically Bounded Staleness) to predict staleness: http://pbs.cs.berkeley.edu/#demo
While you have partitions, you'll get a "maybe" back, because you can't distinguish between the cases of "a write hasn't propagated everywhere" and "it has propagated everywhere but I haven't received an ACK for it"; that's unavoidable, but doesn't mean that this is not useful anyway.
The easiest way to see just how empty the definition is, is by negating it and seing what undesirable property a system would have to have in order to fail to be eventually consistent. Take the definition from wikipedia:
"Given a sufficiently long period of time over which no changes are sent, all updates can be expected to propagate eventually through the system and all the replicas will be consistent."
Negation:
"No matter how long you wait without sending changes, updates may not propagate through the system, and disagreement may continue to exist between replicas indefinitely."
So basically, saying that something is "eventually consistent" just means "we won't completely ignore you forever". Great.
The problem, which I alluded to in my first footnote, is that it's hard to guarantee anything stronger than eventual in the presence of partitions. If you want to make sure your system behaves "correctly" under arbitrary partitioning, you necessarily have to admit the trivial cases. What you also want to guarantee is that, in the absence of partitions, the system still "does the right thing"; this often gets lost in the definition of eventually consistent systems, which is why you should consider both safety and liveness.
I'd also add that the consistency related to ACID semantics from relational databases refers to transactional consistency, not replica consistency. Indeed, distributed RDBMSs often opt for strong (replica) consistency models, but there is no reason a distributed relational database can't be weakly (replica) consistent while maintaining ACID semantics on a single machine. Moreover, if the RDBMS must needs to be available in the presence of partitions, it must be weakly (replica) consistent.