One of the latest challenges in computer science seems to be the CAP theorem. It addresses a perceived impossibility of building large-scale and clustered (web) service architectures. The fact that it (supposedly) has been proven to be true makes what I am going to write here all the more unlikely. Still, read on because I will show that I am right and CAP is not an impossibility after all... While the impossibility proof of CAP is mathematically correct, it is based on assumptions that are too strict. By relaxing these assumptions, I found the solution presented here.

What is CAP?

The CAP theorem (short for consistency, availability, partition-tolerant) essentially states that you cannot have a clustered system that supports all of the following three qualities:

Consistency is a quality meaning (informally speaking) that reads and writes happen correctly. In other words, the overall effect of executing thousands or millions of transactions concurrently is the same as if they had been executed one-at-a-time. Usually, this is done with the help of a transaction manager of some sort.
Availability essentially means that every operation (that makes it to a non-failing node) eventually returns a result.
This quality refers to the possibility of tolerating partitions on the network. Note that we suppose a cluster architecture (which is where the network comes in).

CAP is a conjecture originally formulated by Eric Brewer (Inktomi) and has influenced many of today's larger-scale websites like . In other words, the impact of CAP is very large. To make it worse, the perceived impossibility of a CAP system (one that has all three desirable properties) has lead people to advocate something called BASE (Basically Available, Soft-state and Eventually Consistent) - see this talk by Werner Vogels (CTO at Amazon).

As far as I know (but I could be wrong), a theoretical foundation of BASE does not exist yet (it seems more of an informal approach which to me raises serious questions concerning correctness). In this post I will present:
  • a CAP solution
  • how this conforms to what BASE wants to achieve
  • a "design pattern" for building correct systems that (in a way) offer both CAP and BASE qualities
Because CAP is perceived as impossible and because BASE lacks formal treatment, I consider this to be a signification contribution to the state of today's engineering;-)

What about the proof of Brewer's theorem?

Brewer's proof has been published by Nancy Lynch et al and discussed by me (see my earlier post and also this one).

While the theoretical proof of the impossibility of CAP is valid, it has a big limitation: it assumes that all three CAP properties have to be supplied at the same moment in time. If you drop this assumption, then all of a sudden you get into a new spectrum of possibilities. This is what I will do here.

A CAP solution

Enough talk, let's get to the core of the matter. Here is my solution to CAP. To make it concrete, I will use the concept of a web-shop like Amazon. Here are the rules that are sufficient to ensure CAP:
  1. Process reads from the database if possible, or use a cached value if needed for availability (if the DB is unreachable).
  2. All reads use versioning or another mechanism that allows optimistic locking.
  3. Updates supplied by clients (orders in case of Amazon) are queued for execution, and include the versioning information of the reads that lead to the update.
  4. Queued updates are processed when the number of partitions is low enough to do so. The easiest way to do this is with a cluster-wide distributed transaction across all replicas (more on scalability later), but other more refined ways are possible (such as quorum-based replication or any other smart way of replicating). The version information in the update is used to validate it: if the data in the database has been modified since the original read(s) that lead to the update, the update is rejected and a cancellation is reported back to the client. Otherwise the order is processed and a confirmation is reported back to the client.
  5. The results (confirmation or cancellation) are sent asynchronously to the clients. This can be either email, message queuing, or any other asynchronous delivery method.
That's it. Adhere to these guidelines, and you have a CAP architecture. I will not provide a formal proof here (I intend to do that elsewhere, in a research paper), but intuitively the proof is as follows:
  • This system is consistent because reads are based on snapshots and incorrect updates are rejected before they are applied. In other words: there are no incorrect executions.
  • This system is available since reads always return a value, and so do writes (even though they are queued and it may take a while).
  • This system is partition-tolerant because it allows network and node failures.
Granted, this system does not provide all three at the same moment in time (which is how we go around the impossibility), but nevertheless the result is quite strong IMHO.

The limitations

There are some limitations to this solution - all of which seem reasonable:
  1. Read-only requests may be presented with stale information (due to updates that have yet-to-be-applied). In that sense, their results could be "inconsistent": for instance, the availability of an Amazon item can change between two page views. I do not see this as a major restriction, since no website that I know of will offer read consistency for the duration of a user session. It all depends on what you consider to be within the scope of one transaction;-) Note that this almost corresponds to snapshot isolation found in Oracle.
  2. Partitions should not last forever: in order for this to work, partitions should be resolved within a reasonable time (reasonable being: within the expected confirmation time for updates). The duration of any partitions also affects the time window in which reads can produce stale data.
  3. The updates have to be applied in the same relative order at all cluster nodes. This puts some restrictions on the algorithm used to do this.
Note that updates are always based on correct reads thanks to the versioning check before they are applied. So update transactions are always consistent.

How does this relate to BASE?

You could see this as a design pattern for BASE if you like. The solution adheres to BASE in the sense that it uses cached reads (if needed) and that the updates are delayed (so you could say they are "eventually" applied and the system becomes "consistent").

Reflections in scalability

So far the CAP focus was on possibility. I think my solution shows that it is possible. Now how about scaling up?

The naive solution (a huge distributed transaction to update all cluster nodes in-sync) is unlikely to scale: as you add more nodes, more updates are needed. Now I am a big fan of transactions, but not to use them in an arbitrary matter. So how to propagate these updates through the cluster?

While smarter solutions for this exist (such as the work by Bettina Kemme), a trivial first try would be to push updates (lazily) to all nodes in the cluster. This can be done with a smart queuing mechanism. The disadvantage is that updates are not applied everywhere at once (rather, the all-or-nothing quality just "ripples" through the system). So you get into the "eventually" style again.

Note that this latter suggestion makes the system behave much like the READ COMMITTED isolation level (which, by the way, is the default in Oracle). So this approach sacrifices consistency/isolation a bit in favor of scalability.

Future work

Additional research could/should be done in the following areas:
  • Improving read consistency through session affinity
  • The best way to push the updates through the cluster
  • Performance evaluation in real life implementations
Final note and disclaimer

I did not see Brewer's original presentation of the CAP theorem - so it could be that what he meant with consistency also involved all reads (see the limitations of the solution I presented here). In that case I did not find a solution for CAP but at least it is a framework and proof outline for BASE wink

UPDATE 15/3/2012:

It seems like Greg Young and Udi Dahan have been working along similar lines and gave this pattern/solution a name: CQRS.

Contact Us

Atomikos Corporate Headquarters
Hoveniersstraat, 39/1, 2800
Mechelen, Belgium

T +3215613055

Subscribe to our newsletter

Never miss an update

Copyright 2015 Atomikos BVBA