A Foolish Consistency...

May 27, 2014

How do you scale a computer system?

It seems like such a simple question. Of course, to a certain extent, it is - if you've designed the system properly up front. By this I mean that, if you've broken the overall system down into components which are a) decoupled as much as possible, and b) as internally cohesive as possible.

However, when it comes to distributed computer systems - ones with, say, a UI, a bunch of business logic, and a database - it somehow becomes much more complicated. We know how to scale the UI and business layers, but scaling the database always becomes...messy.

I've recently been thinking a lot about why this is, spurred, in particular, by an interesting video of a discussion between Carl Hewitt, Erik Meijer and Clemens Szyperski.

Consistently Inconsistent

The discussion is mostly about Dr. Hewitt's actor model of computation - which is interesting in and of itself - but the thing that really caught my attention was an almost offhand comment he made about relational databases being a problem in that they are trying to enforce consistency in an inherently non-consistent situation.

To think about this, we first need to figure out what we mean by "consistent" in this context: we would regard a system as consistent when all parts of the system agree on the state of the system at any given time. Another way of putting this is that the system is transactionally atomic at the global scale.

Or, to put it in other words, it's a Turing machine.

Where Do I Put the Paper Tape?

If you'll recall, a Turing machine is a model of computation in which the state of a computer system is recorded on paper tape through a series of consecutive operations. Turing machines are the model computer scientists have used for years to model computer systems. Dr. Hewitt's point (and I think he's right here) is that they aren't Turing machines at all, and that the Turing machine model is entirely the wrong model to use.

Why? Well, because modern distributed systems have two things Turing machines don't have: concurrency and latency. Concurrency is, of course, the ability to have two or more tasks executing at the same time (this is slightly different than parallelism, in that the tasks don't have to be executing simultaneously, but can just be taking turns executing). Latency is the idea that new information is not available to all parts of the system simultaneously. However, by modeling systems as Turing machines, you're assuming that a) the system has a single state that b) is known simultaneously throughout the system; in other words, the system is consistent.

For example, if our system is a banking system, we would want it to be consistent in that the balance of a given account, for example, is always the same everywhere in the system. If we get a simultaneous deposit and withdrawal, the system will somehow process each and arrive at the correct balance all at once.

How do we achieve this magic in systems that may span continents, for instance? Databases. Databases are the transactional glue that holds this together. We treat the database as the "system of record", and, since it's transactional, we use it to enforce the transactionality of our system as a whole.

Which is where everything falls apart.

Modern Systems as an Extension of the Database

Because, what this means is that you're limiting the scale of the system to how well you can scale your database. Rather than increasing cohesion and reducing coupling, you're instead coupling your whole system together - tightly - through the database, because all parts of the system have to ensure that the system of record - the database - gets to the proper, consistent, state before continuing.

You can push the problem off, of course - there are scads of strategies for doing so. But it's still ignoring the fundamental problem, which is that you're trying to scale a fundamentally coupled system (or, to put it another, equivalent, way: you're trying to make an inherently inconsistent system consistent).

For instance, let's take a typical large-scale distributed system, with a UI layer, a business layer, and a database, distributed over several machines, maybe even over several data centers on different continents, with, however, a single, logical database which is the system of record.

Of course, since the network latency plus the database locking is a horrible performance hit, your first "optimization" is to cache the data locally, creating a caching key by hashing some combination of data values based on the table indexes. But, of course, since you don't want the system to get into an inconsistent state, you find ways to make sure the cache and the database are in agreement - write-through caches, for example, or cache replication - which are complicated to maintain in production, but necessary for consistency.

Then the next point of pain is that the database is too large, so you shard the database. But, again, since you want your system to be consistent, you have to make sure that each part of the system has a way of accessing the correct data shard for the correct data, which further complicates the situation. So you suddenly find yourself adding a lot of complexity around keeping your database updated with the correct information. And this is not even thinking about the possibility of network fragmentation.

Jewel of Denial

What to do? Well, to begin with, we can stop being in denial about the consistency of our systems. Once we realize that pursuing consistency is a foolish endeavor, and that eventual consistency is the best we can hope for, it becomes relatively simple: design our systems the way they actually should be designed - decoupled cohesive subsystems which will come to some agreement about the state of the overall system eventually. We need to understand that if one part of the system tells a user that her balance is X because it hasn't gotten the message from another part that it's actually Y it's not the end of the world, because users understand that things can't happen instantaneously.

In the case of the application above, we let each local system keep track of its own ideas of the state of the system-as-a-whole, while providing a mechanism for each subsystem to communicate its idea of the current state with each other subsystem, and some way of the whole system coming to some consensus of what the actual state is. Thus, when subsystem A gets a deposit, while subsystem B gets a withdrawal on the same account, each will have a different idea of the balance on the account until they update each other, at which time both will end up with the correct final balance. In return for this temporary inconsistency we will get much simpler systems which are both more performant (because of not having to come to immediate consensus on the state of the system, with the corresponding network latencies and lock contention) and more scalable (because the subsystems are truly decoupled from each other).

Jack Lund

Jack Lund

comments powered by Disqus