As bad as anything else

It's About the Guarantees

A lot of times, when people think of Erlang supervisors, they think of restarts. This is, despite what the title of this entry may point at, very true. Jim Gray, from Tandem Computers, wrote in a paper that "In the measured period, one out of 132 software faults was a Bohrbug, the remainder Heisenbugs."

In complex production systems, most faults and errors are going to be transient, and retrying an operation will be a good way to do things — Jim Gray's paper quotes Mean Times Between Failures (MTBF) of systems handling transient bugs being better by a factor of 4. Still, supervisors aren't just about restarting.

I'll tell you what you can expect from me.

One very important part of Erlang supervisors and their supervision trees is that their start phase is synchronous. Each OTP Process started has a period during which it can do its own thing, preventing the entire boot sequence of its siblings and cousins to come. If the process dies there, it's retried again, and again, until it works, or fails too often.

That's where people make a very common mistake. There isn't a backoff or cooldown period before a supervisor restarts a crashed child. When people write a network-based application and try to set up a connection in this initialization phase, and that the remote service is down, the application fails to boot after too many fruitless restarts. Then the system may shut down.

Many Erlang developers end up arguing in favor of a supervisor that has a cooldown period. I strongly oppose the sentiment for one simple reason: it's all about the guarantees.

Restarting a process is about bringing it back to a stable, known state. From there, things can be retried. When the initialization isn't stable, supervision is worth very little. An initialized process should be stable no matter what happens. That way, when its siblings and cousins get started later on, they can be booted fully knowing that the rest of the system that came up before them is healthy.

If you don't provide that stable state, or if you were to start the entire system asynchronously, you get very little benefit from this structure that a try ... catch in a loop wouldn't provide.

What Should I do?

Supervised processes provide guarantees in their initialization phase, not a best effort. This means that when you're writing a client for a database or service, you shouldn't need a connection to be established as part of the initialization phase unless you're ready to say it will always be available no matter what happens.

You could force a connection during initialization if you know the database is on the same host and should be booted before your Erlang system, for example. Then a restart should work. In case of something incomprehensible and unexpected that breaks these guarantees, the node will end up crashing, which is desirable: a pre-condition to starting your system hasn't been met. It's a system-wide assertion that failed.

If, on the other hand, your database is on a remote host, you should expect the connection to fail. In this case, the only guarantee you can make in the client process is that your client will be able to handle requests, but not that it will communicate to the database. It could return {error, not_connected} on all calls during a net split, for example.

The reconnection to the database can then be done using whatever cooldown or backoff strategy you believe is optimal, without impacting the stability of the system. It can be attempted in the initialization phase as an optimization, but the process should be able to reconnect later on if anything ever disconnects.

If you expect failure to happen on an external service, do not make its presence a guarantee of your system. We're dealing with the real world here, and failure of external dependencies is always an option. To put it another way, we're going to lean on liveness (something good should eventually happen) as a guarantee, and reduce the safety (something bad will never happen) that we promise when it comes to the availability of services in a distributed system.

Of course, the libraries and processes that call the client will then error out if they didn't expect to work without a database. That's an entirely different issue in a different problem space, but not one that is always impossible to work around. For example, let's pretend the client is to a statistics service for Ops people — then the code that calls that client could very well ignore the errors without adverse effects to the system as a whole. In other cases, an event queue could be added in front of the client to avoid losing state when things go sour.

The difference in both initialization and supervision approaches is that the client's callers make the decision about how much failure they can tolerate, not the client itself. That's a very important distinction when it comes to designing fault-tolerant systems. Yes, supervisors are about restarts, but they should be about restarts to a stable known state.