Remotely loading Erlang modules

Sometimes, you’re connected to another node in an Erlang cluster, and you need to inject a new or replacement module. Here’s how.

Motivation

At Electric Imp, we used “blue/green” deployments: we bring up new servers and once we’re sure they’re working, we send the traffic to them. However, because IoT device connections are potentially long-lived, and because we didn’t just want to drop the connections (because reasons), we needed a way to gently disconnect devices from one of the old servers and send them to the new servers.

Then we realised that we hadn’t built that mechanism into the currently-deployed servers. Which was a problem.

Obviously, using Erlang, you can load new code at runtime. But, rather than pushing the new module to each old server (using scp), and then using the remote console (probably via tmux) to hot-load it, we wondered if there was a better way?

One of us noticed that Erlang allows you to load a module as a binary, which means that it doesn’t need to be a file on the server’s disk. And that you can do this from a remote node.

We wondered: could we just build this into the deploy script, so that it would simply inject the needed deployment module into the old servers? Every time, so we never needed to worry about compatibility between old and new?

Turns out: yes. So we did. In hindsight, there are probably better ways to do what we did, but having discovered that we could, we didn’t stop to think if we should.

Whatever. We’re not in Jurassic Park. Here’s how:

The code

% Load the module locally.
{module, Mod} = code:load_file(Mod).
{Mod, Bin, File} = code:get_object_code(Mod).

% Load the object code into the remote node.
{module, Mod} = rpc:call(Node, code, load_binary, [Mod, File, Bin]).

Yeah, that was a bit of an anti-climax after that build-up. Sorry.