Finitomata FTW

• elixir 81

It’s been a year since I started to implement my own FSM library in Elixir. Now it’s funny to read my writings from the time when I had only a gut feeling I’m doing it right.

• The Proper FSM for Elixir
• First Class Documentation
• Finitomata ❤ Mox

Now I have a proof that my premonitions were correct.

When I shared my approach to keeping the local cache of some external incoming value as a bunch of Finitomata instances to my boss, he replied with the following.

The simple value as an FSM? Have you gone mad man?

No, I haven’t. And here is why.

We do receive the external values via some kind of stream, where we might receive a full snapshot for the key, a delta against the previously received snapshot for that key, and a delta against the snapshot for another key. The flow might contain tens thousands values per a second and the total number of keys might be up to one hundred thousand. This all might need more sophisticated infrastructure than a bare DynamicSupervisor or even a PartitionSupervisor. We need a handy lookup, and we might later decide to scale it horizontally.

Infinitomata runs transparently on the cluster, which means all one needs to use more memory/cores to process state transitions, would be adding new nodes to the cluster. Infinitomata will use them transparently upon addition.

But what’s wrong with leveraging a standard :pg, would probably ask a careful reader. After all, Infinitomata is built on top of :pg and we unlikely would need any FSM goodness to carry simple values. Negative. We definitely might rely on conditionals within values updates, but it would quickly become clumsy making it hard to reason about. On the contrary, we might keep the current state of the value in the FSM state, abstracting it out of the business logic. We won’t need to validate anything, because transitions are either possible (allowed,) or ignored.

As I said, we might receive a delta, which implies to get to the actual value, the snapshot must be already there. Yes, we might do somewhat along

value =
  if state.snapshot?, do: apply_delta(), else: :ignore

or even pattern-match in the function clause, but it’d be cumbersome, because deltas might rely on the snapshots for other keys. Instead, we might handle the transition for “local” delta if and only we already have a snapshot, meaning we are in the state ready. The transition handler for deltas relying on “external” keys would be as easy as

def on_transition(current_state, :foreign_delta, {foreign_key, delta_values}, state) do
  target_state =
    case Infinitomata.state(key) do
      {:error, :not_started} -> Logger.info("…")
      %{state: :idle} -> Logger.warn("…")
      %{state: :ready} -> apply_delta(delta_values)
    end

  if target_state == :ok,
    do: {:ok, current_state, state},
    else: {:ok, target_state, state}
end

This is clean, applies to :foreign_delta event only, and leaves the FSM in the correct state no matter what. We do gain the sequential processing for free as well. Also, there is no chance one forgets about some combination of state and event and leaves it unhandled. The :finitomata compiler would tell the FSM is inconsistent, if there are orphans or unhandled transitions.

All the on_transition/4 callbacks might be as a matter of fact implemented as a with/1 chain, logging the error, if any, and leaving the FSM in the state it was before the event came. There is no chance to end up with some inconsistency.

Even if the FSM definitions is as simple as

  @fsm """
  idle --> |snapshot| ready
  ready --> |snapshot| ready
  ready --> |delta| ready
  ready --> |foreign_delta| ready
  ready --> |shutdown| down 
  """

  use Finitomata, fsm: @fsm, auto_terminate: true

I prefer to use FSM because it abstracts a lot of boilerplate out and makes me to implement business logic only, with all the guarantees and validations, provided by Finitomata and all the necessary on_transition/4 handlers I am to implement, ready to pattern-match on meaningful states and events.

With a sleek learning curve, first-class test support with Finitomata.ExUnit, and plain text FSM declaration as both code and documentation, it’S a pleasure to deal with.

Happy infinite automating!