Can’t Live with It, Can’t Live without It

(This post was originally created for the Erlang Solutions blog. The original can be found here)

I’d like to share some thoughts about Elixir’s with keyword. with is a wonderful tool, but in my experience it is a bit overused. To use it best, we must understand how it behaves in all cases. So, let’s briefly cover the basics, starting with pipes in Elixir.

Pipes are a wonderful abstraction

But like all tools, you should think about when it is best used…

Pipes are at their best when you expect your functions to accept and return basic values. But often we don’t have only simple values because we need to deal with error cases. For example:

region 
|> Module.fetch_companies()
|> Module.fetch_departments()
|> Enum.map(& &1.employee_count)
|> calculate_average()

If our fetch_* methods return list values there isn’t a problem. But often we fetch data from an external source, which means we introduce the possibility of an error. Generally in Elixir this means {:ok, _} tuples for success and {:error, _} tuples for failure. Using pipes that might become:

region
|> Module.fetch_companies()
|> case do
  {:ok, companies} -> Module.fetch_departments(companies)
  {:error, _} = error -> error
end
|> case do
  {:ok, departments} ->
    departments
    |> Enum.map(& &1.employee_count)
    |> calculate_average()
  {:error, _} = error -> error
end

Not horrible, but certainly not beautiful. Fortunately, Elixir has with!

with is a wonderful abstraction

But like all tools, you should think about when it’s best used…

with is at it’s best when dealing with the happy paths of a set of calls which all return similar things. What do I mean by that? Let’s look at what this code might look like using with?

with {:ok, companies} <- Module.fetch_companies(region),
     {:ok, departments} <- Module.fetch_departments(companies) do
  departments
  |> Enum.map(& &1.employee_count)
  |> calculate_average()
end

That’s definitely better!

  • We separated out the parts of our code which might fail (remember that failure is a sign of a side-effect and in functional programming we want to isolate side-effects).
  • The body is only the things that we don’t expect to fail.
  • We don’t need to explicitly deal with the {:error, _} cases (in this case with will return any clause values which don’t match the pattern before <-).

But this is a great example of a happy path where the set of calls all return similar things. But where are some examples of where we might go wrong with with?

Non-standard failure

What if Module.fetch_companies returns {:error, _} but Module.fetch_departments returns just :error? That means your with is going to return two different error results. If your with is the end of your function call then that complexity is now the caller’s responsibility. You might not think that’s a big deal because we can do this:

else
  :error -> {:error, "Error fetching departments"}

But this breaks to more-or-less important degrees because:

  • … once you add an else clause, you need to take care of every non-happy path case (e.g. above we should match the {:error, _} returned by Module.fetch_companies which we didn’t need to explicitly match before) 😤
  • … if either function is later refactored to return another pattern (e.g. {:error, _, _}) – there will be a WithClauseError exception (again, because once you add an else the fallback behavior of non-matching <- patterns doesn’t work) 🤷‍♂️
  • … if Module.fetch_departments is later refactored to return {:error, _} – we’ll then have an unused handler 🤷‍♂️
  • … if another clause is added which also returns :error the message Error fetching departments probably won’t be the right error 🙈
  • … if you want to refactor this code later, you need to understand everything that the called functions might potentially return, leading to code which is hard to refactor. If there are just two clauses and we’re just calling simple functions, that’s not as big of a deal. But with many with clauses which call complex functions, it can become a nightmare 🙀

So the first major thing to know when using with is what happens when a clause doesn’t match it’s pattern:

  • If else is not specified then the non-matching clause is returned.
  • If else is specified then the code for the first matching else pattern is evaluated. If no else pattern matches , a WithClauseError is raised.

As Stratus3D excellently put it: “with blocks are the only Elixir construct that implicitly uses the same else clauses to handle return values from different expressions. The lack of a one-to-one correspondence between an expression in the head of the with block and the clauses that handle its return values makes it impossible to know when each else clause will be used”. There are a couple of well known solutions to address this. One is using “tagged tuples”:

with {:fetch_companies, {:ok, companies} <- {:fetch_companies, Module.fetch_companies(region)},
     {:fetch_departments, {:ok, departments} <- {:fetch_departments, Module.fetch_departments(companies)},
  departments
  |> Enum.map(& &1.employee_count)
  |> calculate_average()
else
  {:fetch_companies, {:error, reason}} -> ...
  {:fetch_departments, :error} -> ...
end

Though tagged tuples should be avoided for various reasons:

  • They make the code a lot more verbose
  • else is now being used, so we need to match all patterns that might occur
  • We need to keep the clauses and else in sync when adding/removing/modifying clauses, leaving room for bugs.
  • Most importantly: the value in an abstraction like {:ok, _} / {:error, _} tuples is that you can handle things generically without needing to worry about the source

A generally better solution is to create functions which normalize the values matched in the patterns. This is covered well in a note in the docs for with and I recommend checking it out. One addition I would make: in the above case you could leave the Module.fetch_companies alone and just surround the Module.fetch_departments with a local fetch_departments to turn the :error into an {:error, reason}.

Non-standard success

We can even get unexpected results when with succeeds! To start let’s look at the parse/1 function from the excellent decimal library. It’s typespec tells us that it can return {Decimal.t(), binary()} or :error. If we want to match a decimal value without extra characters, we could have a with clause like this:

with {:ok, value} <- fetch_value(),
     {decimal, ""} <- Decimal.parse(value) do
  {:ok, decimal}

But if value is given as "1.23 " (with a space at the end), then Decimal.parse/1 will return {#Decimal<1.23>, " "}. Since that doesn’t match our pattern (string with a space vs. an empty string), the body of the with will be skipped. If we don’t have an else then instead of returning a {:ok, _} value, we return {#Decimal<1.23>, " "}.

The solution may seem simple: match on {decimal, _}! But then we match strings like "1.23a" which is what we were trying to avoid. Again, we’re likely better off defining a local parse_decimal function which returns {:ok, _} or {:error, _}.

There are other, similar, situations:

  • {:ok, %{"key" => value}} <- fetch_data(...) – the value inside of the {:ok, _} tuple may not have a "key" key.
  • [%{id: value}] <- fetch_data(...) – the list returned may have more or less than one item, or if it does only have one item it may not have the :id key
  • value when length(value) > 2 <- fetch_data(...) – the when might not match. There are two cases where this might surprise you:
    • If value is a list, the length of the list being 2 or below will return the list.
    • If value is a string, length isn’t a valid function (you’d probably want byte_size). Instead of an exception, the guard simply fails and the pattern doesn’t match.

The problem in all of these cases is that the intermediate value from fetch_data will be returned, not what the body of the with would return. This means that our with returns “uneven” results. We can handle these cases in the else, but again, once we introduce else we need to take care of all potential cases.

I might even go to the extent of recommending that you don’t define with clause patterns which are at all deep in their pattern matching unless you are very sure the success case will be able to match the whole pattern. One example where you might take a risk is when matching %MyStruct{key: value} <- … where you know that a MyStruct value is going to be returned and you know that key is one of the keys defined for the struct. No matter the case, dialyzer is one tool to gain confidence that you will be able to match on the pattern (at least for your own code or libraries which also use dialyzer).

One of the simplest and most standard ways to avoid these issues is to make sure the functions that you are calling return {:ok, variable} or {:error, reason} tuples. Then with can fall through cleanly (definitely check out Chris Keathley’s discussion of “Avoid else in with blocks” in his post “Good and Bad Elixir”).

With all that said, I recommend using with statements whenever you can! Just make sure that you think about fallback cases that might happen. Even better: write tests to cover all of your potential cases! If you can strike a balance and use with carefully, your code can be both cleaner and more reliable.

Tags

Updated: