Skip to content

Latest commit

 

History

History
372 lines (283 loc) · 13.9 KB

runtime.md

File metadata and controls

372 lines (283 loc) · 13.9 KB
Raw
id title
runtime
Enabling Runtime Checks

As we've mentioned before, Sorbet is a gradual system: it can be turned on and off at will. This means the predictions srb makes statically can be wrong.

That's why Sorbet also uses runtime checks: even if a static prediction was wrong, it will get checked during runtime, making things fail loudly and immediately, rather than silently and after the fact.

In this doc we'll answer:

  • What's the runtime effect of adding a sig to a method?
  • Why do we want to have a runtime effect?
  • What are our options if we don't want sigs to affect the runtime?

Runtime-checked sigs

Adding a method signature opts that method into runtime typechecks (in addition to opting it into more static checks). In this sense, sorbet-runtime is similar to libraries for adding runtime contracts.

Concretely, adding a sig wraps the method defined beneath it in a new method that:

  • validates the types of arguments passed in against the types in the sig
  • calls the original method
  • validates the return type of the original method against what was declared
  • returns what the original method returned1

For example:

require 'sorbet-runtime'

class Example
  extend T::Sig

  sig {params(x: Integer).returns(String)}
  def self.main(x)
    "Passed: #{x.to_s}"
  end
end

Example.main([]) # passing an Array!
❯ ruby example.rb
...
Parameter 'x': Expected type Integer, got type Array with unprintable value (TypeError)
Caller: example.rb:11
Definition: example.rb:6
...

In this small example, we have a main method defined to take an Integer, but we're passing an Array at the call site. When we run ruby on our example file, sorbet-runtime raises an exception because the signature was violated.

Why have runtime checks?

Runtime checks have been invaluable when developing Sorbet and rolling it out in large Ruby codebases like Stripe's. Type annotations in a codebase are near useless if developers don't trust them (consider how often YARD annotations fall out of sync with the code... 😰).

Adding a sig to a method is only as good as the predictions it lets srb make about a codebase. Wrong sigs are actively harmful. Specifically, when sigs in our codebase are wrong:

  • we can't use them to find code to refactor. Sorbet will think some code paths can never be reached when they actually can.
  • they're effectively as good as out-of-date documentation, with little added benefit over just comments.
  • we could never use them to make Ruby code run faster. In the future, we hope to use Sorbet types to make Ruby faster, but sigs that lie will actually make code slower than no types at all.

By leveraging runtime checks, we can gain lots of confidence and trust in our type annotations:

  • Automated test runs become tests of our type annotations!
  • Our production observability and monitoring catch bad sigs early, before they propagate false assumptions throughout a codebase.

To drive these points home, let's look at a concrete example:

# typed: true
require 'sorbet-runtime'

class Example
  extend T::Sig

  def self.some_untyped_method
    nil
  end

  sig {params(x: Integer).returns(Integer)}
  def self.add_one(x)
    x + 1
  end
end

Example.add_one(Example.some_untyped_method)

→ View on sorbet.run

In this example, srb tc reports that there were no errors statically. But if we were to run this code with ruby, some_untyped_method would return nil, we'd try to add 1 to it, and Ruby would raise a NoMethodError for +. By adding a sig to add_one, the Sorbet runtime will raise an exception before even starting to execute the method. This makes typing errors from untyped code manifest early and loudly and right at the source, rather than silently, long after a sig was added, and far removed from this line of code.

Most people are either familiar with a completely typed language (Java, Go, etc.) or a completely untyped language like Ruby; a gradual type system can be very foreign at first. Including these runtime checks by default protects typed code from untyped code, making it easier to drive adoption of types in the long run.

Changing the runtime behavior

While having runtime checks is the default, it's possible to change the behavior of the runtime system via configuration settings. These configuration settings live in under the T::Configuration module within the sorbet-runtime gem.

There are two main ways to change Sorbet's runtime:

  • When the runtime checks fail, what to do in response.

    To change this, we use .on_failure in a method signature.

  • Whether the runtime checks run in the first place.

    To change this, we use .checked in a method signature.

In the next sections, we'll give some examples of how to use both.

.on_failure: Changing what happens on runtime errors

By adding .on_failure(...) to a sig and registering a T::Configuration callback, we can change what happens when a sig check fails. For example:

require 'sorbet-runtime'

# (1) Register call_validation_error_handler callback.
# This runs every time a method with a sig fails to type check at runtime.
T::Configuration.call_validation_error_handler = lambda do |signature, opts|
  #                                                     ┌──┐
  if signature.on_failure && signature.on_failure[0] == :log
    puts opts[:pretty_message]
  else
    raise TypeError.new(opts[:pretty_message])
  end
end

class Main
  extend T::Sig

  # (2) Use .on_failure in the sig for a method
  #                                       ┌───────────────┐
  sig {params(argv: T::Array[String]).void.on_failure(:log)}
  def self.main(argv)
    puts argv
  end
end

# (3) When we call main incorrectly, it will print
# with puts instead of raise an exception.
Main.main(42)

# Output:
# ❯ ruby example.rb
# Parameter 'argv': Expected type T::Array[String], got type Integer with value 42
# Caller: example.rb:25
# Definition: example.rb:18
# 42

We defined our own meaning for .on_failure with T::Configuration. Without doing this, .on_failure has no effect. Because of this, .on_failure can be completely customized within any codebase to change what it means to fail. For example at Stripe, we use .on_failure to attach team ownership information to a failure.

The T::Configuration handler we wrote branches on whether .on_failure was provided, which means the logging behavior is opt in. This is nice because it means the default behavior is still to make problems with types fail loudly and early, rather than silently as a log. If we wanted, we could have inverted this:

T::Configuration.call_validation_error_handler = lambda do |signature, opts|
  #                                                     ┌────┐
  if signature.on_failure && signature.on_failure[0] == :raise
    raise TypeError.new(opts[:pretty_message])
  else
    puts opts[:pretty_message]
  end
end

# ...

#                                       ┌─────────────────┐
sig {params(argv: T::Array[String]).void.on_failure(:raise)}

With this T::Configuration handler, the default is to log, and we can use .on_failure to opt specific sigs into raising on failure.

We haven't depicted it here, but the T::Configuration handler will get an array of every argument that was provided to .on_failure for this sig—specifically there's no restriction that .on_failure must be given only one arg nor that the arg is a Symbol. For more on the various T::Configuration handlers, see Runtime Configuration.

.checked: Whether to check in the first place

Careful! Opting out of runtime checks can significantly degrade the trustworthiness of type signatures. Only disable the runtime after understanding the tradeoffs. See Gradual Type Checking to learn more.

In our examples above with .on_failure, every method call had the runtime type checks run to determine whether the T::Configuration handler should be called in the first place. This comes with an overhead, and we carefully audit and monitor the performance of sorbet-runtime as a result. The overhead of these checks are usually very small.

But in some cases, especially when calling certain methods in tight loops or other latency-sensitive paths, the overhead of even doing the checks (regardless of what happens on failure) is prohibitively expensive. To handle these cases, Sorbet offers .checked(...) which declares in what environments a sig should be checked:

# (1) Runtime checks always run.
sig {params(xs: T::Array[String]).void.checked(:always)}

# (2) Runtime checks only run in "tests" (see below).
# In non-tests, this sig specifically has no runtime overhead.
sig {params(xs: T::Array[String]).void.checked(:tests)}

# (3) Never runs the runtime checks. Careful!
sig {params(xs: T::Array[String]).void.checked(:never)}

If .checked(...) is omitted on a sig, the default is .checked(:always). The default checked level can also be configured. For example:

T::Configuration.default_checked_level = :tests

This can also be set via the SORBET_RUNTIME_DEFAULT_CHECKED_LEVEL environment variable, see Environment Variables for more.

Writing this will make it so that any sig which does not have a .checked(...) call in it will behave as if the user had written .checked(:tests). To prevent accidental misuse, sorbet-runtime will require that this setting is changed before any sig is evaluated at runtime.

Note: For .checked(:tests) to work correctly, checking in tests must be enabled in every entry point into the tests. To declare that a certain entry point is a test, run:

T::Configuration.enable_checking_for_sigs_marked_checked_tests

This can also be set via the SORBET_RUNTIME_ENABLE_CHECKING_IN_TESTS environment variable, see Environment Variables for more.

For example, this should probably be placed as the first line of any rake test target, as well as any other entry point to a project's tests. If this line is absent, .checked(:tests) sigs behave as if they had been .checked(:never).

T::Sig::WithoutRuntime.sig

Even with .checked(:never), there is some slight runtime overhead. The block for a sig {...} above a method is not evaluated until the first time that method is called. But Sorbet can only know whether a sig is .checked(:never) or not until the block is evaluated. So even if a sig is marked .checked(:never), Sorbet will still wrap the method. The first time the method is called, Sorbet will discover the .checked(:never) and put back the original method.

Sometimes even this tiny amount of runtime metaprogramming is unacceptable at runtime. To completely eliminate all runtime side effects when defining a signature, replace sig with T::Sig::WithoutRuntime.sig when annotating methods:

# typed: true
require "sorbet-runtime"

class Foo
  extend T::Sig
  # This signature will raise both statically and at runtime
  # (because `NotFound` doesn't exist, and causes a NameError)
  sig { params(x: NotFound).void.checked(:never) }
  def foo(x); end

  # This signature will only raise statically
  # (the sig block is ignored completely at runtime)
  T::Sig::WithoutRuntime.sig { params(x: NotFound).void }
  def bar(x); end
end

Note: For consistency in large codebases, it's best to avoid T::Sig::WithoutRuntime unless it can't be avoided, because the performance win versus simply using .checked(:never) is marginal. Some examples:

  • Circumstances make it impossible to put extend T::Sig in the class, but sorbet-runtime is still loaded.
  • The method is method_missing or respond_to_missing?, as sorbet-runtime does not support wrapping these methods.
  • Some code not under your control is trying to reflect on the result of .parameters or .arity of the sig'ed method. (Note that this precludes code under your control: Sorbet provides T::Utils.signature_for_method, which exposes the arity of the underlying method.)
  • After careful measurements, even the first call to a method is performance sensitive. This is rare, but can happen when trying to optimize speed of code loading specifically.

If none of these reasons apply, it's best to avoid T::Sig::WithoutRuntime, because it will stand out versus most other usages of sig in a codebase.

T.let, T.cast, T.must, T.bind

Type assertions like T.let, T.cast, T.must, and T.bind are normally checked at runtime, just like sig annotations on methods, unless runtime checks have been disabled.

Unlike method signatures, type assertions always have a performance cost, whether or not they are checked at runtime. See Type Assertions for tips on patterns that reduce or avoid this cost.

What's next?

  • Signatures

    Method signatures are the primary way that we add static and dynamic type checking in our code. Learn the available syntax advanced features of signatures.

  • Runtime Configuration

    Learn how to change or disable Sorbet's runtime type checks via settings and callbacks.

Footnotes

  1. The case for .void in a sig is slightly different. See the docs for void. ↩