Back to Flow

Annotation Requirement

website/docs/lang/annotation-requirement.md

0.312.116.8 KB
Original Source

Note: As of version 0.199 Flow uses Local Type Inference as its inference algorithm. The rules in this section reflect the main design features in this inference scheme.

Flow tries to avoid requiring type annotation for parts of programs where types can easily be inferred from the immediate context of an expression, variable, parameter, etc.

Variable declarations

Take for example the following variable definition

js
const len = "abc".length;

All information necessary to infer the type of len is included in the initializer "abc".length. Flow will first determine that "abc" is a string, and then that the length property of a string is a number.

The same logic can be applied for all const-like initializations. Where things get a little more complicated is when variable initialization spans across multiple statements, for example in

js
declare const maybeString: ?string;

let len;
if (typeof maybeString === "string") {
  len = maybeString.length;
} else {
  len = 0;
}

Flow can still determine that len is a number, but in order to do so it looks ahead to multiple initializer statements. See section on variable declarations for details on how various initializer patterns determine the type of a variable, and when an annotation on a variable declaration is necessary.

Function Parameters

Unlike variable declarations, this kind of "lookahead" reasoning cannot be used to determine the type of function parameters. Consider the function

js
function getLength(x) {
  return x.length;
}

There are many kinds of x on which we could access and return a length property: an object with a length property, or a string, just to name a few. If later on in the program we had the following calls to getLength

js
getLength("abc");
getLength({length: 1});

one possible inference would be that x is a string | { length: number }. What this implies, however, is that the type of getLength is determined by any part of the current program. This kind of global reasoning can lead to surprising action-at-a-distance behavior, and so is avoided. Instead, Flow requires that function parameters are annotated. Failure to provide such a type annotation manifests as a [missing-local-annot] error on the parameter x, and the body of the function is checked with x: any:

js
function getLength(x) { // Error
  return x.length;
}

const n = getLength(1); // no error since getLength's parameter type is 'any'

To fix this error, one can simply annotate x as

js
function getLength(x: string) {
  return x.length;
}

The same requirement holds for class methods

js
class WrappedString {
  data: string;
  setStringNoAnnotation(x) { // Error
    this.data = x;
  }
  setString(x: string) {
    this.data = x;
  }
}

Contextual Typing {#toc-contextual-typing}

Function parameters do not always need to be explicitly annotated. In the case of a callback function to a function call, the parameter type can easily be inferred from the immediate context. Consider for example the following code

js
const arr = [0, 1, 2];
const arrPlusOne = arr.find(x => x % 2 === 1);

Flow infers that the type of arr is Array<number>. Combining this with the builtin information for Array.find, Flow can determine that the type of x => x % 2 === 1 needs to be number => unknown. This type acts as a hint for Flow and provides enough information to determine the type of x as number.

Any attendant annotation can potentially act as a hint to a function parameter, for example

js
const fn1: (x: number) => number = x => x + 1;

However, it is also possible that an annotation cannot be used as a function parameter hint:

js
const fn2: unknown = x => x + 1; // Error

In this example the unknown type simply does not include enough information to extract a candidate type for x.

Flow can infer the types for unannotated parameters even when they are nested within other expressions like objects. For example in

js
const fn3: {f: (number) => void} = {f: (x) => {x as string}}; // Error

Flow will infer number as the type of x, and so the cast fails.

Function Return Types {#toc-function-return-types}

Unlike function parameters, a function's return type does not need to be annotated in general. So the above definition of getLength won't raise any Flow errors.

There are, however, a couple of notable exceptions to this rule. The first one is class methods. If we included to the WrappedString class a getString method that returns the internal data property:

js
class WrappedString {
  data: string;
  getString(x: string) { // Error
    return this.data;
  }
}

Flow would complain that getString is missing an annotation on the return.

The second exception is recursive definitions. A trivial example of this would be

js
function foo() { // Error
  return bar();
}

function bar() {
  return foo();
}

The above code raises a [definition-cycle] error, which points to the two locations that form a dependency cycle, the two missing return annotations. Adding a return annotation to either function would resolve the issue.

Effectively, the requirement on an annotation for method returns is a special-case of the recursive definition restriction. The recursion is possible through access on this.

Generic Calls {#toc-generic-calls}

In calls to generic functions the type of the result may depend on the types of the values passed in as arguments. This section discusses how this result is computed, when type arguments are not explicitly provided.

Consider for example the definition

js
declare function map<T, U>(
  f: (T) => U,
  array: ReadonlyArray<T>,
): Array<U>;

and a potential call with arguments x => x + 1 and [1, 2, 3]:

js
map(x => x + 1, [1, 2, 3]);

Here Flow infers that the type of x is number.

Some other common examples of generic calls are calling the constructor of the generic Set class or calling useState from the React library:

js
const set = new Set([1, 2, 3]);

import {useState} from 'react';
component Example() {
  const [num, setNum] = useState(42);
  return null;
}

Flow here infers that the type of set is Set<number>, and that num and setNum are number and (number) => void, respectively.

Computing a Solution

Computing the result of a generic call amounts to:

  1. coming up with a solution for T and U that does not contain generic parts,
  2. replacing T and U with the solution in the signature of map, and
  3. performing a call to this new signature of map.

This process is designed with two goals in mind:

  • Soundness. The results need to lead to a correct call when we reach step (3).
  • Completeness. The types Flow produces need to be as precise and informative as possible, to ensure that other parts of the program will be successfully checked.

Let's see how these two goals come into play in the map example from above.

Flow detects that ReadonlyArray<T> needs to be compatible with the type of [1, 2, 3]. It can therefore infer that T is number.

With the knowledge of T it can now successfully check x => x + 1. The parameter x is contextually typed as number, and thus the result x + 1 is also a number. This final constraint allows us to compute U as a number too.

The new signature of map after replacing the generic parts with the above solution is

js
(f: (number) => number, array: ReadonlyArray<number>) => Array<number>

It is easy to see that the call would be successfully checked.

Errors during Polymorphic Calls

If the above process goes on smoothly, you should not be seeing any errors associated with the call. What happens though when this process fails?

There are two reasons why this process could fail:

Under-constrained Type Parameters

There are cases where Flow might not have enough information to decide the type of a type parameter. Let's examine again a call to the builtin generic Set class constructor, this time without passing any arguments:

js
const set = new Set(); // Error
set.add("abc");

During the call to new Set, we are not providing enough information for Flow to determine the type for T, even though the subsequent call to set.add clearly implies that T will be a string. Remember that inference of type arguments is local to the call, so Flow will not attempt to look ahead in later statements to determine this.

In the absence of information, Flow would be at liberty to infer any type as T: any, mixed, empty, etc. This kind of decision is undesirable, as it can lead to surprising results. For example, if we silently decided on Set<empty> then the call to set.add("abc") would fail with an incompatibility between string and empty, without a clear indication of where the empty came from.

So instead, in situations like this, you'll get an [underconstrained-implicit-instantiation] error. The way to fix this error is by adding a type annotation. There a few potential ways to do this:

  • Add an annotation at the call-site in one of two ways:

    • an explicit type argument
      js
      const set = new Set<string>();
      
    • an annotation on the initialization variable:
      js
      const set: Set<string> = new Set();
      
  • Add a default type on the type parameter T at the definition of the class:

    js
    declare class SetWithDefault<T = string> extends ReadonlySet<T> {
      constructor(iterable?: ?Iterable<T>): void;
      // more methods ...
    }
    

    In the absence of any type information at the call-site, Flow will use the default type of T as the inferred type argument:

    js
    const defaultSet = new SetWithDefault(); // defaultSet is SetWithDefault<string>
    

Incompatibility Errors

Even when Flow manages to infer non-generic types for the type parameters in a generic call, these types might still lead to incompatibilities either in the current call or in code later on.

For example, if we had the following call to map:

js
declare function map<T, U>(f: (T) => U, array: ReadonlyArray<T>): Array<U>;
map(x => x + 1, [{}]); // Error

Flow will infer T as {}, and therefore type x as {}. This will cause an error when checking the arrow function since the + operation is not allowed on objects.

Finally, a common source of errors is the case where the inferred type in a generic call is correct for the call itself, but not indicative of the expected use later in the code. For example, consider

js
import {useState} from 'react';
component Example() {
  const [str, setStr] = useState("");

  declare const maybeString: ?string;
  setStr(maybeString); // Error
  return null;
}

Passing the string "" to the call to useState makes Flow infer string as the type of the state. So setStr will also expect a string as input when called later on, and therefore passing a ?string will be an error.

Again, to fix this error it suffices to annotate the expected "wider" type of state when calling useState:

js
const [str, setStr] = useState<?string>("");

Module Exports {#toc-module-exports}

Flow builds type signatures for each module based solely on the module's exports, without analyzing the module's internal implementation. This process is called signature extraction: Flow constructs a "typed interface" for each module — a summary of every exported value's type — by looking only at the export site, not the module body. This is what enables Flow to typecheck modules in parallel and provide fast IDE feedback: each module's type can be understood without typechecking its dependencies' implementations.

For example, an exported function needs parameter and return annotations:

js
// Works: Flow can build a signature from the annotations
export function getLength(x: string): number {
  return x.length;
}

export function getLength2(x) { // Error: Flow cannot determine the parameter or return type
  return x.length;
}

Exported variables typically don't need annotations when their type can be determined from the initializer:

js
export const name = "Alice"; // OK: type is clearly 'string'
export const count = items.length; // OK: type determined from .length

If Flow cannot determine the type of an export, you'll see a [signature-verification-failure] error with the message "Cannot build a typed interface for this module." The fix is to add a type annotation to the export.

Unsupported expression forms in exports {#toc-unsupported-expression-forms}

Beyond missing annotations, certain expression forms are not supported in exported positions because Flow cannot determine their types during signature extraction. These typically produce a [signature-verification-failure] error with a specific sub-message:

  • Unexpected object key — exported objects must use simple literal keys (string or number literals, identifiers). Computed keys prevent Flow from knowing the shape of the object at the export site.

    js
    const key = "x";
    export const obj = {[key]: 1}; // Error: Expected simple key in object
    

    Fix: annotate the export, e.g., export const obj: {x: number} = {[key]: 1}.

  • Unexpected array spread — spreading into an exported array literal is not supported because Flow would need to evaluate the spread to determine the array's element type.

    js
    const xs = [1, 2];
    export const ys = [...xs, 3]; // Error: Unexpected spread in array
    

    Fix: add a type annotation, e.g., export const ys: Array<number> = [...xs, 3].

  • Array hole — exported arrays with holes (elisions like [1, , 3]) are not supported.

    js
    export const arr = [1, , 3]; // Error: Unexpected array hole
    

    Fix: fill in the hole or annotate the export, e.g., export const arr: Array<number | void> = [1, , 3].

  • Empty array — an empty array [] in an export position has no elements for Flow to infer an element type from.

    js
    export const items = []; // Error
    

    Fix: add a type annotation, e.g., export const items: Array<string> = [].

  • Unsupported expression — some complex expressions (e.g., conditionals, function calls, logical expressions) cannot be evaluated during signature extraction.

    js
    declare const condition: boolean;
    export const value = condition ? 1 : "a"; // Error
    

    Fix: add a type annotation, e.g., export const value: number | string = condition ? 1 : "a".

Generic Context {#toc-generic-context}

When a variable declared in an outer scope is only assigned inside a generic function, Flow cannot safely infer the variable's type. The assignments happen in a "generic context" — they involve type parameters that could take on any concrete type at each call site. Since the variable lives outside the generic function, Flow needs to know its type independently of any particular call.

You'll see an [invalid-declaration] error like: "Variable X should be annotated, because it is only initialized in a generic context."

js
let lastSeen; // Error: only initialized in a generic context
function remember<T>(value: T): T {
  lastSeen = value;
  return value;
}
console.log(lastSeen);

The fix is to add a type annotation to the variable:

js
let lastSeen: mixed; // Works!
function remember<T>(value: T): T {
  lastSeen = value;
  return value;
}
console.log(lastSeen);

This also applies when a variable is initialized to null and only otherwise assigned in a generic context — you'll see: "Variable X should be annotated, because it is only ever assigned to by null and in generic context."

js
let cached = null; // Error: only assigned by null and in generic context
function cache<T>(value: T): T {
  cached = value; // Error: T is incompatible with null
  return value;
}
console.log(cached);

As with the previous example, the fix is to annotate the variable:

js
let cached: mixed = null; // Works!
function cache<T>(value: T): T {
  cached = value;
  return value;
}
console.log(cached);

Empty Array Literals {#toc-empty-array-literals}

Empty array literals ([]) are handled specially in Flow. You can read about their behavior and requirements.

See Also {#toc-see-also}

  • Variables — how variable declarations determine when annotations are needed
  • Generics — generic return types are a common trigger for annotation requirements
  • Functions — function parameters typically require annotations