This is a set of best practices to keep in mind while writing a libdef. These best practices should also be considered when reviewing new libdefs or changes to existing libdefs.
Using the any type for a variable or interface results in the loss of type information as types pass through it. That means if a type passes through any before propogating on to other code, the any will potentially cause Flow to miss type errors!
In many places it is better (but also stricter) to use the mixed type rather than the any type. The mixed type is safer in that it allows anything to flow in to it, but can never be used downstream without dynamic type tests that verify the type at runtime.
Consider this code:
const cache = new Map();
function setCache(key: string, value: any) {
cache.set(key, value);
}
function getCache(key: string) {
cache.get(key);
}
setCache('foo', 42);
var someString = getCache('foo'); // oops, 'foo' isn't a string!
someString.length; // but because we used `any`, we don't know toerror here!Because we've typed the values of the map as any, we open ourselves up to type errors:
If, however, we had used mixed instead of any, Flow would ask us to verify that the type we got back from getCache() is in fact a string before using it as one:
const cache = new Map();
function setCache(key: string, value: mixed) {
cache.set(key, value);
}
function getCache(key: string) {
cache.get(key);
}
setCache('foo', 42);
var someString = getCache('foo');
if (typeof someString === 'string') {
someString.length;
} else {
throw new Error('`foo` is unexpectedly something other than a string!');
}Because we used a dynamic type test in the second example, we've proved to Flow that the mixed result must be a string. If we hadn't done this, Flow would have given us an error reminding us that the value could be anything and it isn't safe to assume without checking first.
There is sometimes a trade-off with using mixed, though. Using mixed means that if your variable ever does flow downstream, you'll always have to prove what kind of type it is (with a type test) before you can use it. Sometimes this is just too annoying and the risk involved with any is just worth the trade-off.
In the case of library definitions, it is almost always better to use mixed for function and method parameters because the trade-off is hidden from users:
declare function setCache(key: string, value: mixed): void;
setCache('number', 42); // Ok
setCache('string', 'asdf'); // OkIt is also almost always ok to use mixed as the return type of a callback:
declare function getUser(cb: (user: User) => mixed): void;
getUser((user) => console.log('Got the user!'));Using mixed in place of any for the return type of a function or the type of a variable is a judgement call, though. Return types and declared variables flow into users' programs, which means that users will have to prove the type of mixed before they can use them.
Right now we don't have a good way to write types inside declare module {} bodies that aren't exported. This problem is being worked on, but in the meantime the best option is to just put a declaration outside the declare module {} and reference it.
Because this effectively creates a global type, please be sure to namespace these types with something like $npm$ModuleName$:
type $npm$MyModule$Options = {
option1: number,
option2: string,
};
declare module "MyModule" {
declare function doStuff(options: $npm$MyModule$Options): void;
}All definitions are contained in the definitions/ directory and are named and structured very explicitly.
Let's say you've created a definition for an npm package called left-pad at
version 4.1.0. You've followed all of the
best practices for writing high quality definitions,
and now you'd like to contribute it:
We call this the "package version directory".
The naming scheme of this directory must be formatted as
${packageName}_v${packageVersion}. This convention is enforced by the
library-definition test runner.
We call this the "flow version directory".
This specifies that the definition you are contributing is compatible with all
versions of Flow. If it is not, you can specify a version range with names like
flow_>=v0.21.x ("any version at or after v0.21.x") or
flow_>=v0.21.x_<=v0.24.x ("any version inclusively between v0.21.x and
v0.24.x").
If you aren't sure which versions of Flow your definition is compatible with,
start with flow_all and the test runner (which we'll run in a later step) will
tell us if there are problems in some versions of Flow.
You may create multiple flow version directories to target different versions of Flow if necessary.
The name of the definition file must match the name of the package version directory two levels up. This, too, will be enforced by the library-definition test runner.
The purpose of a test file is to exercise the library definition to some reasonable degree. At minimum your tests should:
- Import at least one thing from each of the modules
declared in the definition. - Use the library definition in a couple of common ways that should not produce a type error.
- Use the library definition in a couple of ways that are expected to produce
a type error. To indicate that a line in a test file expected to cause an
error just put a
// $ExpectErrorcomment above the error-causing line.
Here is an example of a nice and thorough test file. You don't necessarily have to be this thorough, but the more thorough you are the better!
You can write as many test files as you like as long as their names start with
test_.
All test files that sit in the package version directory will be run by
the test-runner for all versions of Flow the package version supports. If you
ever need to write a test for a particular version of Flow, you can put the
test_ file in the appropriate flow version directory.
You may also leave off the argument to run all tests (this takes a while). Please note that this test (and the one on Travis-CI) only will be able to run if the name of the repo folder is still "flow-typed".