The `waitUntilCalled` utility now has a timeout. It will now throw an
error if the stub is not called enough times, rather than blocking
forever.
The return type had to be changed to a function, so that we could throw
when the timeout is triggered. I tried returning an error that rejected
first, but if you don't handle the error synchronously Node.js will
consider it to be an unhandled Promise rejected (even if it _is_
handled later on).
I worked around this by resolving in the timeout case as well, so that
there is never a "deferred" Promise exception in the timeout case. The
returned function re-throws the error if it's given. That way there is
never any unhandled Promise rejection.
Some of the unit tests for the incoming transaction controller included
a 1 second wait. The wait was to ensure that a state update did not
occur, as it happens asynchronously.
The tests work equally well using a `setTimeout` with a zero second
wait, because the asynchronous block update is guaranteed to have been
queued up by the time this timeout function is called. The timeout has
been reduced to `0` to speed up the tests.
Additionally, `undefined` has been added to the list of network names
used to construct the fake API responses. This is to ensure that the
API returns a valid response, so that the test fails when it should.
The incoming transactions controller now uses the `chainId` for the
current network instead of the `networkId`. This ensures that custom
RPC endpoints for the built-in supported networks do correctly receive
incoming transactions.
As part of this change, the incoming transactions controller will also
cease keeping track of the "last block fetched" for networks that are
not supported. This piece of state never really represented the last
block fetched, as _no_ blocks were fetched for any such networks. It
been removed.
Unit tests have been added to the incoming transactions controller to
ensure that block updates are correctly resulting in state updates when
incoming transactions are enabled. All other events that trigger state
updates are tested as well.
The tests were written to be minimally dependent upon implementation
details of the controller itself. `nock` was used to mock the API
response from Etherscan. Each event is triggered asynchronously by
`sinon`, as in production they are likely only triggered
asynchronously.
This was extracted from #9583
This PR includes a new `wait-until-called` module meant to help with
writing asynchronous tests. It allows you to wait until a stub has been
called.
The shared mocks used previously in the incoming transaction controller
tests have been replaced with functions that can generate a new mock
for each test.
We should avoid ever sharing mocks between tests. It's quite easy for
a mock to get accidentally mutated or not correctly "reset" for the
next test, leading to test inter-dependencies and misleading results.
In particular, it is unsafe to share a `sinon` fake (e.g. a spy or
stub) because they can't be fully reset between tests. Or at least it's
difficult to reset them property, and it can't be done while also
following their recommendations for preventing memory leaks.
The spy API and all related state can be reset with `resetHistory`,
which can be called between each test. However `sinon` also recommends
calling `restore` after each test, and this will cause `sinon` to drop
its internal reference to the fake object, meaning any subsequent call
to `resetHistory` would fail. This is intentional, as it's required to
prevent memory from building up during the test run, but it also means
that sharing `sinon` fakes is particularly difficult to do safely.
Instead we should never share mocks in the first place, which has other
benefits anyway.
This was discovered while writing tests for #9583. I mistakenly
believed that `sinon.restore()` would reset the spy state, and this was
responsible for many hours of debugging test failures.
Previously all browser globals were allowed to be used anywhere by
ESLint because we had set the `env` property to `browser` in the ESLint
config. This has made it easy to accidentally use browser globals
(e.g. #8338), so it has been removed. Instead we now have a short list
of allowed globals.
All browser globals are now accessed as properties on `window`.
Unfortunately this change resulted in a few different confusing unit
test errors, as some of our unit tests setup assumed that a particular
global would be used via `window` or `global`. In particular,
`window.fetch` didn't work correctly because it wasn't patched by the
AbortController polyfill (only `global.fetch` was being patched).
The `jsdom-global` package we were using complicated matters by setting
all of the JSDOM `window` properties directly on `global`, overwriting
the `AbortController` for example.
The `helpers.js` test setup module has been simplified somewhat by
removing `jsdom-global` and constructing the JSDOM instance manually.
The JSDOM window is set on `window`, and a few properties are set on
`global` as well as needed by various dependencies. `node-fetch` and
the AbortController polyfill/patch now work as expected as well,
though `fetch` is only available on `window` now.
* Use @metamask/eslint-config@1.1.0
* Use eslint-plugin-mocha@6.2.2
* Mark root ESLint config as root
* Update Mocha ESLint rules with shared ESLint config
* Fix styling of `.transaction-list`
* Filter `incomingTxListSelector` by network as well
* Start and stop block tracker polling in incoming tx controller
* Add fetch with abort in bg for `IncomingTxController`
