At this point in the stack, we're not in control over what our
autocomplete providers are, what actions they'll do in order to
field our queries, etc. For example, some providers may hit the disk
and perform expensive DB queries internally. Some may even hit the
network, in theory!
In order to keep things perceptively speedy, let's run the actual work
off the main thread. This patch sets up a new pool thread to process
autocomplete requests. More than one thread is selected so that we maintain
liveliness during quick user input. Background tasks are cancelled as new
queries come in, and stale results are discarded.
- Remove a few unnecessary dependencies
- Add some comments around usage of 'api'
- Mark some of the dependencies as 'api' if they're necessary for use of the module
A storage implementation might need to cleanup its allocated resources,
and this provides an entry point for that to happen.
An example might be closing a native database connection that's maintained
by the Rust Places library.
Underlying HistoryStorage implementations might be blocking on IO (Rust Places, room),
or migth be non-blocking (in-memory). A Deferred return type for get* methods makes
it convenient to wrap both regular and `async` operations. Use of 'suspend' functions
serves the same role for write* methods.
This change makes it obvious at the engine call-site that the underlying implementation
might suspend, at which point we can wrap calls in `runBlocking` or use another coroutine builder.
