Summary:
This diff is designed to dramatically speed up new window load time for
all window types and reduce memory consumption of our hot windows.
Before this diff, windows loaded in ~3 seconds. They now boot in a couple
hundred milliseconds without requiring to keep hot windows around for
each and every type of popout window we want to load quickly.
One of the largest bottlenecks was the `require`ing and initializing of
everything in `NylasExports`.
I changed `NylasExports` to be entirely lazily-loaded. Drafts and tasks
now register their constructors with a `StoreRegistry` and the
`TaskRegistry`. This lets us explicitly choose a time to activate these
stores in the window initalization instead of whenever nylas-exports
happens to be required first.
Before, NylasExports was required first when components were first
rendering. This made initial render extremely slow and made the proposed
time picker popout slow.
By moving require into the very initial window boot, we can create a new
scheme of hot windows that are "half loaded". All of the expensive
require-ing and store initialization is done. All we need to do is
activate the packages for just the one window.
This means that the hot window scheme needs to fundamentally change from
have fully pre-loaded windows, to having half-loaded empty hot windows
that can get their window props overridden again.
This led to a major refactor of the WindowManager to support this new
window scheme.
Along the way the API of WindowManager was significantly simplifed.
Instead of a bunch of special-cased windows, there are now consistent
interfaces to get and `ensure` windows are created and displayed. This
DRYed up a lot of repeated logic around showing or creating core windows.
This also allowed the consolidation of the core window configurations into
one place for much easier reasoning about what's getting booted up.
When a hot window goes "live" and gets populated, we simply change the
`windowType`. This now re-triggers the loading of all of the packages for
the window. All of the loading time is now just for the packages that
window requires since core Nylas is there thanks to the hot window
mechanism.
Unfortunately loading all of the packages for the composer was still
unnaceptably slow. The major issue was that all of the composer plugins
were taking a long time to process and initialize. The solution was to
have the main composer load first, then trigger another window load
settings change to change the `windowType` that loads in all of the
plugins.
Another major bottleneck was the `RetinaImg` name lookup on disk. This
requires traversing the entire static folder synchronously on boot. This
is now done once when the main window loads and saved in a cache in the
browser process. Any secondary windows simply ask the backend for this
cache and save the filesystem access time.
The Paper Doc below is the current set of manual tests I'm doing to make
sure no window interactions (there are a lot of them!) regressed.
Test Plan: https://paper.dropbox.com/doc/Window-Refactor-UYsgvjgdXgVlTw8nXTr9h
Reviewers: juan, bengotow
Reviewed By: bengotow
Differential Revision: https://phab.nylas.com/D2916
Summary:
- Removes controlled focus in the composer!
- No React components ever perfom focus in lifecycle methods. Never again.
- A new `Utils.schedule({action, after, timeout})` helper makes it easy to say "setState or load draft, etc. and then focus"
- The DraftStore issues a focusDraft action after creating a draft, which causes the MessageList to focus and scroll to the desired composer, which itself decides which field to focus.
- The MessageList never focuses anything automatically.
- Refactors ComposerView apart — ComposerHeader handles all top fields, DraftSessionContainer handles draft session initialization and exposes props to ComposerView
- ComposerHeader now uses a KeyCommandRegion (with focusIn and focusOut) to do the expanding and collapsing of the participants fields. May rename that container very soon.
- Removes all CommandRegistry handling of tab and shift-tab. Unless you preventDefault, the browser does it's thing.
- Removes all tabIndexes greater than 1. This is an anti-pattern—assigning everything a tabIndex of 0 tells the browser to move between them based on their order in the DOM, and is almost always what you want.
- Adds "TabGroupRegion" which allows you to create a tab/shift-tabbing group, (so tabbing does not leave the active composer). Can't believe this isn't a browser feature.
Todos:
- Occasionally, clicking out of the composer contenteditable requires two clicks. This is because atomicEdit is restoring selection within the contenteditable and breaking blur.
- Because the ComposerView does not render until it has a draft, we're back to it being white in popout composers for a brief moment. We will fix this another way - all the "return unless draft" statements were untenable.
- Clicking a row in the thread list no longer shifts focus to the message list and focuses the last draft. This will be restored soon.
Test Plan: Broken
Reviewers: juan, evan
Reviewed By: juan, evan
Differential Revision: https://phab.nylas.com/D2814
