Summary:
Until now, we've been hiding transactions beneath the surface. When you call persistModel, you're implicitly creating a transaction.
You could explicitly create them with `atomically`..., but there were several critical problems that are fixed in this diff:
- Calling persistModel / unpersistModel within a transaction could cause the DatabaseStore to trigger. This could result in other parts of the app making queries /during/
the transaction, potentially before the COMMIT occurred and saved the changes. The new, explicit inTransaction syntax holds all changes until after COMMIT and then triggers.
- Calling atomically and then calling persistModel inside that resulted in us having to check whether a transaction was present and was gross.
- Many parts of the code ran extensive logic inside a promise chained within `atomically`:
BAD:
```
DatabaseStore.atomically =>
DatabaseStore.persistModel(draft) =>
GoMakeANetworkRequestThatReturnsAPromise
```
OVERWHELMINGLY BETTER:
```
DatabaseStore.inTransaction (t) =>
t.persistModel(draft)
.then =>
GoMakeANetworkRequestThatReturnsAPromise
```
Having explicit transactions also puts us on equal footing with Sequelize and other ORMs. Note that you /have/ to call DatabaseStore.inTransaction (t) =>. There is no other way to access the methods that let you alter the database. :-)
Other changes:
- This diff removes Message.labels and the Message-Labels table. We weren't using Message-level labels anywhere, and the table could grow very large.
- This diff changes the page size during initial sync from 250 => 200 in an effort to make transactions a bit faster.
Test Plan: Run tests!
Reviewers: juan, evan
Reviewed By: juan, evan
Differential Revision: https://phab.nylas.com/D2353
Electron 0.35.1 includes the tray fixes we contributed last week but also includes API restructuring and improvements. Most importantly, modules from electron are now imported via `require('electron')`
Summary:
Fixes T4291
If I made a final edit to a pre-existing draft and sent, we'd queue a
`SyncbackDraftTask` before a `SendDraftTask`. This is important because
since we have a valid draft `server_id`, the `SendDraftTask` will send by
server_id, not by POSTing the whole body.
If the `SyncbackDraftTask` fails, then we had a very serious issue whereby
the `SendDraftTask` would keep on sending. Unfortunately the server never
got the latest changes and sent the wrong version of the draft. This
incorrect version would show up later when the `/send` endpoint returned
the message that got actually sent.
The solution was to make any queued `SendDraftTask` fail if a dependent
`SyncbackDraftTask` failed.
This meant we needed to make the requirements for `shouldWaitForTask`
stricter, and block if tasks failed.
Unfortunatley there was no infrastructure in place to do this.
The first change was to change `shouldWaitForTask` to `isDependentTask`.
If we're going to fail when a dependent task fails, I wanted the method
name to reflect this.
Now, if a dependent task fails, we recursively check the dependency tree
(and check for cycles) and `dequeue` anything that needed that to succeed.
I chose `dequeue` as the default action because it seemed as though all
current uses of `shouldWaitForTask` really should bail if their
dependencies fail. It's possible you don't want your task dequeued in this
dependency case. You can return the special `Task.DO_NOT_DEQUEUE_ME`
constant from the `onDependentTaskError` method.
When a task gets dequeued because of the reason above, the
`onDependentTaskError` callback gets fired. This gives tasks like the
`SendDraftTask` a chance to notify the user that it bailed. Not all tasks
need to notify.
The next big issue was a better way to determine if a task truely errored
to the point that we need to dequeue dependencies. In the Developer Status
area we were showing tasks that had errored as "Green" because we caught
the error and resolved with `Task.Status.Finished`. This used to be fine
since nothing life-or-death cared if a task errored or not. Now that it
might cause abortions down the line, we needed a more robust method then
this.
For one I changed `Task.Status.Finished` to a variety of finish types
including `Task.Status.Success`. The way you "error" out is to `throw` or
`Promise.reject` an `Error` object from the `performRemote` method. This
allows us to propagate API errors up, and acts as a safety net that can
catch any malformed code or unexpected responses.
The developer bar now shows a much richer set of statuses instead of a
binary one, which was REALLY helpful in debugging this. We also record
when a Task got dequeued because of the conditions introduced here.
Once all this was working we still had an issue of sending old drafts.
If after a `SyncbackDraftTask` failed, now we'd block the send and notify
the users as such. However, if we tried to send again, there was a
separate issue whereby we wouldn't queue another `SyncbackDraftTask` to
update the server with the latest information. Since our changes were
persisted to the DB, we thought we had no changes, and therefore didn't
need to queue a `SyncbackDraftTask`.
The fix to this is to always force the creation of a `SyncbackDraftTask`
before send regardless of the state of the `DraftStoreProxy`.
Test Plan: new tests. Lots of manual testing
Reviewers: bengotow
Reviewed By: bengotow
Subscribers: mg
Maniphest Tasks: T4291
Differential Revision: https://phab.nylas.com/D2156
Summary:
Caveat: If you submit feedback from one account, and then switch accounts and
submit feedback again, the data from the old account still gets passed. It
looks like that was true with the previous compose-a-draft-with-feedback action
too. Maybe you know how to fix this? I couldn't figure it out.
Test Plan: Try it out! Not sure how to usefully write specs.
Reviewers: bengotow
Reviewed By: bengotow
Differential Revision: https://phab.nylas.com/D2106
Summary: Adds a prominent blue button to email us feedback
Test Plan: Run tests
Reviewers: dillon, evan
Reviewed By: dillon, evan
Differential Revision: https://phab.nylas.com/D2081
- Do not sort and then reverse the request history every time it's requested
- Do not allow the request history to grow forever (cap at 200)
- Do not stringify tasks until they're clicked
Summary: Fixes T3625
Test Plan: Easy to test by opening/closing reply options dropdown as a large image attachment is downloaded
Reviewers: evan, dillon
Reviewed By: dillon
Maniphest Tasks: T3625
Differential Revision: https://phab.nylas.com/D2034
Summary: Also enhancements to the developer toolbar
Test Plan: edgehill --test
Reviewers: dillon, bengotow
Reviewed By: bengotow
Differential Revision: https://phab.nylas.com/D1976
Summary:
Move sync workers and Edgehill token checks to work window
Move the task queue and database setup to the work window
Move ContactStore background refresh to work window
Store the task queue in the database
WIP
The TaskQueue now puts tasks in the database instead of in a file, which also means it can be observed
Move all delta sync and initial sync to a package, make NylasSyncStore which exposes read-only sync state
DraftStore no longer reads task status. Once you set the "sending" bit on a draft, it never gets unset. But that's fine actually.
If your package lists windowTypes, you *only* get loaded in those windowTypes. If you specify no windowTypes, you get loaded in the root window.
This means that onboarding, worker-ui, worker-sync, etc. no longer get loaded into the main window
ActivitySidebar has a special little store that observes the task queue since it's no longer in the window
Move "toggle component regions" / "toggle react remote" to the Developer menu
Move sync worker specs, update draft store specs to not rely on TaskQueue at all
Test Plan: Run existing tests, all pass
Reviewers: dillon, evan
Reviewed By: evan
Differential Revision: https://phab.nylas.com/D1936
