Our undo-redo system before this diff is based on commands. A command
is:
- A function that produces some data required to perform and undo a
change
- A function that actually performs the change, based on the data
- Another function that undoes the change, based on the data
- Optionally, a function to _redo_ the change, although in practice we
never use this
Each command that gets run is added to the undo/redo stack unless it
says it shouldn't be.
This diff replaces this system of commands with a new one where all
changes to the store are automatically recorded in the undo/redo stack.
You can imagine the new history manager like a tape recorder - it
automatically records everything that happens to the store in a special
diff, unless you "pause" the recording and ask it not to. Undo and redo
rewind/fast-forward the tape to certain marks.
As the command concept is gone, the things that were commands are now
just functions that manipulate the store.
One other change here is that the store's after-phase callbacks (and the
after-phase side-effects as a result) are now batched up and called at
the end of certain key operations. For example, `applyDiff` would
previously call all the `afterCreate` callbacks before making any
removals from the diff. Now, it (and anything else that uses
`store.atomic(fn)` will defer firing any after callbacks until the end
of an operation. before callbacks are still called part-way through
operations.
## Design options
Automatic recording is a fairly large big semantic change, particularly
to the standalone `store.put`/`store.remove` etc. commands. We could
instead make not-recording the default, and make recording opt-in
instead. However, I think auto-record-by-default is the right choice for
a few reasons:
1. Switching to a recording-based vs command-based undo-redo model is
fundamentally a big semantic change. In the past, `store.put` etc. were
always ignored. Now, regardless of whether we choose record-by-default
or ignore-by-default, the behaviour of `store.put` is _context_
dependant.
2. Switching to ignore-by-default means that either our commands don't
record undo/redo history any more (unless wrapped in
`editor.history.record`, a far larger semantic change) or they have to
always-record/all accept a history options bag. If we choose
always-record, we can't use commands within `history.ignore` as they'll
start recording again. If we choose the history options bag, we have to
accept those options in 10s of methods - basically the entire `Editor`
api surface.
Overall, given that some breaking semantic change here is unavoidable, I
think that record-by-default hits the right balance of tradeoffs. I
think it's a better API going forward, whilst also not being too
disruptive as the APIs it affects are very "deep" ones that we don't
typically encourage people to use.
### Change Type
- [x] `sdk` — Changes the tldraw SDK
- [x] `improvement` — Improving existing features
- [x] `galaxy brain` — Architectural changes
### Release Note
#### Breaking changes
##### 1. History Options
Previously, some (not all!) commands accepted a history options object
with `squashing`, `ephemeral`, and `preserveRedoStack` flags. Squashing
enabled/disabled a memory optimisation (storing individual commands vs
squashing them together). Ephemeral stopped a command from affecting the
undo/redo stack at all. Preserve redo stack stopped commands from wiping
the redo stack. These flags were never available consistently - some
commands had them and others didn't.
In this version, most of these flags have been removed. `squashing` is
gone entirely (everything squashes & does so much faster than before).
There were a couple of commands that had a special default - for
example, `updateInstanceState` used to default to being `ephemeral`.
Those maintain the defaults, but the options look a little different now
- `{ephemeral: true}` is now `{history: 'ignore'}` and
`{preserveRedoStack: true}` is now `{history:
'record-preserveRedoStack'}`.
If you were previously using these options in places where they've now
been removed, you can use wrap them with `editor.history.ignore(fn)` or
`editor.history.batch(fn, {history: 'record-preserveRedoStack'})`. For
example,
```ts
editor.nudgeShapes(..., { ephemeral: true })
```
can now be written as
```ts
editor.history.ignore(() => {
editor.nudgeShapes(...)
})
```
##### 2. Automatic recording
Previously, only commands (e.g. `editor.updateShapes` and things that
use it) were added to the undo/redo stack. Everything else (e.g.
`editor.store.put`) wasn't. Now, _everything_ that touches the store is
recorded in the undo/redo stack (unless it's part of
`mergeRemoteChanges`). You can use `editor.history.ignore(fn)` as above
if you want to make other changes to the store that aren't recorded -
this is short for `editor.history.batch(fn, {history: 'ignore'})`
When upgrading to this version of tldraw, you shouldn't need to change
anything unless you're using `store.put`, `store.remove`, or
`store.applyDiff` outside of `store.mergeRemoteChanges`. If you are, you
can preserve the functionality of those not being recorded by wrapping
them either in `mergeRemoteChanges` (if they're multiplayer-related) or
`history.ignore` as appropriate.
##### 3. Side effects
Before this diff, any changes in side-effects weren't captured by the
undo-redo stack. This was actually the motivation for this change in the
first place! But it's a pretty big change, and if you're using side
effects we recommend you double-check how they interact with undo/redo
before/after this change. To get the old behaviour back, wrap your side
effects in `editor.history.ignore`.
##### 4. Mark options
Previously, `editor.mark(id)` accepted two additional boolean
parameters: `onUndo` and `onRedo`. If these were set to false, then when
undoing or redoing we'd skip over that mark and keep going until we
found one with those values set to true. We've removed those options -
if you're using them, let us know and we'll figure out an alternative!
8151e6f586