This PR revamps how errors in signia are handled.
This was brought about by a situation that @MitjaBezensek encountered
where he added a reactor to a shape util class. During fuzz tests, that
reactor was being executed at times when the Editor was not in a usable
state (we had a minor hole in our sync rebase logic that allowed this,
fixed elsewhere) and the reactor was throwing errors because it
dereferenced a parent signal that relied on the page state
(getShapesInCurrentPage or whatever) when there were no page records in
the store.
The strange part was that even if we wrapped the body of the reactor
function in a try/catch, ignoring the error, we'd still see the error
bubble up somehow.
That was because the error was being thrown in a Computed derive
function, and those are evaluated independently (i.e. outside of the
reactor function) by signia as it traverses the dependency graph from
leaves to roots in the `haveParentsChanged()` internal function.
So the immediate fix was to make it so that `haveParentsChanged` ignores
errors somehow.
But the better fix involved completely revamping how signia handles
errors, and they work very much like how signia handles values now. i.e.
- signia still assumes that deriver functions are pure, and that if a
deriver function throws once it will throw again unless its parent
signals change value, so **it caches thrown errors for computed values**
and throws them again if .get() is called again before the parents
change
- it clears the history buffer if an error is thrown
- it does not allow errors to bubble during dirty checking i.e. inside
`haveParentsChanged` or while calculating diffs.
### Change Type
- [x] `patch` — Bug fix
- [ ] `minor` — New feature
- [ ] `major` — Breaking change
- [ ] `dependencies` — Changes to package dependencies[^1]
- [ ] `documentation` — Changes to the documentation only[^2]
- [ ] `tests` — Changes to any test code only[^2]
- [ ] `internal` — Any other changes that don't affect the published
package[^2]
- [ ] I don't know
[^1]: publishes a `patch` release, for devDependencies use `internal`
[^2]: will not publish a new version
### Test Plan
1. Add a step-by-step description of how to test your PR here.
2.
- [x] Unit Tests
- [ ] End to end tests
### Release Notes
- Add a brief release note for your PR here.
Finally removing all these deprecated getters ahead of the full release.
### Change Type
- [x] `major` — Breaking change
### Release Notes
- (Breaking) Removed deprecated getters.
This PR opts to split the big singleton out into other smaller
singletons so that we can revert the moving of the tsdoc comments that
happened in #2322
### Change Type
- [x] `patch` — Bug fix
[^1]: publishes a `patch` release, for devDependencies use `internal`
[^2]: will not publish a new version
One minor issue with signia is that it uses global state for
bookkeeping, so it is potentially disastrous if there is more than one
version of it included in a bundle.
To prevent that being an issue before we had a warning that would
trigger if signia detects multiple initializations.
> Multiple versions of @tldraw/state detected. This will cause
unexpected behavior. Please add "resolutions" (yarn/pnpm) or "overrides"
(npm) in your package.json to ensure only one version of @tldraw/state
is loaded.
Alas I think this warning triggers too often in development
environments, e.g. during HMR or janky bundlers.
Something that can prevent the need for this particular warning is
having a global singleton version of signia that we only instantiate
once, and then re-use that one on subsequent module initializations. We
didn't do this before because it has a few downsides:
- breaks HMR if you are working on signia itself, since updated modules
won't be used and you'll need to do a full refresh.
- introduces the possibility of breakage if we remove or even add APIs
to signia. We can't rely on having the latest version of signia be the
first to instantiate, and we can't allow later instantiations to take
precedence since atoms n stuff may have already been created with the
prior version. To mitigate this I've introduced a `apiVersion` const
that we can increment when we make any kind of additions or removals. If
there is a mismatch between the `apiVersion` in the global singleton vs
the currently-initializing module, then it throws.
Ultimately i think the pros outweigh the cons here, i.e. far fewer
people will see and have to deal with the error message shown above, and
fewer people should encounter a situation where the editor appears to
load but nothing changes when you interact with it.
### Change Type
- [x] `patch` — Bug fix
[^1]: publishes a `patch` release, for devDependencies use `internal`
[^2]: will not publish a new version
### Release Notes
- Make a global singleton for tlstate.
Follow up to #2189
### Change Type
- [ ] `patch` — Bug fix
- [ ] `minor` — New feature
- [ ] `major` — Breaking change
- [ ] `dependencies` — Changes to package dependencies[^1]
- [x] `documentation` — Changes to the documentation only[^2]
- [ ] `tests` — Changes to any test code only[^2]
- [ ] `internal` — Any other changes that don't affect the published
package[^2]
- [ ] I don't know
[^1]: publishes a `patch` release, for devDependencies use `internal`
[^2]: will not publish a new version
This PR replaces the `.value` getter for the atom with `.get()`
### Change Type
- [x] `major` — Breaking change
---------
Co-authored-by: David Sheldrick <d.j.sheldrick@gmail.com>
