Summary:
It's time to ship the new updater! `IGListExperimentalAdapterUpdater` has been running on Instagram for a couple months with better performance and stability. In this diff, we're renaming `IGListExperimentalAdapterUpdater` (and related classes) to `IGListAdapterUpdater`.
Here's a recap:
## Stability
* `[IGListAdapter setDataSource]` isn't safe
* We're changing the underlying data without telling the `UICollectionView`.
* Fix: Lets invalidate the `UICollectionView` data by changing its dataSource.
* `[IGListAdapter setCollectionView]` isn't safe
* This is synchronous, but we might have pending or on-going updates. The `UICollectionView` might get synced before the pending update actually start executing, so the diff results will be off.
* Fix: Lets wrap updates in a transaction that can be cancelled.
* Returning a nil `IGListSectionController` from `IGListAdapterDataSource` could crash
* The `IGListAdapterUpdater` will still perform the diffing assuming that all the objects will have a section, which isn't the case.
* Fix: Lets generate the `IGListSectionController` before the diffing.
* Other improvements
* Lets ask for the `fromObject` just before diffing, instead of asking when scheduling the update.
* If the `UICollectionView` section count doesn't match `fromObject`, lets fallback to a reload.
## Performance
* Re-test background diffing
* `IGListExperimentBackgroundDiffing` and coalescing updates wasn't safe because of the issues mentioned above. The longer we wait, the more likely we'll end up in a race condition. Lets try re-testing with the stability improvements.
* Unblocks background layout calculation
* This is a larger project, but these improvements are required to make background work safe.
* Only create the `backgroundView` if needed (although this doesn't really require the new updater)
## Other
* Transactions
* `IGListAdapterUpdater` is the workhorse of `IGListKit` and has become a bit hard to follow over the years. We want to break it apart into simpler, more manageable parts.
* Avoid blocks
* There's a lot of blocks flying around, making crash logs hard to read. Lets try to use methods/functions where possible.
Reviewed By: Haud, lorixx
Differential Revision: D25884782
fbshipit-source-id: 1357fa23513a239051d5b1766823effa3199f656
Summary: This makes sure the `UICollectionView` section count matches what we expect before applying a diff results, if not, we fallback to a `reloadData`. This doesn't decrease the crash rate significantly, but it's nice to have a last line of defense. We do tradeoff crashes for performance issues, but it's better that an app works at all and we'll see the issue via asserts.
Reviewed By: lorixx
Differential Revision: D25884778
fbshipit-source-id: 5011d0907ce0f971ea3a0bf95c1549d52f615982
Summary: This experiment shows a slight decrease in frame-drops on features that have more expensive hashing, so lets ship!
Reviewed By: lorixx
Differential Revision: D25884785
fbshipit-source-id: a5e33abe6f129166ab9b75de80636db801599b74
Summary: `NSIndexPath` is in UIKit, so we should be importing its header.
Reviewed By: lorixx
Differential Revision: D25334717
fbshipit-source-id: 5a134a0ddbe88fddb33adfa182d7aaf437bc47d0
Summary:
Another source-of-truth inconsistency issue.
I kept getting this assertion T65423827 task.
It turned out the the root cause is that we use the `_viewSectionControllerMap` to keep track of the mapping between view and section controller.
Which is dangerous.
Because the UICV's source of truth is the `sectionMap`, now we are adding another source of truth between view <-> sectionController to be `_viewSectionControllerMap`, this is just waiting for bugs to happen.
We should always refers to sectionMap as the source of truth for everything here: since we can get the section index, we can replace all the `-sectionControllerForView` to use `-sectionControllerForSection:` instead, which is safer.
Multiple source of truth is the bug for this unexpected assertion.
Once we can cleanup this, then we dont need to use that weird: `- (void)mapView:(UICollectionReusableView *)view toSectionController:(IGListSectionController *)sectionController` API which we might forget to map, that can cause bug.
Reviewed By: elfredpagan
Differential Revision: D24173911
fbshipit-source-id: 6138cafc0b382fc569d17b14b13a6b50d85d342d
Summary: We don't actually need to call `[UICollectionVew performBatchUpdates ...]` if we don't have changes.
Reviewed By: patters
Differential Revision: D23386181
fbshipit-source-id: 469a22a35b9ead5181e95f36a8bc59ba36373bbe
Summary:
We're seeing some crashes where the `UICollectionView`'s section count doesn't match the `fromObjets`. Ideally that shouldn't happen, but in case it does, lets fallback to a `reloadData` instead of crashing. Lets `IGFailAssert()` to keep an eye on it.
EDIT: Turns out `UICollectionView` will sometimes already do that for item-level inconsistencies! Example error message:
> The number of items contained in an existing section after the update (9) must be equal to the number of items contained in that section before the update (10), plus or minus the number of items inserted or deleted from that section (0 inserted, 0 deleted) and plus or minus the number of items moved into or out of that section (0 moved in, 0 moved out). - will perform reloadData.
But it still throws a `NSInternalInconsistencyException` for section-level inconsistencies.
Reviewed By: patters
Differential Revision: D23386178
fbshipit-source-id: a9930b619701cffd37d0d6e2bcb1268ef7c9d0ce
Summary:
For diffing, `IGListAdapterUpdater` uses the full `fromObjects` instead of `validObjects`, which can be different if we're missing `IGListSectionController`. The diffing results won't match what the `IGListAdapter` tells the `UICollectionView`, so we can crash.
To fix this, lets try to generate the `IGListSectionController` before performing the diffing, and keep them in a container object (`IGListTransitionData`) alonside the `fromObjects` and `toObjects`. This might also unblock other cool features in the future, like giving `IGListSectionControllers` time for background work, but that's for another day.
Since this change is relatively large, lets create a temporary protocol `IGListUpdatingDelegateExperimental` which will have a slightly different API from `IGListUpdatingDelegate`. When/if the new implementation ships, we can merge both protocols.
Why not edit `IGListAdapterUpdater` directly?
* I don't want to mess up exiting updaters just yet.
Reviewed By: patters
Differential Revision: D23145776
fbshipit-source-id: 21642cafa64e2a26a173e2ba683ef5c6cede17d7
Summary: We pass `experiments` to places that don't need it, so lets clean that up.
Reviewed By: patters
Differential Revision: D23145782
fbshipit-source-id: affcfe0f38d1b8e544940af89e5692dbdc36dd76
Summary:
On `IGListAdapater` data update:
1) Pass the capacity count, so that arrays don't have to re-size.
2) Avoid using a set, so that we don't need to deal with hashes and equality. The updater should have dealt with duplicates already.
Reviewed By: patters
Differential Revision: D23145771
fbshipit-source-id: 2ed93231e15ddcd66cfe4d1f7384c563c77caa8e
Summary: We call `[collectionView layoutIfNeeded]` before performing updates, but it's not clear why. It leads to some strange animation issues when performing updates while scrolling. So lets try to skip it?
Reviewed By: Haud
Differential Revision: D22244053
fbshipit-source-id: 0c01a73860ebc16f430fed04d5c29a5bde038286
Summary: This experiment shipped so we can remove it. We still need a way to disable it with `allowsReloadingOnTooManyUpdates` in case we don't want to resign the first responder on large updates.
Reviewed By: Haud
Differential Revision: D22219238
fbshipit-source-id: 98e78f3ed040809db6c4b4c8da8fd0e976aca0a2
Summary: Actually this won't work and causes even more crashes. For example, if we reload cell 0 and insert at 0, we end up with just a single delete & insert at 0, which causes an inconsistency exception. `IGListApplyUpdatesToCollectionView` already takes care of avoiding multiple reloads of the same `NSIndexPath` by using a `NSMutableSet` when collecting the `reloadDeletePaths`.
Reviewed By: apadalko
Differential Revision: D21822502
fbshipit-source-id: c2382951ffe013f82c1de2492286d2911f40444d
Summary: Ship this experiment. `IGListCollectionViewLayout` should get the section/index counts via `UICollectionView` to stay in sync, instead of the `dataSource`.
Reviewed By: apadalko
Differential Revision: D21789871
fbshipit-source-id: 9fe069bd793b36680cd8562e69c7367d69a11ec7
Summary: This allows us to potentially enable `-Wundef` in the future. Behavior is the same, undefined macros are false.
Differential Revision: D20906603
fbshipit-source-id: d5798861f0c9dfcd7a87936d575621d52c0ef7c7
Summary:
**Context**
Recently an IGListKit experiment was introduced, `IGListExperimentPerformUpdatesWithoutDeferringCATransactionCommit`. When enabled, we opt to no longer defer a call to `[CATransaction commit]`, because it could feasibly end a different transaction than intended. In practice, this was leading to issues with `UIViewPropertyAnimator`, where deffered commits were ending in-progress animators.
**This Change**
The results we have seen from enabling this fix show no changes to performance and stability, so this seems safe to roll out. This change removes the experiment, and enables the new, non-deffered behavior.
Reviewed By: lorixx
Differential Revision: D20120169
fbshipit-source-id: 0473652020a250d67b02b860fb74c73e43615aef
Summary:
**Context**
When performing non-animated updates, `IGListKit` does so by beginning a `CATransaction` before `performBatchUpdates:`, and ending it in the completion block. While this disables animations for the collection view updates, it is feasible that some other changes might be bundled into the same transaction, given that we have no guarantee of the completion block being called inline.
It seems that this is happening in practice and occasionally disrupting `UIViewPropertyAnimator` animations, forcing the completion immediately.
**This Change**
To fix this issue, we can commit the transaction inline. Here we use `performWithoutAnimation:`, although it's possible that this is just a wrapper around `CATransaction` calls anyway.
Because the change may have unknown side effects, this change is gated by an experiment.
Reviewed By: maxolls
Differential Revision: D19150307
fbshipit-source-id: f7cff46ac7141609f42caafaf69bad71a980ad29
Summary:
Pull Request resolved: https://github.com/Instagram/IGListKit/pull/1377
This splits the diffing-only components from `IGListKit` out into a separate library, `IGListDiffKit`, allowing them to be used independently of the rest of `IGListKit`. We've found that the diffing components of the library are useful independently of the rest of the `UICollectionView` infrastructure, and separating the libraries allows apps to take advantage of those components without paying the full binary size cost of `IGListKit`. The diffing components are available as a subspec in Cocoapods, and as an independent framework in Carthage and direct Xcode project import (i.e. Git submodules).
This is a breaking change, necessitating a major version bump, although it should have only minor impact for projects using the umbrella header `IGListKit.h`, as that header continues to import the diffing utilities. In particular, for a project using Cocoapods and importing `IGListKit.h`, there should be no code changes required.
Reviewed By: calimarkus
Differential Revision: D18114249
fbshipit-source-id: 363b5a427e32800bbc6e82f897230963b4167245
