Pass prevContext param to componentDidUpdate

[bvaughn]

See this comment for my latest thoughts about this issue.

This PR resolves the issue of the missing prevContext param to componentDidUpdate calls.

Note that this approach is support option 1 below.

Options Considered

It seems like we have a high-level decision to make regarding how and if we support the prevContext parameter going forward. Options include:

1. Drop support for it: If we do this we should update the ReactContextValidator-test test to assert that the prevContext parameter is not passed and update to Stack to mirror this behavior.
2. Support it (option A): Store it as an expando property on instance (eg __reactInternalPrevContext). Adding to the instance avoids bloating Fibers for all of the types that don't need this property and keeps our Fiber-consuming functions monomorphic.
3. Support it (option B): Use stateNode to store a wrapper object around instance and previous context for the case of ClassComponents. This prevents all non-ClassComponent fibers from growing while also keeping fiber-consuming methods monomorphic. Since this only impacts ClassComponent it should hopefully introduce minimal forking.

Regarding continued support: At this time it does not seem like this parameter is being used by React components within Facebook but a quick Github search seems to suggest the parameter is being used externally (eg 1, 2, 3, etc.).

[sebmarkbage]

Sorry. We explicitly did not choose this strategy for the reason that it doesn't make sense to add to everything when only a few places change and it can be recreated while traversing anyway.

[bvaughn]

I've updated this PR description with some more context based on our face-to-face chat about this.

[gaearon]

a quick Github search seems to suggest the parameter is being used externally (eg 1, 2, 3, etc.).

Most of these appear to be copies of React, type definitions, and a few one-off projects with no community activity that use context directly despite us explicitly saying in the docs that it’s not supported. I wouldn’t worry too much about them because updates are already broken for context.

Edit: this is getting some traffic so I want to point out that if you relied on this feature, the “fix” to get it back is three lines of code. The benefit of dropping it is that it was broken in the first place (so if you relied on it, your program was likely buggy) and that it introduced an unnecessary memory cost for all components when only a tiny minority used it.

[gaearon]

"But there still exist some problems" — is this ready to get merged or do you want to add failing cases?

[bvaughn]

Hey Dan. 😄 I temporarily set this down because of the holidays. However I also kind of wanted to wait until others on the team weighed in regarding the 3 options above. Do you have any thoughts or preferences?

Sebastian and I chatted about this for a good hour or two on Thursday and it was not clear at the end of that talk which option was most appropriate. I thought this PR was worth considering because it fixes a backwards compatibility problem. I didn't realize the was such a big unopened question around it.

As for the issues hinted at- there are some problems with context|props being stale if an update is bailed out on midway (after render but before commit) and then later updated again by a higher priority change. I will write a failing test or two for that on Tuesday even if we don't fix it right away. (Started writing one already but haven't finished yet.) 😁

Happy holidays!

[gaearon]

Yea, no rush, just curious.
I’m inclined towards either this fix or not fixing it at all.

[bvaughn]

No worries. I should have left a clearer status on the PR before heading out.

I was starting to feel bike-sheddy about this on Thursday and I was hoping a few days away from it might lend some clarity.

[bvaughn]

Note to self: PR #8655 has implications on this that probably pushes us toward either option B or not passing prevContext to componentDidUpdate at all.

I haven't wrapped my head fully around this yet but I think that if it's possible for instance.props and instance.state inputs to be invalidated by swapping child and progressedChild then the same applies to any property stored on the instance.

[bvaughn]

I've been thinking more about the lifecycle methods and specific concerns about context in fiber.

First some assumptions I have:

• componentWillReceiveProps only exists to enable components to update state in reaction to a change in inputs (props and, unofficially, context).
• componentWillUpdate only exists to enable components to compare prev/next inputs to decide if anything internally needs to be set/reset in response to a specific input change (eg in react-virtualized I reset cached size and position data if certain inputs change).
• componentDidUpdate exists to enable components to do side-effect things like DOM manipulations (that rely on a render having already been done) or network requests (in response to a change in inputs).

Assuming the above assumptions are correct, I have some concerns about fiber and context.

I think it's not controversial to say that side-effect things (eg DOM manipulation, network requests) should only be done in componentDidUpdate. Other lifecycle methods (eg componentWillUpdate) can't be relied on because the work they do may be bailed out on.

What about other instance attributes? Using react-virtualized as an example again- I store some derived data on the instance. This data is outside of React's management/awareness and so it won't be reset when a Fiber is bailed out on. I think that PR #8655 mitigates this by explicitly resetting instance.props and instance.state before calling lifecycle hooks (or at least it will by the time it is merged). As long as I'm using the correct lifecycle methods, my instance attributes will be eventually consistent with inputs.

But what about instance attributes that are derived from context?

• If we remove the prevContext parameter from componentDidUpdate then how can a component safely react to a change in context?
• Even with that parameter present- if context isn't managed in the same way as props and state (PR [Fiber] Move memoization to begin phase #8655) then it could be inconsistent between the instance and the fiber.

I believe @sebmarkbage has some ideas/plans for context going forward that are pretty different from the current, unstable implementation. Hopefully these ideas do away with the way we're managing next/prev context and this whole comment becomes invalid- but I wanted to write it down while it was in my mind in case. @sebmarkbage, maybe we can chat more about this topic next week when you're back?

Update: Chatted in meatspace with @acdlite about this. He pointed out another existing stack limitation with fiber WRT shouldComponentUpdate. (Seems like context is already flaky but fiber will make it more so.) The only safe way to use context is to pass singleton/static props (eg the way react-redux Provider passes down store) that components subscribe to for changes.

I think I'm now leaning toward option 1 above (removing the context param entirely from the lifecycle hooks). I'm not sure about the timing for this though. Would we want to do this for the next release or hold off? If we want to maintain a semblance of backward compatibility in this area, we could move forward with this PR (given that context remains an unofficial/unsupported feature) and deprecate for both stack and fiber in a subsequent release.

[bvaughn]

I want to expand on my concern about instance attributes a little more while I'm thinking about it.

I think it's unsafe to do anything with side effects (even setting or updating "private" instance attributes) in componentWillReceiveProps; this method should only be used for updating state in response to a change in props. Anything else can be problematic.

(There may be operations that should only be done once- when an input changes from X to Y. Because we might call componentWillReceiveProps on the same instance twice with the same X-to-Y transition- due to how we reset inputs a la PR #8655- this type of operation is unsafe.)

This is also a potential problem if we store anything references to non-primitive types in state. Because our instance may run componentWillReceiveProps more than once for the same previous inputs- it's possible we take an action too many times (eg increment a counter more than once).

@sebmarkbage I'd like to chat with you about this next week if you're available. 😄

[bvaughn]

Ping @sebmarkbage. Now that you're back, can we chat about this?