Bug 1649926 - Always enqueue an async task to change upload and deprecate getUploadEnabled

[badboy]

Replicating the commit messages for information:

Always enqueue an async task to change upload

That way it follows what a user expect when calling it in between other calls:
It executes in the right order.

Because the dispatch queue is halted until Glean is fully initialized
we can safely enqueue here and it will execute after initialization.

and

Deprecate `getUploadEnabled()` across all implementations

Due to Glean's asynchronous initialization the return value can be incorrect.
Applications should not rely on Glean's internal state.
Upload enabled status should be tracked by the application and communicated to Glean if it changes.
Note: The method was removed from the C# implementation.

---

This probably deserves some extra eyes on it. Putting you 3 on it, but maybe one or 2 should be enough.

I'm now reasonably sure it does what we expect it to do and the tests capture that.
However the tests are a bit brittle and (at least in some cases) might pass just fine on the old code base too.

Getting rid of getUploadEnabled means we also don't need to additionally track that upload enabled flag anymore.

Unfortunately I can't write a C# test because we don't have all the necessary utilities in place yet.
I did verify that adding GleanInstance.SetUploadEnabled(false); in the sample app just after Initialize does NOT send the custom ping, but DOES send the deletion-request ping.

[travis79]

I'd say this approach looks good to me (provided the iOS tests pass, which I assume they do). The one problem I see is that Lockwise-iOS does rely on getUploadEnabled in a weird way, that can be fixed if they ever update Glean.

[mdboom]

What happens in the following order of events?

Glean.setUploadEnabled(false)
Glean.initialize(uploadEnabled=true)

I would expect the state to be upload is enabled, but since the first call would be deferred until after the second, that might not happen (we should confirm with a test -- I could be wrong).

[Dexterp37]

I'm not sure if it's important to call out or not somewhere, but consider this scenario:

1. Initialize(uploadEnabled: true) is called
2. somePing.submit() is called right after
3. SetUploadEnabled(false) is called.

The ping at (2) might still get through, and I believe it's fine-ish. What do you think @badboy ?

[badboy]

Yes, that's the behavior I would expect. From a user perspective uploadEnabled was set to false only after the ping was submitted.

However ... (gosh this is complicated) ... because of the delay of the uploader picking it up it could still get cancelled :/

[badboy]

What happens in the following order of events?

Glean.setUploadEnabled(false)
Glean.initialize(uploadEnabled=true)

I would expect the state to be upload is enabled, but since the first call would be deferred until after the second, that might not happen (we should confirm with a test -- I could be wrong).

Damn, you are probably right. This is the one case why we had to preserve the state in an instance variable.
I need to think about it.

[badboy]

Ok, following up on what Mike said.

The positive first: It's easy to write a test!
Now the rest:
I think the only sensible solution is to actually ignore setUploadEnabled completely before Glean is initialized.
If that happens the order is (mostly) clear and the initialize should be the last operation, so is what should take precedence.
We already expect the uploadEnabled passed to initialize be correct. The whole reason we introduce that parameter was because people kept doing it wrong with setUploadEnabled.

I propose we add this to the top of setUploadEnabled:

        if (!isInitialized()) {
            Log.w(LOG_TAG, "Changing upload enabled before Glean is initialized is not supported.")
            return
        }

cc @Dexterp37 @mdboom

[badboy]

I spoke too soon.

We can't differentiate between

Glean.setUploadEnabled(false)
Glean.initialize(...)

and

Glean.initialize(...)
Glean.setUploadEnabled(false)

right now.

Do we really need more state? -_-

[badboy]

So yeah, if we track that initialize was called at all we might be able to prevent this properly.

wdyt?

[Dexterp37]

I propose we add this to the top of setUploadEnabled:

        if (!isInitialized()) {
            Log.w(LOG_TAG, "Changing upload enabled before Glean is initialized is not supported.")
            return
        }

cc @Dexterp37 @mdboom

I support this approach and like it. My only recommendation would be to have a more comprehensive message or add a link to our docs that explain the initial state should be provided with Glean.Initialize.

[badboy]

As I won't get this done today here's what's left:

• Incorporate the comments from above
  • true -> false in getUploadEnabled across all implementations
  • In Python make getUploadEnabled private
• Add a new variable initWasCalled (name bikesheddable), set to true before returning from initialize
• Check that new variable in setUploadEnabled
  • Print warning with extended message if initWasCalled == false

[Dexterp37]

r+ with a few optional nits! Great work on this one, thanks for your efforts. Please DO NOT MERGE just yet. I'd love to cut a release before this gets merged :)

[badboy]

Waiting with the rebase, given that there's a release incoming and thus I need to change the changelog anyway.