feat!: internally maintain provider status

[toddbaert]

This PR implements a few things from spec 0.8.0:

• implements internal provider status (already implemented in JS)
  • the provider no longer updates its status to READY/ERROR, etc after init (the SDK does this automatically)
  • the provider's state is updated according to the last event it fired
• adds PROVIDER_FATAL error and code
• adds "short circuit" feature when evaluations are skipped if provider is NOT_READY or FATAL
• removes some deprecations that were making the work harder since we already have pending breaking changes.

Fixes: #250

[kinyoklion]

Maybe more related to the spec than to this specific implementation. But I did find it somewhat hard to handle the runtime status after initialization status when the SDK automatically implements ready.

I think it makes it easy for simple providers, and somewhat cumbersome for more complex providers. I would actually prefer to be able to opt out.

Basically an underlying SDK may already emit these events, but we have to connect and translate these events conditionally around initialization. So after initialization if a provider fails we want to emit that event, but if during initialization it fails we need to not emit that event from the provider because the SDK also will emit it.

Example of attempting to reconcile this in a python provider: https://github.com/launchdarkly/openfeature-python-server/blob/243a8dc515bcd16fe8c7d5a0913fbdfe6e073aa0/ld_openfeature/provider.py#L46

[toddbaert]

So after initialization if a provider fails we want to emit that event, but if during initialization it fails we need to not emit that event from the provider because the SDK also will emit it.

I can see the issue here. We have some similar logic in flagd since it also supports automatic reconnection. I think for the majority of providers though, such events around reconnection etc, are not supported; it does indeed seem to be something featured in the most rigorous implementations.

I'm open to some kind of opt-out for emitting events on READY, but I'm not sure if the ROI is there. 🤔 I suppose it would be fairly easy to implement. 🤔

[kinyoklion]

So after initialization if a provider fails we want to emit that event, but if during initialization it fails we need to not emit that event from the provider because the SDK also will emit it.

I can see the issue here. We have some similar logic in flagd since it also supports automatic reconnection. I think for the majority of providers though, such events around reconnection etc, are not supported; it does indeed seem to be something featured in the most rigorous implementations.

I'm open to some kind of opt-out for emitting events on READY, but I'm not sure if the ROI is there. 🤔 I suppose it would be fairly easy to implement. 🤔

I think an opt-out would work.

I've not thought about it extensively, but another option may be to have automatic events be part of a base class that more complex providers could choose not to use instead of it being in the API. Though that path isn't the most sustainable unless you have a language that supports more "mixin" type classes. So maybe not.

Maybe also we can leave it as is and think if the situation can be improved. Overall I feel like getting initialization correct has become more complex, so maybe the abstraction needs adjusted some.

[toddbaert]

So after initialization if a provider fails we want to emit that event, but if during initialization it fails we need to not emit that event from the provider because the SDK also will emit it.

I can see the issue here. We have some similar logic in flagd since it also supports automatic reconnection. I think for the majority of providers though, such events around reconnection etc, are not supported; it does indeed seem to be something featured in the most rigorous implementations.
I'm open to some kind of opt-out for emitting events on READY, but I'm not sure if the ROI is there. 🤔 I suppose it would be fairly easy to implement. 🤔

I think an opt-out would work.

I've not thought about it extensively, but another option may be to have automatic events be part of a base class that more complex providers could choose not to use instead of it being in the API. Though that path isn't the most sustainable unless you have a language that supports more "mixin" type classes. So maybe not.

Maybe also we can leave it as is and think if the situation can be improved. Overall I feel like getting initialization correct has become more complex, so maybe the abstraction needs adjusted some.

I thought a bit about an opt-out. I think we could add some fields/metadata to the provider to signal this and possibly other behaviors. Perhaps we could open an new issue to discuss this in the spec.

[toddbaert]

I had to rebase this and resolve conflicts on the recently merged concellationToken/Async chnage.

[askpt]

I added a few comments and concerns about the ValueTask.

[askpt]

LGTM. I don't have any more concerns about this PR. Thanks @toddbaert