Refactor context

[reyang]

This is to support census-instrumentation/opencensus-specs#247.

1. Provide a generic context, which is maintained collaboratively by the Python runtime and the OpenCensus library.
2. Provide asyncio/async/await support.
3. Refactor existing context (span_context, trace/execution_context, tags/execution_context, correlation_context).
4. Provide callback API for span start/end.

@c24t @bogdandrutu @discostu105

[bogdandrutu]

I think this should be under context/ not under contrib. The core library should not depend on a contrib package.

[reyang]

@bogdandrutu

1. Do you mean we want to have this package under context/ instead contrib/ to avoid the confusion?
2. Or, do you mean this context should be part of the opencensus package instead of having its own package?

Having a small package of the generic context encourages other libraries to interop (similar like gRPC Context in Java) without having to depend on OpenCensus.

[c24t]

I just talked to @bogdandrutu, he means putting the package under opencensus/context/ or a new top level context/.

[reyang]

Thanks for clarification @c24t ! Please let me know your preference on the naming and folder structure.

Considering we will need to move CorrelationContext and TraceContext as well, probably we just rename contrib to a general name packages?

[reyang]

Personally I like the way how OpenCensus Node.js is organized, so we avoid inventing 1st class and 2nd class citizens. It also gives us flexibility on having opencensus-api and opencensus-impl in the future.

[c24t]

Moving everything into a top-level packages/ sounds fine to me, and would make it easy for people browsing the library to find a particular package quickly. But I'm more than happy having first-class opencensus/ and context/ dirs and a second-class contrib/. Up to you.

[bogdandrutu]

I preferred the way @c24t proposed.

@reyang why do we need to move the CorrelationContext there? Correlation context is part of opencensus/tags. I am not sure I understand.

[reyang]

@bogdandrutu CorrelationContext is a separate package which implements the W3C spec.

The opencensus/tags will be implemented leveraging CorrelationContext, while CorrelationContext itself could be used for other purpose (e.g. application developer can put TransactionId and UserIdentity).

[reyang]

@bogdandrutu @c24t I don't have strong opinion on the naming and folder structure, so let's use the way @c24t proposed.

[bogdandrutu]

@reyang I am very confused, my understanding is completely different. The current opencensus/tags are used exactly for this purpose: can be used to construct stats (metrics) or appended to logs or to Spans.

[reyang]

@bogdandrutu If the user wants to propagate a TransactionId within/across the process, which has nothing to do with trace/logs/metrics, do we expect CorrelationContext to be used, or we expect opencensus/tags to be used?

[bogdandrutu]

If TransactionId is used to correlate different telemetry events than that should be in what we currently call tags, if that is for a different purpose I expect them to use a different header.

[reyang]

It is for a different purpose (not to correlate telemetry, but to tell the transaction system about the distributed transaction scope).
Do we expect user to use W3C correlation context, or another header should be used (which we don't cover)?
According to the spec, such context could be user's identity, flight name. The spec doesn't require things in the context to be consumed by the telemetry system.