Properly define CorrelationContext as an optional interface.

[carlosalberto]

From #420

It is not clear to me why we choose to not have all the functions interact explicitly with a "CorrelationContext" object and we want this path of interacting explicitly with Context. In my opinion this case is more natural for languages that explicitly pass Context around (like Go) but not that natural for Python where for example the ContextVar library does not work very well with this API.

This is a very important distinction that we want to make so languages keep a certain uniformity, and I do agree this the statement above, so we must add a note or a section that fully clarifies this one.

[carlosalberto]

Added it to the (incoming) milestone 0.4.0.

[carlosalberto]

Copying/pasting a comment from @bogdandrutu on why we should recommend implementations to implement this (from #420):

Here are some reasons why we need an explicit object called Correlation exposed:

• Current API makes impossible to move one item from a Context object to another. For example in case of a batch operation if there are 5 operations batched and I want to extract the Correlation from one object I need to extract all elements in a dictionary, then iterate over them and manually add 1 by 1.
• It makes a clear separation between the Context functionality and the Correlation functionality.
• Allow users to manually propagate individual items in the Context.
• Consistency with trace where we expose the Span object.
• Consistency with languages that requires to have an object defined like Python if you want to use ContextVar API.
• To avoid creating long Context chains (if implemented using a reference to the parent instead of copying all the elements), we can build instrumentation that creates a Context object once an adds all the items (like Span + CorrelationContext).

For an experiment you can look at https://github.com/grpc/grpc-java/blob/b7ccc0d14297f255557e67a4f0364480f4a80bcf/census/src/main/java/io/grpc/census/CensusStatsModule.java#L102 to see why a TagContext was needed in order to implement the tags propagation in gRPC.

[andrewhsu]

From the SIG mtg today, sounds like this one would be best to be bumped to the next milestone.

[andrewhsu]

from the issue triage mtg today with TC, allowing changes related to this issue for GA if editorial changes only