Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Moving contextvars and threadlocal context implementations to the API #419

Merged
merged 4 commits into from
Feb 18, 2020

Conversation

codeboten
Copy link
Contributor

@codeboten codeboten commented Feb 14, 2020

While keeping the flexibility of the interface to support additional implementations of the RuntimeContext, moving the ContextVarsRuntimeContext and ThreadLocalRuntimeContext into the API ensures a useful context implementation for users without requiring additional configuration. This changes checks the version number to determine which implementation to use, unless it is overridden by the OPENTELEMETRY_CONTEXT environment variable. Removing the DefaultRuntimeContext as part of this change.

The discussion suggesting this change can be found here

While keeping the flexibility of the interface to support additional implementations of the RuntimeContext, moving the ContextVarsRuntimeContext and ThreadLocalRuntimeContext into the API ensures a useful context implementation for users without requiring additional configuration. This changes checks the version number to determine which implementation to use, unless it is overridden by the OPENTELEMETRY_CONTEXT environment variable. Removing the DefaultRuntimeContext as part of this change.

Signed-off-by: Alex Boten <aboten@lightstep.com>
@codeboten codeboten requested a review from c24t February 14, 2020 00:12
@codeboten codeboten requested a review from a team as a code owner February 14, 2020 00:12
@Oberon00
Copy link
Member

I think we should follow what Java does in open-telemetry/opentelemetry-java#720 and partially has already done (https://github.com/open-telemetry/opentelemetry-java/tree/master/context_prop): Move the context to its own distribution that depends on nothing and on which API depends.

@codeboten
Copy link
Contributor Author

@Oberon00 I like this, were you thinking another top level directory for this?

Signed-off-by: Alex Boten <aboten@lightstep.com>
@ocelotl
Copy link
Contributor

ocelotl commented Feb 14, 2020

I think we should follow what Java does in open-telemetry/opentelemetry-java#720 and partially has already done (https://github.com/open-telemetry/opentelemetry-java/tree/master/context_prop): Move the context to its own distribution that depends on nothing and on which API depends.

@Oberon00 are you talking about moving the abstract RuntimeContext class into its own distribution or about moving the ContextVarsRuntimeContext and ThreadLocalRuntimeContext implementations or the three of them?

Signed-off-by: Alex Boten <aboten@lightstep.com>
@codecov-io
Copy link

codecov-io commented Feb 14, 2020

Codecov Report

Merging #419 into master will decrease coverage by 0.1%.
The diff coverage is 90%.

Impacted file tree graph

@@            Coverage Diff             @@
##           master     #419      +/-   ##
==========================================
- Coverage    88.3%   88.19%   -0.11%     
==========================================
  Files          42       41       -1     
  Lines        2060     2042      -18     
  Branches      236      234       -2     
==========================================
- Hits         1819     1801      -18     
  Misses        170      170              
  Partials       71       71
Impacted Files Coverage Δ
...elemetry-api/src/opentelemetry/context/__init__.py 67.39% <100%> (+3.1%) ⬆️
...i/src/opentelemetry/context/threadlocal_context.py 100% <100%> (ø)
...i/src/opentelemetry/context/contextvars_context.py 87.5% <66.66%> (ø)
...etry-sdk/src/opentelemetry/sdk/metrics/__init__.py 98.29% <0%> (-0.18%) ⬇️
...elemetry-api/src/opentelemetry/metrics/__init__.py 95.58% <0%> (ø) ⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 72862c9...57b214e. Read the comment docs.

Copy link
Member

@c24t c24t left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks great!

install_requires=["typing; python_version<'3.5'"],
install_requires=[
"typing; python_version<'3.5'",
"aiocontextvars; python_version<'3.7' and python_version>='3.5'",
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good call on the conditional requirements.

context.set_current(self.previous_context)

@patch(
"opentelemetry.context._RUNTIME_CONTEXT", ContextVarsRuntimeContext() # type: ignore
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure why the CLI black isn't wrapping these. FWIW black.format_file_contents via the vim plugin does.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It could be line length isn't being honored. I see it in the pyproject.toml, but IIRC my editor wasn't always honoring that configuration either.

@c24t
Copy link
Member

c24t commented Feb 15, 2020

@Oberon00 are you talking about moving the abstract RuntimeContext class into its own distribution or about moving the ContextVarsRuntimeContext and ThreadLocalRuntimeContext implementations or the three of them?

I'd assume all three. This is what we did with the opencensus-context package too.

@Oberon00
Copy link
Member

Oberon00 commented Feb 15, 2020

Exactly. It is illuminating to read OTEP 66, in which the context API was introduced, especially the motivation:

Separation of concerns

  • Cleaner package layout results in an easier to learn system. It is possible to
    understand Context Propagation without needing to understand Observability.
    [...]

Extensibility

  • A clean separation allows the context propagation mechanisms to be used on
    their own [...]

-- https://github.com/open-telemetry/oteps/blob/master/text/0066-separate-context-propagation.md

So one could install opentelemetry-context without even using telemetry, to e.g. transport identity information across contexts and threads. Basically that package would contain our re-implementation/abstraction of contextvars plus the inter-process propagation.


if (3, 5, 3) <= version_info < (3, 7):
import aiocontextvars # type: ignore # pylint:disable=unused-import
import aiocontextvars # type: ignore # pylint:disable=unused-import,import-error

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please note that without this patch in aiocontextvars fantix/aiocontextvars#66 behavior of async code in python<3.7 differs from original contextvars in python 3.7.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hello @condorcet

What specific behavior difference are you referring to?

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @ocelotl !
Even with aiocontextvars call_soon and similar methods of loop don't create their own context. Hence such scheduled callbacks are "stealing" outer context.
For example:

import asyncio
from contextvars import ContextVar

var = ContextVar('var')

def callback(name):
    print(f'{name} var before: {var.get()}')
    var.set(name)
    print(f'{name} var after: {var.get()}')

if __name__ == '__main__':
    var.set('root')
    loop = asyncio.get_event_loop()
    loop.call_soon(callback, 'callback_1')
    loop.call_soon(callback, 'callback_2')
    asyncio.get_event_loop().run_forever()

In python 3.7 (proper behavior):

callback_1 var before: root
callback_1 var after: callback_1
callback_2 var before: root
callback_2 var after: callback_2

In python 3.6 (with aiocontextvars):

callback_1 var before: root
callback_1 var after: callback_1
callback_2 var before: callback_1
callback_2 var after: callback_2

Here you can see that first callback changed context var, and next callback used modified value rather than initial.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for clarifying @condorcet, I'll open an issue to track this.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Opened #436

Copy link
Member

@toumorokoshi toumorokoshi left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nice!

@@ -84,9 +85,14 @@ def get_current() -> Context:
if _RUNTIME_CONTEXT is None:
# FIXME use a better implementation of a configuration manager to avoid having
# to get configuration values straight from environment variables
if version_info < (3, 5):
# contextvars are not supported in 3.4, use thread-local storage
default_context = "threadlocal_context"
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

does this mean that async context won't work for python3.4?

context.set_current(self.previous_context)

@patch(
"opentelemetry.context._RUNTIME_CONTEXT", ContextVarsRuntimeContext() # type: ignore
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It could be line length isn't being honored. I see it in the pyproject.toml, but IIRC my editor wasn't always honoring that configuration either.

…t.py

Co-Authored-By: Chris Kleinknecht <libc@google.com>
@toumorokoshi
Copy link
Member

There was a good conversation in the Gitter around breaking this out into it's own package now.

@c24t points that there will be complications around version mismatches.

E.g. today, if someone hard pins the API / SDK, they will also have to be aware that there is a context package that must also be hard pinned.

so consumers will probably hard pin API / SDK, pick up a backwards-incompatible context change (since they're unaware of it's existence or role), then have to go figure out what went wrong.

Probably good enough motivation to not split the context package out until it becomes more stable.

@toumorokoshi toumorokoshi merged commit d824f19 into open-telemetry:master Feb 18, 2020
@Oberon00
Copy link
Member

@toumorokoshi, @c24t: With the same argument, we shouldn't split the SDK from the API.

@toumorokoshi
Copy link
Member

Yes, I think ideally, we should have kept it together as long as we could have, until we had a strong need to (e.g. someone started implementing their own SDK). But doing it from the start I think isn't the end of the world, it also sets the precedence that both SDK and API packages should be called out specifically.

I could be convinced either way, hopefully context changes are few and far between after the first cycle or two. But I don't intend to be a voice of authority on this one, I'll defer to those who are more directly contributing, and also have a need for a separate context object.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

8 participants