fix DinD host detection when using local socket

[ikravets]

When using IPC socket (unix/npipe) to communicate with local docker server docker-py implementation rewrites the base url to HTTP, see https://github.com/docker/docker-py/blob/e901eac7a8c5f29c7720eafb9f58c8356cca2324/docker/api/client.py#L143-L168

We cannot rely on URL scheme in order to detect such connections. Instead, we detect connection adapters added by docker-py api client:

• UnixHTTPAdapter https://github.com/docker/docker-py/blob/e901eac7a8c5f29c7720eafb9f58c8356cca2324/docker/transport/unixconn.py

• NpipeHTTPAdapter https://github.com/docker/docker-py/blob/e901eac7a8c5f29c7720eafb9f58c8356cca2324/docker/transport/npipeconn.py

As NpipeHTTPAdapter type may not be available outside of Windows we rely on adapter-specific attributes.

[codecov-commenter]

Codecov Report

❗ No coverage uploaded for pull request base (main@928af5a). Click here to learn what that means.
The diff coverage is n/a.

❗ Current head afa057e differs from pull request most recent head f9d4508. Consider uploading reports for the commit f9d4508 to get more accurate results

@@           Coverage Diff           @@
##             main     #283   +/-   ##
=======================================
  Coverage        ?   87.30%           
=======================================
  Files           ?       30           
  Lines           ?      843           
  Branches        ?       57           
=======================================
  Hits            ?      736           
  Misses          ?       76           
  Partials        ?       31           

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[ikravets]

Rebased to the master. @tillahoffmann, any chance for a review?

[ikravets]

rebased to master

[tillahoffmann]

Thank you for these improvements! Are there tests that we could add that fail before the change is introduced to avoid future regressions?

[tillahoffmann]

This is great, thank you for having such a detailed look into improving the host detection. Host detection is the challenge that seems to cause most trouble for our users so far.

Do you know whether this change would also resolve the issue of localnpipe appearing in the connection string, as discussed in #108? You're much more clued up on this than I am.

[tillahoffmann]

Even if the NpipeHTTPAdapter is only used on windows, would it still be available for an isinstance check independent of the platform? Or does the docker package not even expose NpipeHTTPAdapter if it's not on windows?

Alternatively, would checking adapter.__class__.name == 'NpipeHTTPAdapter' be easier to understand than checking the attributes of the adapter?

[ikravets]

I don't think that NpipeHTTPAdapter symbol is available outside of windows. Checking the class name would break for NpipeHTTPAdapter subclasses. So I went on with a classic duck-typing.

[CarliJoy]

I think something like this:

try:
    from docker.transport.npipeconn import NpipeHTTPAdapter
except ImportError:
    class NpipeHTTPAdapter: ...

Would be a more proper solution.

It is much clearer what happens, does not depend the design of the Adapter and respects subclasses.

[ikravets]

@tillahoffmann Hi, it's been more than half a year since I've opened this PR. Can we maybe merge this?

[ikravets]

@tillahoffmann a kind ping

[CarliJoy]

This should be obsolete now that #714 is merged.
Could you check?
If you think the detection is still required maybe add a is_socket_connection method or property and include the results of the check to either host or get_connection_mode of the DockerClient.