Continue to support Kernel._parent_ident for backward compatibility
#1427
Conversation
|
I have confirmed locally on Linux that this fixes the problematic JupyterLab tests, which are: $ cd packages/services
$ jlpm test test/kernel/ikernel.spec.ts
<snip>
Test Suites: 1 passed, 1 total
Tests: 71 passed, 71 total
Snapshots: 0 total
Time: 118.164 s
Ran all test suites matching /test\/kernel\/ikernel.spec.ts/i.
|
| @@ -313,6 +318,15 @@ def __init__(self, **kwargs): | |||
| self._shell_parent_ident = ContextVar("shell_parent_ident") | |||
| self._shell_parent_ident.set(b"") | |||
|
|
|||
| # For backward compatibility so that _parent_ident["shell"] and _parent_ident["control"] | |||
| # work as they used to for ipykernel >= 7 | |||
| self._parent_ident = LazyDict( | |||
There was a problem hiding this comment.
Do we want to make this a property with a deprecation warning pointing to a stable public API to use instead?
There was a problem hiding this comment.
We could but as far as I can tell there isn't currently a public API for this and I don't personally want to extend the public API.
I think that ipykernel 8 is likely to support concurrent execution of async code cells (via anyio or not), and the timeline for version 7 to 8 will be much faster than for 6 to 7. For that the concepts of a current parent header and ident won't exist so we'd either have to remove the caching of such information and pass it around as arguments, or at least make the caching more complex such than any public API we could invent now would have to be changed again. So allowing downstream libraries to continue to use _parent_ident for the 7.x series seems sensible.
| @@ -146,6 +146,12 @@ Submodules | |||
| :show-inheritance: | |||
|
|
|||
|
|
|||
| .. automodule:: ipykernel.utils | |||
There was a problem hiding this comment.
do we want LazyDict to be a publicly documented part of the API? I would think we'd want it to be a private implementation detail.
PR #1414 replaced the private member variable
Kernel._parent_identwith threadsafe support of separate parent idents per thread (subshell) usingContextVars. This caused a failure downstream in JupyterLab (jupyterlab/jupyterlab#17785) as it reads the private_parent_ident. Upon investigation there are a number of downstream projects that read this variable, presumably due to copy-paste-modify ofipykernelcode to override the behaviour. So rather than break those projects, here I am reintroducingKernel._parent_identas a shim for backward compatibility that can be called the same way as before. It uses a lazy-evaluated read-only dictionary.I have also added a new test to check that
Kernel.get_parentandKernel._parent_identperform as expected.