ENH allow caching coroutine functions

[gsakkis]

Addresses #889, allowing coroutine functions (available in Python 3.5+) to be decorated with joblib.memory. In short, the decorated function's __call__ and call_and_shelve are coroutine functions that when awaited return the output or MemorizedResult, respectively, that would normally return if the original function was not a coroutine. A small subset of tests were ported from test_memory to their asyncio equivalent to verify the expected behavior.

In order to keep the asyncio-specific implementation as small as possible, I had to refactor a large portion of memory.py first. It's a separate commit so it can be reviewed and tested independently.

[codecov]

Codecov Report

Merging #894 into master will decrease coverage by 3.06%.
The diff coverage is 98.64%.

@@            Coverage Diff             @@
##           master     #894      +/-   ##
==========================================
- Coverage   95.15%   92.09%   -3.07%     
==========================================
  Files          45       47       +2     
  Lines        6440     6549     +109     
==========================================
- Hits         6128     6031      -97     
- Misses        312      518     +206

Impacted Files	Coverage Δ
joblib/test/test_memory.py	98.01% <100%> (-0.16%)	⬇️
joblib/_memory_async.py	100% <100%> (ø)
joblib/test/test_memory_async.py	100% <100%> (ø)
joblib/_store_backends.py	90.5% <90.9%> (-0.03%)	⬇️
joblib/memory.py	95.46% <97.64%> (-0.24%)	⬇️
joblib/_memory_helpers.py	3.07% <0%> (-70.77%)	⬇️
joblib/backports.py	33.33% <0%> (-60.42%)	⬇️
joblib/_multiprocessing_helpers.py	69.23% <0%> (-11.54%)	⬇️
joblib/format_stack.py	71.77% <0%> (-11.49%)	⬇️
joblib/my_exceptions.py	88.67% <0%> (-9.44%)	⬇️
... and 25 more

---

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 0957e31...c4e8948. Read the comment docs.

[codecov]

Codecov Report

Attention: Patch coverage is 97.09544% with 7 lines in your changes are missing coverage. Please review.

❗ No coverage uploaded for pull request base (main@c2087db). Click here to learn what that means.

Files	Patch %	Lines
joblib/memory.py	95.45%	5 Missing ⚠️
joblib/_store_backends.py	93.75%	2 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #894   +/-   ##
=======================================
  Coverage        ?   95.21%           
=======================================
  Files           ?       45           
  Lines           ?     7672           
  Branches        ?        0           
=======================================
  Hits            ?     7305           
  Misses          ?      367           
  Partials        ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[gsakkis]

Any feedback on this? Btw as commented in another PR, the AppVeyor failure looks unrelated.

[pierreglaser]

Thanks for putting this together. I can make a pass on it if @tomMoral or @ogrisel are on board with this PR.

[WestXu]

It's been about a year since last update on this merge. Is this repo abandoned somehow? What's the substitute library guys plz?

[ckchow]

joblib.memory is great but it is painful that it doesn't work quite right with async.

[davidbrochart]

I might be interested in using joblib.memory to automatically cache the execution of Jupyter notebook cells in akernel. It already supports reactive programming, but that doesn't play well with big data. Caching cell execution would allow to "restart and run all" basically for free.
akernel does some AST transformation on the executed code, and already wraps the cell code into an async function. It could use joblib.memory if coroutines were supported.
Any plan on continuing that effort?

[leshchenko1979]

You can try using perscache (https://github.com/leshchenko1979/perscache) - it supports persistent caching for async functions.

[rundef]

I wish this was part of the joblib's package !

[gsakkis]

I'm not using joblib for some years now but I could try rebasing this PR if there's interest in merging it, otherwise just lmk and I'll close it.
cc @ogrisel

[tomMoral]

I think this could be a nice addition, if you are up to rebasing the PR, i can take a look to help merge it.

[gsakkis]

@tomMoral i rebased it, please take a look. Not sure what's the issue with the pypy timeout, the related tests are passing.

[tomMoral]

Thanks a lot for rebasing and sorry for the delay in answering, I will take a look at this PR.

I could not look at the pypy timeout but most probable is that it is due to a random failure of the pipeline on pypy that is unrelated to your changes.
I relaunched the tests to see the status here.

[tomMoral]

This looks quite good!

I did a change on the name path -> call_id, did not expect this change to be in so many places 😅
Some cosmetic changes but overall I think this PR is very nice, thanks for the refactoring which makes the code much easier to read.

[gsakkis]

@tomMoral thanks for the review! I noticed in some places you rename path to cache_id and others to call_id; is this intentional?

[tomMoral]

@tomMoral thanks for the review! I noticed in some places you rename path to cache_id and others to call_id; is this intentional?

no, I think call_id was my goal 😁

[tomMoral]

The failure is due to outdated install of sklearn's nightly build, which is fixed in #1567
I will merge this one once the CI is green.

[tomMoral]

Thanks a lot @gsakkis !! That's a nice feature!

Sorry for the (very) slow process and thanks for keeping up with it 🙏

[gsakkis]

Better late than never, many thanks @tomMoral!