[dask] speed up tests

[jmoralez]

This aims to reduce the runtime of the dask tests. Following #6816 (comment), the first step was to replace the client fixture with one that reuses the same cluster and just creates new clients in every test, and makes the least changes to the existing code.

There are some other tests that are building clusters instead of using the client fixture that could be benefited by this, however adding this only reduced the runtime by about a minute, so I ran pytest with --durations=0 and got this:

=============================================================== slowest durations ================================================================
155.91s call     python/test_with_dask.py::TestWithDask::test_approx
138.92s call     python/test_with_dask.py::TestWithDask::test_hist
22.91s call     python/test_with_dask.py::test_from_dask_array
9.20s call     python/test_with_dask.py::TestWithDask::test_feature_weights
7.02s call     python/test_with_dask.py::test_dask_ranking
6.61s call     python/test_with_dask.py::test_with_asyncio
6.44s call     python/test_with_dask.py::test_parallel_submit_multi_clients

Will investigate the ones that take the most time.

[codecov-commenter]

Codecov Report

Merging #7020 (f159c58) into master (655e699) will increase coverage by 0.15%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #7020      +/-   ##
==========================================
+ Coverage   81.71%   81.86%   +0.15%     
==========================================
  Files          13       13              
  Lines        3916     3916              
==========================================
+ Hits         3200     3206       +6     
+ Misses        716      710       -6     

Impacted Files	Coverage Δ
python-package/xgboost/core.py	82.96% <0.00%> (+0.10%)	⬆️
python-package/xgboost/dask.py	82.02% <0.00%> (+0.66%)	⬆️

---

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 655e699...f159c58. Read the comment docs.

[trivialfis]

The slow test is probably not caused by dask but by hypothesis.

[jmoralez]

I see. The total runtime for the tests in this PR was 14 minutes and the current master is around 17 minutes. Should I add the client fixture to tests that currently build a cluster internally like these?

xgboost/tests/python/test_with_dask.py

Line 67 in 7beb2f7

def test_from_dask_dataframe() -> None:

xgboost/tests/python/test_with_dask.py

Line 110 in 7beb2f7

def test_from_dask_array() -> None:

That could probably reduce the runtime by a couple more minutes. Or if you have any other suggestions for maybe improving the hypothesis ones I could look into it.

[trivialfis]

hould I add the client fixture to tests that currently build a cluster internally like these?

Some tests use specific number of workers so that they have to define their own cluster.

Or if you have any other suggestions for maybe improving the hypothesis ones I could look into it.

Sorry I don't have any suggestion, you know these better than me. ;-)

[jmoralez]

I ran the tests that take the most time and pretty much all the time is spent training so I don't think there's something that could be improved there. There are a couple of clusters that get created with kWorkers and no threads_per_worker, I could create an additional fixture like kCluster and pass that to those tests so that they reuse that one.

This didn't improve as much as I hoped haha so please let me know if its useful at all.

[trivialfis]

I ran the tests that take the most time and pretty much all the time is spent training so I don't think there's something that could be improved there.

That's fine, the hypothesis tests take longer than we would like but highly effective at catching bugs.

There are a couple of clusters that get created with kWorkers and no threads_per_worker

Thank you for looking into them!

could create an additional fixture like kCluster and pass that to those tests so that they reuse that one.

I will follow up on making those changes since I wrote most of the tests, I should cleanup my own mess.

This didn't improve as much as I hoped haha so please let me know if its useful at all.

Of course it's useful and thank you! After merging the PR we know that we should focus on other places.

[jmoralez]

@trivialfis I think this is ready for review, looking forward to your thoughts.