Handle `import`ing relocated dispatch functions by jakirkham · Pull Request #623 · rapidsai/dask-cuda

jakirkham · 2021-05-24T23:48:07Z

As these functions recently got relocated, handle importing from the new location with a fallback to the old location.

xref: dask/dask#7503
xref: dask/dask#7505

dask_cuda/proxy_object.py

This allows us to smooth over differences between Dask before the dispatch refactor and after.

quasiben · 2021-05-25T00:20:29Z

Thanks @jakirkham for the work and @galipremsagar for the review!

quasiben · 2021-05-25T00:20:33Z

@gpucibot merge

jakirkham · 2021-05-25T03:01:19Z

Seeing this on CI

19:52:07 ________________________ test_dataframes_share_dev_mem _________________________
19:52:07 
19:52:07     def test_dataframes_share_dev_mem():
19:52:07         cudf = pytest.importorskip("cudf")
19:52:07     
19:52:07         df = cudf.DataFrame({"a": range(10)})
19:52:07 >       grouped = shuffle_group(df, "a", 0, 2, 2, False, 2)
19:52:07 
19:52:07 dask_cuda/tests/test_proxify_host_file.py:109: 
19:52:07 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
19:52:07 dask_cuda/proxify_device_objects.py:142: in wrapper
19:52:07     ret = func(*args, **kwargs)
19:52:07 /opt/conda/envs/rapids/lib/python3.7/site-packages/dask/dataframe/shuffle.py:816: in shuffle_group
19:52:07     ind = hash_object_dispatch(df[cols] if cols else df, index=False)
19:52:07 /opt/conda/envs/rapids/lib/python3.7/site-packages/dask/utils.py:511: in __call__
19:52:07     meth = self.dispatch(type(arg))
19:52:07 /opt/conda/envs/rapids/lib/python3.7/site-packages/dask/utils.py:499: in dispatch
19:52:07     return self.dispatch(cls)  # recurse
19:52:07 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
19:52:07 
19:52:07 self = <dask.utils.Dispatch object at 0x7fe9f6b57fd0>
19:52:07 cls = <class 'cudf.core.dataframe.DataFrame'>
19:52:07 
19:52:07     def dispatch(self, cls):
19:52:07         """Return the function implementation for the given ``cls``"""
19:52:07         # Fast path with direct lookup on cls
19:52:07         lk = self._lookup
19:52:07         try:
19:52:07             impl = lk[cls]
19:52:07         except KeyError:
19:52:07             pass
19:52:07         else:
19:52:07             return impl
19:52:07         # Is a lazy registration function present?
19:52:07         toplevel, _, _ = cls.__module__.partition(".")
19:52:07         try:
19:52:07             register = self._lazy.pop(toplevel)
19:52:07         except KeyError:
19:52:07             pass
19:52:07         else:
19:52:07             register()
19:52:07             return self.dispatch(cls)  # recurse
19:52:07         # Walk the MRO and cache the lookup result
19:52:07         for cls2 in inspect.getmro(cls)[1:]:
19:52:07             if cls2 in lk:
19:52:07                 lk[cls] = lk[cls2]
19:52:07                 return lk[cls2]
19:52:07 >       raise TypeError("No dispatch for {0}".format(cls))
19:52:07 E       TypeError: No dispatch for <class 'cudf.core.dataframe.DataFrame'>
19:52:07 
19:52:07 /opt/conda/envs/rapids/lib/python3.7/site-packages/dask/utils.py:505: TypeError

jakirkham · 2021-05-25T03:02:56Z

cc @madsbk

galipremsagar · 2021-05-25T03:07:01Z

This error is because hash_object_dispatch location has changed and dask-cudf code is not yet updated with correct location. dask-cudf will silently fail as the import is in try/catch and the dispatch is never registered for cudf.DataFrame. Fix for this is already included in : https://github.com/rapidsai/cudf/pull/8342/files#diff-8cccf94d2261588b89ab7cd02573abfe37a19aab98fb6d210184c5ce293e9b3eR249

galipremsagar · 2021-05-26T13:07:38Z

rerun tests

jakirkham · 2021-05-26T14:38:38Z

rerun tests

(looks like packages are up now)

kkraus14 · 2021-05-26T15:54:55Z

Looks like there's a failure: https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci-v0.20/job/dask-cuda/job/prb/job/dask-cuda-gpu-test/CUDA=11.2,GPU_LABEL=gpu,OS=centos7,PYTHON=3.8/104/testReport/junit/dask_cuda.tests/test_ucx_options/test_global_option/

jakirkham · 2021-05-26T15:58:10Z

Yeah saw that. Though it's not related to this change. Looks like some UCX configuration test failure. My inclination is to mark it as xfail so we can get this out and unblock people on CI who still need these changes. Thoughts?

jakirkham · 2021-05-26T16:03:35Z

dask_cuda/tests/test_ucx_options.py

+@pytest.mark.xfail(reason="https://github.com/rapidsai/dask-cuda/issues/627")
 def test_global_option():
    for seg_size in ["2K", "1M", "2M"]:
        p = mp.Process(target=_test_global_option, args=(seg_size,))


Ok marked this test with xfail so we can get these changes out-the-door to fix other CI issues. Though raised issue ( #627 ) to track/investigate

I think this is ok. @pentschev do you have any concerns? we are still in burn down so if something does crop up we have time to handle

For clarity cuDF's and RAFT's CI still fails without these changes, which is blocking a fair number of people. No objection to revisiting, but I do think we need to do this short term.

This is weird, it never failed before, including in my nightly runs with both UCX 1.9 and UCX master. I'm actually concerned that this may be some error from a previous test that pytest is spilling over. I opened a PR to trigger CI in #628 , let's hold on and see if that fails too before merging this.

Agree it is weird. I don't recall seeing this error on this PR before that build either.

Ok though JFYI before this came up, Ben had used a merge command upthread ( #623 (comment) ). So the bot may wind up merging this anyways if passes.

Haha, yeah, it was too quick, so no worries. I'm just now wondering if this isn't a bit dangerous, one could do any changes to the PR after someone told the bot to merge, and unless somebody sees it in time, it would be merged.

Yeah it probably should be associated with a commit hash

Or just invalidated based on timestamp. If commit timestamp > gpucibot merge comment, invalidate automerge.

I mean the time the commit was pushed, not the local commit timestamp.

Submitted PR ( #630 ) to revert this piece. Probably still fails, but at least that gives us something to work from

codecov-commenter · 2021-05-26T16:46:50Z

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

Attention: Patch coverage is 50.00000% with 2 lines in your changes missing coverage. Please review.

Please upload report for BASE (branch-21.06@ab1d35c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
dask_cuda/proxy_object.py	50.00%	2 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@               Coverage Diff               @@
##             branch-21.06     #623   +/-   ##
===============================================
  Coverage                ?   90.56%           
===============================================
  Files                   ?       15           
  Lines                   ?     1622           
  Branches                ?        0           
===============================================
  Hits                    ?     1469           
  Misses                  ?      153           
  Partials                ?        0

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

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

jakirkham requested a review from a team as a code owner May 24, 2021 23:48

github-actions bot added the python python code needed label May 24, 2021

jakirkham added bug Something isn't working non-breaking Non-breaking change labels May 24, 2021

Handle importing relocated dispatch functions

f168cc1

jakirkham force-pushed the fix_dask_dispatch branch from d31caa5 to f168cc1 Compare May 24, 2021 23:52

galipremsagar reviewed May 24, 2021

View reviewed changes

dask_cuda/proxy_object.py Outdated Show resolved Hide resolved

Update dispatch function modules

132ea66

This allows us to smooth over differences between Dask before the dispatch refactor and after.

galipremsagar approved these changes May 25, 2021

View reviewed changes

quasiben approved these changes May 25, 2021

View reviewed changes

Handle the movement of concat_pandas

74768b8

This was referenced May 25, 2021

Fix meta creation incase of object dask/dask#7586

Merged

Adding profiling to dask shuffle #625

Merged

[REVIEW] Add support for make_meta_obj dispatch in dask-cudf rapidsai/cudf#8342

Merged

make_meta is in utils

47a75c1

jakirkham mentioned this pull request May 26, 2021

test_global_option fails #627

Closed

Mark test_global_option xfail

6e32eca

jakirkham commented May 26, 2021

View reviewed changes

rapids-bot bot merged commit 1ef38b4 into rapidsai:branch-21.06 May 26, 2021

jakirkham deleted the fix_dask_dispatch branch May 26, 2021 17:13

jakirkham mentioned this pull request May 26, 2021

Fix import path changes in Dask #626

Closed

Conversation

jakirkham commented May 24, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

quasiben commented May 25, 2021

Uh oh!

quasiben commented May 25, 2021

Uh oh!

jakirkham commented May 25, 2021

Uh oh!

jakirkham commented May 25, 2021

Uh oh!

galipremsagar commented May 25, 2021

Uh oh!

galipremsagar commented May 26, 2021

Uh oh!

jakirkham commented May 26, 2021

Uh oh!

kkraus14 commented May 26, 2021

Uh oh!

jakirkham commented May 26, 2021

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented May 26, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

jakirkham commented May 24, 2021 •

edited

Loading

codecov-commenter commented May 26, 2021 •

edited

Loading