Regression in use with dask-awkward #155
Comments
|
Dmitry Kalinkin ***@***.***> writes:
```python
import boost_histogram as bh
import dask_awkward as dak
import dask_histogram as dh
arr = dak.from_lists([list(range(10))] * 3)
axis = bh.axis.Regular(10, 0., 10.)
hist = dh.factory(
arr,
axes=(axis,),
).compute()
print(hist.values())
```
Works on dask 2024.11.2, distributed 2024.11.2 and dask-awkward 2024.9.0:
```output
[3. 3. 3. 3. 3. 3. 3. 3. 3. 3.]
```
Broken with dask 2024.12.0, distributed 2024.12.0 and dask-awkward 2024.12.1:
```python
Traceback (most recent call last):
File "/tmp/regr/regr.py", line 10, in <module>
).compute()
^^^^^^^^^
File "/nix/store/nbqgvfbciwpxdazx5zn72i31wckkp5ad-python3.12-dask-2024.12.0/lib/python3.12/site-packages/dask/base.py", line 372, in compute
(result,) = compute(self, traverse=False, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/nbqgvfbciwpxdazx5zn72i31wckkp5ad-python3.12-dask-2024.12.0/lib/python3.12/site-packages/dask/base.py", line 660, in compute
results = schedule(dsk, keys, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/9x4i6b87rx9w6y26732msn50rnn6fbr8-python3.12-dask-histogram-2024.9.1/lib/python3.12/site-packages/dask_histogram/core.py", line 290, in _blocked_dak
return thehist.fill(thedata, weight=theweights, sample=thesample)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/3h45q4z2cg38204ja42696kbik3j2wxb-python3.12-boost-histogram-1.5.0/lib/python3.12/site-packages/boost_histogram/_internal/hist.py", line 511, in fill
self._hist.fill(*args_ars, weight=weight_ars, sample=sample_ars) # type: ignore[arg-type]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: Wrong number of args
```
Likely due to blockwise changes that just made it into dask-awkward
(which were made to be compatible with the new Task spec in upstream
Dask). Blockwise is responsible for unpacking arguments, and it looks
like `args_ars` in that call to a boost-histogram function is
incorrectly packed by the blockwise layer in the graph.
The `_partitionwise` function in core.py is probably the place that
needs patching (that's where dask.blockwise is used).
|
|
Unfortunately only interactions with delayed objects seemed to need any wrapping with TaskRef or the like... I'll take a further look into it tomorrow. |
|
@veprbl just released dask-awkward 2024.12.0, should be good to go now. Thanks @douglasdavis! |
|
I can confirm that this solves the example above, but there are more issues. Consider: import boost_histogram as bh
import dask_awkward as dak
import dask_histogram as dh
arr = dak.from_lists([list(range(10))] * 3)
axis = bh.axis.Regular(10, 0., 10.)
hist = dh.factory(
arr,
axes=(axis,),
weights=arr,
).compute()
print(hist.values())Should give: but instead gives: Traceback (most recent call last):
File "this.py", line 11, in <module>
).compute()
^^^^^^^^^
File "/nix/store/nbqgvfbciwpxdazx5zn72i31wckkp5ad-python3.12-dask-2024.12.0/lib/python3.12/site-packages/dask/base.py", line 372, in compute
(result,) = compute(self, traverse=False, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/nbqgvfbciwpxdazx5zn72i31wckkp5ad-python3.12-dask-2024.12.0/lib/python3.12/site-packages/dask/base.py", line 660, in compute
results = schedule(dsk, keys, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/qm9y3w8x8j4zfyh4jh5843bdnv0x901x-python3.12-dask-histogram-2024.12.0/lib/python3.12/site-packages/dask_histogram/core.py", line 290, in _blocked_dak
return thehist.fill(thedata, weight=theweights, sample=thesample)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nix/store/3h45q4z2cg38204ja42696kbik3j2wxb-python3.12-boost-histogram-1.5.0/lib/python3.12/site-packages/boost_histogram/_internal/hist.py", line 504, in fill
weight_ars = _fill_cast(weight)
^^^^^^^^^^^^^^^^^^
File "/nix/store/3h45q4z2cg38204ja42696kbik3j2wxb-python3.12-boost-histogram-1.5.0/lib/python3.12/site-packages/boost_histogram/_internal/hist.py", line 82, in _fill_cast
return np.asarray(value)
^^^^^^^^^^^^^^^^^
File "/nix/store/6r7ij4cd6fgzmpcs25iw4p7p18z3c2rs-python3.12-dask-awkward-2024.12.1/lib/python3.12/site-packages/dask_awkward/lib/core.py", line 1718, in __array__
raise NotImplementedError
NotImplementedError |
|
Yeah, the approach using a partial doesn't cover all cases, just |
|
Ah that’s right, the if statements need to be there because the collections need to still resolve to collections (the partial kills the collection discovery in dask-awkward), this is why I originally wrote all the specific function implementations to begin with, it’s an easy fix but yields a decent amount of code duplication. I’ll write up the fix later today.
…On Tue, Dec 17, 2024, at 13:18, Lindsey Gray wrote:
Yeah, the approach using a partial doesn't cover all cases, just `None`s.
—
Reply to this email directly, view it on GitHub
<#155 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAYNYKUEQQWAX3726OL6KZT2GB2GRAVCNFSM6AAAAABTXLE2L6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNBZGQYDEMRVGY>.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
So the real issue is that |
|
And it's specifically happening during the optimization pass. (i.e. these |
|
Yeah, blockwise optimization has gotten more liberal with culling (most things that don't have a TaskRef will get removed in optimization!!) and we have to be much more careful what's actually exposed in the graph. I think, however, it's a good thing in the long run. culling especially touchy because of things like: https://github.com/dask/dask/blob/main/dask/blockwise.py#L675-L684 |
|
Closing again with #157 in |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Works on dask 2024.11.2, distributed 2024.11.2 and dask-awkward 2024.9.0:
Broken with dask 2024.12.0, distributed 2024.12.0 and dask-awkward 2024.12.1:
The text was updated successfully, but these errors were encountered: