FSStore: use `ensure_bytes()` #1285

madsbk · 2022-12-06T11:48:47Z

TODO:

Add unit tests and/or doctests in docstrings
Add docstrings and API docs for any new/modified user-facing classes and functions
Changes documented in docs/release.rst
GitHub Actions have all passed
Test coverage is 100% (Codecov passes)

codecov · 2022-12-06T12:03:53Z

Codecov Report

Merging #1285 (cfab20c) into main (385b5d3) will not change coverage.
The diff coverage is 100.00%.

@@            Coverage Diff            @@
##              main     #1285   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           35        35           
  Lines        14388     14404   +16     
=========================================
+ Hits         14388     14404   +16

Impacted Files	Coverage Δ
zarr/storage.py	`100.00% <100.00%> (ø)`
zarr/tests/test_core.py	`100.00% <100.00%> (ø)`
zarr/util.py	`100.00% <100.00%> (ø)`

madsbk · 2022-12-06T12:58:05Z

@normanrz can you check if this fixes the issue?

normanrz · 2022-12-06T13:02:55Z

Yes! That looks good to me. Thanks for the quick fix!

rabernat · 2022-12-06T13:12:13Z

Thanks for this PR @madsbk! I recently hit a similar issue involving zero-dimensional arrays (which have compression turned off by default).

I read the discussion in #1279, and I am left wondering why we leave it up to the store to call ensure_bytes. The abstraction in Zarr is very clear: the store is a key-value store where the keys are strings and the values are bytes. Therefore, IMO, it should be up to zarr.core to ensure that the data it is sending to the store are always bytes, not the store implementation.

What do others think about this?

madsbk · 2022-12-06T13:27:30Z

The abstraction in Zarr is very clear: the store is a key-value store where the keys are strings and the values are bytes.

We changed this with #934, now we also allow NDArrayLike instances. The motivation is to support memory objects such as device memory. E.g., KvikIO implements a GDSStore that use GDS to speedup IO between device and host memory.

Having said that, I was surprised that fsspec.FSMap didn't support F-ordered numpy arrays. It might be possible to fix the issue in fsspec.FSMap as well.

rabernat · 2022-12-06T13:58:32Z

We changed this with #934, now we also allow NDArrayLike instances.

I see that this has been implemented in the zarr-python implementation. And I thank everyone who worked on this amazing feature! 🚀 But that doesn't mean it is consistent with what the spec says:

The store interface defines a set of operations involving keys and values. In the context of this interface, a key is any string... A value is a sequence of bytes.

It is certainly straightforward to convert an uncompressed ndarray into a sequence of bytes. So maybe I am overthinking this... Or maybe the spec needs to be updated to reflect these new implementation changes? 🤔 My only goal here is to make sure that the spec and the implementation are consistent with each other. Otherwise we risk interoperability with other zarr implementations.

If we are fine with saying "a value is a sequence of bytes or a raw uncompressed array" then I am fine with leaving it up to the store implementation whether to convert arrays to bytes.

madsbk · 2022-12-06T14:21:07Z

Good point and I agree, the spec and implementation should match up!

I must admit I am reading "A value is a sequence of bytes" in a more general sense, not as a reference to the bytes class specifically, which will also exclude types such as bytearray or memoryview.

jakirkham · 2022-12-06T19:14:17Z

Hey Mads, thanks for following up on this issue.

That said, don't think we should be making this change. In general it is already the expectation in zarr-python & Numcodecs that Python Buffer Protocol supporting objects are used (memoryviews, ndarrays, bytearrays, bytes, etc.). Part of the reason for this is it allows zero-copy operations on the data (like casting, reshaping, etc.). It also can support things like memory-mapping natively (as mmap supports the Python Buffer Protocol).

How to word this in the spec is nuanced. The spec itself is not meant to be Python specific, but to support other languages as well. They may not have bytearray or memoryview for example or if they do it may be named differently and/or behave differently. So think we should consider language agnostic wording. It sounds like this is how you understood it as well.

Perhaps others who found this wording in the spec confusing can chime in on how to clarify it. Maybe we can say "a value contains binary data"?

madsbk · 2022-12-07T07:22:00Z

@jakirkham, I am not sure I understand. Are you saying that we shouldn't use ensure_bytes() in FSStore (this PR) or that the spec should stay language agnostic?

I have no strong opinion about the specifications, but I think it is important that we allow NDArrayLike objects in values. If we move ensure_bytes() to the zarr.core level, we are effectively saying that a value is always of type bytes.

jakirkham · 2022-12-07T10:13:02Z

Could we call ensure_ndarray_like instead?

normanrz · 2022-12-07T10:23:56Z

Could we call ensure_ndarray_like instead?

I'm not sure that would work. I think the trick is that at some point, tobytes needs to be called with value.tobytes(order='A') to preserve the order. https://numpy.org/doc/stable/reference/generated/numpy.ndarray.tobytes.html

jakirkham · 2022-12-08T10:58:03Z

There's ensure_contiguous_ndarray_like, which can flatten the array with respect to order

jstriebel · 2022-12-12T14:14:23Z

Since most other stores use ensure_bytes atm, I think this fix is valid for the moment, and important to fix an existing bug. I agree, that there should be a discussion whether every store should do this or not, but this should not block this PR IMO.

rabernat

I am fine with this fix for now, since it seems consistent with what most other stores do.

We can revisit the question of whether this ensure_bytes logic should live elsewhere in the future.

jakirkham · 2022-12-13T09:43:56Z

Sorry to be a stickler. The main reason I bring up ensure_bytes here is this typically incurs the cost of a copy (with additional memory being used). Given this is happening with multiple memory segments that are being written together, this could result in more significant memory usage. Especially as we start to rely more on batching multiple reads & writes. While it may not always be possible to avoid these copies, they could be deferred until absolutely necessary (with the resulting bytes objects quickly discarded) to keep memory usage as low as possible (since not all memory segments would needing copying at once). Hope that clarifies the concern a bit more.

zarr/storage.py

normanrz · 2022-12-19T10:17:58Z

Any chance we could get this merged and released?

rabernat · 2022-12-19T11:44:27Z

I'd like to hear from @jakirkham - John, would you be you okay with addressing ensure_contiguous_ndarray_like vs. ensure_bytes in a follow-on PR (where we can also address it for all the other stores that call ensure_bytes)? Or do you think this needs to be resolved before we merge?

normanrz · 2022-12-19T11:53:16Z

I think @jakirkham's latest suggestion is a straightforward compromise that could be incorporated as part of this PR.

…Store_ensure_bytes

madsbk · 2022-12-19T12:46:27Z

I think @jakirkham's latest suggestion is a straightforward compromise that could be incorporated as part of this PR.

Done

joshmoore · 2022-12-21T09:08:17Z

Re-opening to try to fix the codecov build.

joshmoore · 2022-12-21T10:28:15Z

Now green. @jakirkham: does this match your expectations?

If so, happy to get it in and prep a release during the holidays. (Help with the release notes will be appreciated!)

rabernat · 2022-12-21T14:40:12Z

If we're doing a release, let's get #1304 in too.

joshmoore · 2023-01-16T14:58:24Z

If we're doing a release, let's get #1304 in too.

see #1309 (comment)

Leaving this PR for @jakirkham to review.

jakirkham

Thanks Mads! 🙏

LGTM. Made a few comments that we can follow up on separately

jakirkham · 2023-01-19T18:50:30Z

zarr/util.py

@@ -696,3 +701,28 @@ def all_equal(value: Any, array: Any):
            # using == raises warnings from numpy deprecated pattern, but
            # using np.equal() raises type errors for structured dtypes...
            return np.all(value == array)
+
+
+def ensure_contiguous_ndarray_or_bytes(buf) -> Union[NDArrayLike, bytes]:


We may want to move this to numcodecs.compat, but that can be follow-on work

jakirkham · 2023-01-19T18:55:04Z

zarr/storage.py

+        values = {
+            self._normalize_key(key): ensure_contiguous_ndarray_or_bytes(val)
+            for key, val in values.items()
+        }


Something we may consider is delaying the instantiation of the dict so these copies occur as these values are requested. This can be a follow-on though

joshmoore · 2023-01-19T19:04:11Z

@jakirkham: ok to get this in the 2.13.x series?

@madsbk: I'll get this out ASAP but might wait on any emergency needs for re: #1324

Suggestions for the docs/release.rst statement would be welcome. 🙏🏽

jakirkham · 2023-01-19T19:30:55Z

Suggestions for the docs/release.rst statement would be welcome. 🙏🏽

Oops sorry should have asked for that here (or just added it). Submitted PR ( #1325 )

setitem: use ensure_bytes

3213876

github-actions bot added the needs release notes Automatically applied to PRs which haven't added release notes label Dec 6, 2022

madsbk changed the title ~~setitem: use ensure_bytes~~ FSStore: use ensure_bytes() Dec 6, 2022

test

4811181

madsbk marked this pull request as ready for review December 6, 2022 13:31

rabernat approved these changes Dec 12, 2022

View reviewed changes

jakirkham reviewed Dec 14, 2022

View reviewed changes

zarr/storage.py Outdated Show resolved Hide resolved

madsbk added 2 commits December 19, 2022 13:42

impl. and use ensure_contiguous_ndarray_or_bytes()

5caca83

Merge branch 'main' of github.com:zarr-developers/zarr-python into FS…

f983814

…Store_ensure_bytes

joshmoore closed this Dec 21, 2022

joshmoore reopened this Dec 21, 2022

MSanKeys963 mentioned this pull request Jan 14, 2023

2.13.4 release notes #1316

Merged

madsbk added 2 commits January 16, 2023 17:02

Merge branch 'main' into FSStore_ensure_bytes

3afda2b

Merge branch 'main' into FSStore_ensure_bytes

cfab20c

jakirkham approved these changes Jan 19, 2023

View reviewed changes

jakirkham merged commit b9e9f5a into zarr-developers:main Jan 19, 2023

jakirkham mentioned this pull request Jan 19, 2023

Add FSStore contiguous bug fix note #1325

Merged

6 tasks

madsbk deleted the FSStore_ensure_bytes branch January 20, 2023 09:55

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FSStore: use `ensure_bytes()` #1285

FSStore: use `ensure_bytes()` #1285

madsbk commented Dec 6, 2022 •

edited

Loading

codecov bot commented Dec 6, 2022 •

edited

Loading

madsbk commented Dec 6, 2022

normanrz commented Dec 6, 2022

rabernat commented Dec 6, 2022

madsbk commented Dec 6, 2022 •

edited

Loading

rabernat commented Dec 6, 2022 •

edited

Loading

madsbk commented Dec 6, 2022 •

edited

Loading

jakirkham commented Dec 6, 2022

madsbk commented Dec 7, 2022

jakirkham commented Dec 7, 2022

normanrz commented Dec 7, 2022

jakirkham commented Dec 8, 2022 •

edited

Loading

jstriebel commented Dec 12, 2022

rabernat left a comment

jakirkham commented Dec 13, 2022

normanrz commented Dec 19, 2022

rabernat commented Dec 19, 2022

normanrz commented Dec 19, 2022 •

edited

Loading

madsbk commented Dec 19, 2022

joshmoore commented Dec 21, 2022

joshmoore commented Dec 21, 2022

rabernat commented Dec 21, 2022

joshmoore commented Jan 16, 2023

jakirkham left a comment

jakirkham Jan 19, 2023

jakirkham Jan 19, 2023

joshmoore commented Jan 19, 2023

jakirkham commented Jan 19, 2023

FSStore: use ensure_bytes() #1285

FSStore: use ensure_bytes() #1285

Conversation

madsbk commented Dec 6, 2022 • edited Loading

codecov bot commented Dec 6, 2022 • edited Loading

Codecov Report

madsbk commented Dec 6, 2022

normanrz commented Dec 6, 2022

rabernat commented Dec 6, 2022

madsbk commented Dec 6, 2022 • edited Loading

rabernat commented Dec 6, 2022 • edited Loading

madsbk commented Dec 6, 2022 • edited Loading

jakirkham commented Dec 6, 2022

madsbk commented Dec 7, 2022

jakirkham commented Dec 7, 2022

normanrz commented Dec 7, 2022

jakirkham commented Dec 8, 2022 • edited Loading

jstriebel commented Dec 12, 2022

rabernat left a comment

Choose a reason for hiding this comment

jakirkham commented Dec 13, 2022

normanrz commented Dec 19, 2022

rabernat commented Dec 19, 2022

normanrz commented Dec 19, 2022 • edited Loading

madsbk commented Dec 19, 2022

joshmoore commented Dec 21, 2022

joshmoore commented Dec 21, 2022

rabernat commented Dec 21, 2022

joshmoore commented Jan 16, 2023

jakirkham left a comment

Choose a reason for hiding this comment

jakirkham Jan 19, 2023

Choose a reason for hiding this comment

jakirkham Jan 19, 2023

Choose a reason for hiding this comment

joshmoore commented Jan 19, 2023

jakirkham commented Jan 19, 2023

FSStore: use `ensure_bytes()` #1285

FSStore: use `ensure_bytes()` #1285

madsbk commented Dec 6, 2022 •

edited

Loading

codecov bot commented Dec 6, 2022 •

edited

Loading

madsbk commented Dec 6, 2022 •

edited

Loading

rabernat commented Dec 6, 2022 •

edited

Loading

madsbk commented Dec 6, 2022 •

edited

Loading

jakirkham commented Dec 8, 2022 •

edited

Loading

normanrz commented Dec 19, 2022 •

edited

Loading