-
-
Notifications
You must be signed in to change notification settings - Fork 18.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
PDEP-15: Reject PDEP-10 #58623
base: main
Are you sure you want to change the base?
PDEP-15: Reject PDEP-10 #58623
Conversation
/preview |
Website preview of this PR available at: https://pandas.pydata.org/preview/pandas-dev/pandas/58623/ |
Ugh, looks like the bullets aren't rendering correctly. |
Did I miss an official vote on rejecting this? I am not sure yet that I would want to reject, and am still leaning towards keeping in spite of some negative feedback |
Nope, just opening since I said I would in the discussion issue. We'll still need a formal vote - I'm just kicking off the discussion here. |
|
||
2) Many of the benefits presented in this PDEP can be materialized even with payrrow as an optional dependency. | ||
|
||
For example, as detailed in PDEP-14, it is possible to create a new string data type with the same semantics |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
PDEP 14 does not change performance or memory savings if you do not have pyarrow installed
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
added a note in parentheses at the end of that sentence.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Did you push this up? I don't see anything in parentheses.
The way I am interpreting this now is "we don't need/care for pyarrow strings because we have always had a string data type using Python strings" - is that correct?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I updated the PDEP-15 text, and forgot to remove the PDEP-10 changes.
I've removed the PDEP-10 changes now.
The primary reasons for rejecting this PDEP are twofold: | ||
|
||
1) Requiring pyarrow as a dependency causes installation problems. | ||
- Pyarrow does not fit or has a hard time fitting in space-constrained environments |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think what we could learn from this process is what caused this to change our minds? These issues were discussed leading up to the acceptance of PDEP-10.
The way this is written I think reads more as "we discovered this after the fact" instead of "we decided that X amount of negative feedback on these points was enough to revert". I think there is some value to future PDEPs to set expectations around the latter
cc @pandas-dev/pandas-core @pandas-dev/pandas-triage xref #57073 (comment) for context |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks @lithomas1 for making the updates needed to formally reject PDEP-10.
As discussed in dev meeting on 5/8/24, suggestion is to do a new PDEP that reverts PDEP-10, and keeps any parts we want to keep. |
I'm now leaning towards approving the rejection. My approval of the original PDEP was based solely on improvements to default inference for other dtypes. Despite some recent comments about this, no discussion/clarification has followed on this topic. I'd need to see some positive evidence that the original PDEP-10 authors still intend to support delivering the promised enhancements in this area. Now that the implications of using pd.NA as a default has been discussed in more depth, I suspect that any improved inference would need a couple of dtype variants. |
Yep, I'm planning on updating this current PR to do that, so if anyone has any objections or whatever, we can still discuss here. |
Yeah, I'll probably change the name to PDEP-15 once I get around to moving this to a separate PDEP (probably tomorrow). I was travelling the past week, so didn't really have time then. |
The primary reasons for rejecting this PDEP are twofold: | ||
|
||
1) Requiring pyarrow as a dependency causes installation problems. | ||
- Pyarrow does not fit or has a hard time fitting in space-constrained environments |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Within the context of recent conversation I don't think this comment about AWS is true. AWS distributes an official pandas image for lambda which already includes pyarrow, pandas, and NumPy. This is all required by their own "AWS SDK on pandas" library.
The issue more finely scoped I think is that the default wheel installation via pip into a lambda image exceeds the 256 MB limit. Either using the official AWS provided image or using miniconda should not exceed the space limits
|
||
2) Many of the benefits presented in this PDEP can be materialized even with payrrow as an optional dependency. | ||
|
||
For example, as detailed in PDEP-14, it is possible to create a new string data type with the same semantics |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Did you push this up? I don't see anything in parentheses.
The way I am interpreting this now is "we don't need/care for pyarrow strings because we have always had a string data type using Python strings" - is that correct?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My main comment is that PDEP-10 should be minimally modified, and that PDEP-15 has all the discussion about why we did the rejection.
While both of these reasons are mentioned in the drawbacks section of this PDEP, at the time of the writing | ||
of the PDEP, we underestimated the impact this would have on users, and also downstream developers. | ||
|
||
2) Many of the benefits presented in this PDEP can be materialized even with payrrow as an optional dependency. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I personally don't find this point very convincing. Saying Many of the benefits
but then following it up with one bullet point seems to miss the mark - what are the other many benefits that we don't need pyarrow for? Without pyarrow users are forgoing:
- High performance string operations
- Direct string creation from I/O routines (i.e. no intermediate copies)
- Zero copy data exchange through Arrow C Data Interface
- Performant, memory efficient, and consistent NA handling
On the larger roadmap of pandas this moves us away from tighter Arrow integration, which means we move further away from Arrow compute algorithms / joins and the larger ecosystem of tools that includes streaming, query optimizers, planners, data engines, etc...
I think this argument in its current form is saying "we don't need a car because we have a horse and buggy"
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I personally don't find this point very convincing. Saying
Many of the benefits
but then following it up with one bullet point seems to miss the mark - what are the other many benefits that we don't need pyarrow for? Without pyarrow users are forgoing:
- High performance string operations
- Direct string creation from I/O routines (i.e. no intermediate copies)
- Zero copy data exchange through Arrow C Data Interface
- Performant, memory efficient, and consistent NA handling
On the larger roadmap of pandas this moves us away from tighter Arrow integration, which means we move further away from Arrow compute algorithms / joins and the larger ecosystem of tools that includes streaming, query optimizers, planners, data engines, etc...
I think this argument in its current form is saying "we don't need a car because we have a horse and buggy"
In PDEP-10, there are 3 benefits listed
- pyarrow strings (possible to provide users this benefit without making pyarrow required)
- Nested datatypes (can't have this without arrow, but this is a bit niche)
- Interopability (the alternative is the dataframe interchange protocol, which is more widely adopted at the moment. Not sure about the zero-copy stuff for that, though. I think it also might be possible to implement Arrow C Data interface support without taking on a hard dep on pyarrow)
- Also, the primary beneficiary of this is other dataframe libraries (as opposed to us).
So, IMO, this argument is accurate, in that most of the benefits in PDEP-10 can be made possible (for those user that have pyarrow installed) without making pyarrow required.
The future benefits of Arrow are very compelling, but decisions on making a dependency required should be based on immediate and not future benefits. Like I said before, it is easy to reconsider this decision in a years time if those future benefits are materialize.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If you think points 1 and 3 are possible without pyarrow then the alternatives for that should be laid out in this PDEP, at least at a super high level. I'm assuming point 1 refers to the nanoarrow POC I was sharing; point 3 requires reimplementing the conversions that pyarrow already has. (I personally don't think building either of those from scratch is a good long term solution but it can at least be discussed)
For point 2 how do you know those are niche applications? Its easy to dismiss things that don't exist today as not worthwhile, but I get the feeling that there could be plenty of use cases for the aggregate types, since they have a natural fit with many of the Python containers.
On interoperability the long term prospects for the dataframe interchange protocol seem dubious, and we have even discussed moving that out of pandas (see #56732).
- Also, the primary beneficiary of this is other dataframe libraries (as opposed to us).
The Arrow interchange protocol can be used by any library that needs to work with Arrow data - there is no limit to it being used by other dataframe libraries. It provides a standardized API so that third parties don't need to hack into our internals, which is a direct benefit for us. It also works in two directions - we can be a consumer just as much as a producer.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nested datatypes (can't have this without arrow, but this is a bit niche)
Also wanted to point out that arrow has a decimal128 and decimal256 type which is especially useful for financial calculations where floating point inaccuracies cannot be tolerated, and the arrow decimal types are an extremely significant improvement over using object.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sure, will update and add a note in the PDEP when I get time again.
Co-authored-by: Irv Lustig <irv@princeton.com>
Co-authored-by: Irv Lustig <irv@princeton.com>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm fine with this PDEP, although I'm unsure whether using language as "we, the core team", should appear in a PDEP.
This pull request is stale because it has been open for thirty days with no activity. Please update and respond to this comment if you're still interested in working on this. |
@lithomas1 I think you should update the PDEP based on comments from @WillAyd so we can move this forward and have a vote. |
My vote on this will heavily depend if we can come to an agreement on PDEP 14 |
Thanks for the reminder, this slipped off my radar.
Yeah, ideally, we'd vote on these at the same time. |
Marking as draft as although I've pushed up another commit, I haven't fully finished and refined my thoughts yet. I also need to fix the formatting issues in this PDEP. |
@lithomas1 now that PDEP-14 is accepted, we don't need to require |
I was out for the summer, but I think this PDEP should be updated such that all comments were addressed. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the changes.
@pandas-dev/pandas-core Please provide any additional comments so we can move this to a vote. If no additional comments are requesting any changes within 15 days, we'll open up the voting period.
|
||
1) Requiring pyarrow as a dependency causes installation problems. | ||
|
||
- Pyarrow does not fit or has a hard time fitting in space-constrained environments |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this really exaggerates the problem on AWS. AWS has long distributed its own AWS SDK for pandas library (formerly called awswrangler) which uses pyarrow to better integrate with many of its services (ex: pyarrow is used for high performance data exports to/from AWS Redshift, rather than using a traditional ODBC driver)
The issue here is really just scoped to users that don't want to use the AWS Lambda Managed Layer, but instead want to build their environment from scratch, assumedly without the AWS SDK for pandas. Even then, it may not be a current issue given the drastic reductions in the binary size of pyarrow through both pip and conda
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think it's only conda that's been drastically reduced - install pyarrow and pandas in a fresh venv and it already hits 295 MB. And the wheel size on PyPI is still ~40MB, so we'd be noticeably increasing the load on PyPI by making PyArrow required
My current stance is: if the desire was to make PyArrow dtypes the default for all dtypes, then ok, maybe that'd be justified. But probably not if it's just for the sake of strings, for which I think that
- recommending
pip install pandas[pyarrow]
in all instructions - auto-inferring pyarrow strings if pyarrow is installed
should be enough
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sounds good to me.
I'll add this to the PDEP.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
And the wheel size on PyPI is still ~40MB, so we'd be noticeably increasing the load on PyPI by making PyArrow required
Out of curiosity - is this something that the PyPi maintainers have mentioned as a problem?
If so, the plot twist is then that we should really be pushing AWS users towards using the pre-provided image, rather than building their own from scratch. I believe that would forgo hitting PyPi altogether, and even if it doesn't, it is still smaller than what people are building themselves (see #54466 (comment))
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If so, the plot twist is then that we should really be pushing AWS users towards using the pre-provided image, rather than building their own from scratch. I believe that would forgo hitting PyPi altogether, and even if it doesn't, it is still smaller than what people are building themselves (see #54466 (comment))
Wouldn't we then be asking our users to be dependent on the AWS people updating their pre-provided image whenever we created a new release of pandas
or the arrow team did a new release of pyarrow
?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yea I think that is generally how lambda works, even with the overall Python version. It's not quite an "anything goes" type of execution environment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There's this post from 2021, in which they write about the monthly PyPI bill being almost 2 million US dollars
https://dustingram.com/articles/2021/04/14/powering-the-python-package-index-in-2021/
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There's this post from 2021, in which they write about the monthly PyPI bill being almost 2 million US dollars
https://dustingram.com/articles/2021/04/14/powering-the-python-package-index-in-2021/
I'm not too worried about this - PyPI's bill is subsidized anyways. Of course, it is nice to reduce our load on PyPI, but I don't think we are close to being the worse offenders here (those would probably be something like tensorflow and pytorch), and it's important to keep in mind that a lot of people have pyarrow installed for whatever reason anyways.
I would mostly only be concerned about an increase in size in our own pandas package (since PyPI does limit the total size of all packages uploaded by a project, and raising the limit is a manual and annoying process)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I've made another go at clarifying point1, incorporating the feedback here.
PTAL @WillAyd @MarcoGorelli
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the feedback, but I still disagree with calling the AWS-provided layer a "workaround" - essentially it is the canonical approach to solve an issue that has existed for years with pandas and lambda functions.
A quick google search for something how "how to run pandas on aws lambda" yields a slew of conflicting results on how to get this to work. If the AWS-provided layer is a workaround, then what are we calling the proper approach?
https://stackoverflow.com/questions/36054976/pandas-aws-lambda
https://stackoverflow.com/questions/53824556/how-to-install-numpy-and-pandas-for-aws-lambdas
https://medium.com/swlh/how-to-add-python-pandas-layer-to-aws-lambda-bab5ea7ced4f
https://medium.com/@johnnymao/how-to-use-pandas-in-your-aws-lambda-function-c3ce29f6f189
https://medium.com/@shimo164/lambda-layer-to-use-numpy-and-pandas-in-aws-lambda-function-8a0e040faa18
https://www.youtube.com/watch?v=1UDEp90S9h8
that support Arrow in a zero-copy manner. | ||
|
||
- Support for the Arrow C Data interface in pandas and other libraries in the ecosystem is still very new, though, | ||
(support in pandas itself was only added as of pandas 2.2), and the dataframe interchange protocol, which allows |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't know that this is a true statement about the dataframe interchange protocol being better supported in other libraries. Very recently, the Arrow C Data interface has been adopted by a ton of new libraries.
@MarcoGorelli is probably the most knowledgeable on this overall
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yup - pandas also requires pyarrow in some cases for the interchange protocol anyway
pandas/pandas/core/interchange/from_dataframe.py
Lines 461 to 469 in 56ea76a
assert length is not None, "`length` must be specified for a bit-mask buffer." | |
pa = import_optional_dependency("pyarrow") | |
arr = pa.BooleanArray.from_buffers( | |
pa.bool_(), | |
length, | |
[None, pa.foreign_buffer(buffer.ptr, length)], | |
offset=offset, | |
) | |
return np.asarray(arr) |
Besides, the interchange protocol doesn't support nested data types nor dates, and the progress on the project seems to have completely stalled. I'd suggest not mentioning it at all if it's OK
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Cool, if I understand corrrectly, though, the main benefit is to other dataframe libraries, as of right now.
(doing a quick search: it seems like lots of downstream libraries are using this as a mechanism to gain support for polars and co. in addition to pre-existing pandas support)
Are there any libraries that didn't support pandas before, but now support pandas due to the interchange protocol?
This could be something good to highlight.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Are there any libraries that didn't support pandas before, but now support pandas due to the interchange protocol?
nope, there's been no such case
arrays to those dtypes by default, without forcing a pyarrow requirement on users, | ||
as there is no Python/numpy equivalent for these dtypes). | ||
|
||
- Interoperability |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think @MarcoGorelli point was to remove the Interoperability section entirely, but if that's not true then I don't understand the point this is trying to make in its current form.
The beneficiary of the Arrow C Data interface is not just other dataframe libraries - a decent listing can be found here:
For a direct benefit to pandas, it helps with I/O to boost performance, ensure proper data types, and reduce the amount of code burden. We have already seen this benefit with ADBC drivers, and from the link above, it looks like there is some near-term potential for it to help Excel I/O via fastexcel
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sure, will take it out the next go.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The fact that we don't require PyArrow might put is in a bind for downstream libraries that want interchange with pandas, but themselves probably aren't in a position to require PyArrow. In particular, this conversation is happening with seaborn:
mwaskom/seaborn#3782 (comment)
Somewhat unfortunately, this may mean that we are asked to put more maintenance work into the interchange protocol
This pull request is stale because it has been open for thirty days with no activity. Please update and respond to this comment if you're still interested in working on this. |
doc/source/whatsnew/vX.X.X.rst
file if fixing a bug or adding a new feature.Just something quick and dirty I threw up that we should talk about tomorrow at the dev call.