You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Numpy 2.0 is a sufficiently breaking change that it feels like trying to support it in a bugfix release in 6.0 is not really sensible any more: some things are just going to break and that is unexpected for a bugfix release.
But it would be very good to have an official astropy release that does work with numpy 2.0. Hence, @taldcroft and I wondered whether it might make sense to aim for a somewhat earlier than usual 6.1. It is not completely clear how much earlier, but the beta version of numpy 2.0 has been released and the expectation is that the final version will be about a month from now. Does this make sense?
My own feeling is that 6.1 in itself is looking like a rather minor release, with not too much that is new (which is not illogical, since bigger things have to wait for 7.0 anyway). So, from that respect, perhaps it is not so crazy to do feature-freeze in about a week?
Alternative
As noted by @pllim below, we can just add a big fat warning in 6.0.1 that the numpy 2.0 changes are sufficiently large that we cannot guarantee that astropy will run the same way with numpy 2.0 as it did with 1.x in all cases.
The obvious advantage is that in that case we can just stick to our own schedule.
The disadvantage is that people may not read very carefully and those just trying to use the latest versions of everything may get hurt.
I haven't been following closely so I am confused. I thought #16142 already took care of "copy" in a backward compatible way and all the extra stuff are "nice to have"?
@pllim - I think the conclusion is dawning on us that numpy 2.0 is enough of a break (mostly because of the changed meaning of copy) that making all code work consistently between numpy 1 and numpy 2 is not going to work. Instead, people have to adjust. Since that means that astrop 6.0.x will work somewhat differently with numpy 1.x and 2.0 (if one passes in copy arguments), it seems cleanest just to say that for 6.0, numpy 2.0 is not supported, and for 6.1 to explain that one has to be careful.
But perhaps we are being too careful? After all, most code will just work fine, and most breakage gives a clear exception.
That would push v6.1 ahead for 2-3 months. I am not sure if I am comfortable with that. 🤔 🤪
Is it an option to issue a memo that "yes you can use v6.0.1 with numpy 2.0 but..." and then they can wait a bit more for v6.1 without us changing our release schedule? Or is that bad because we are changing the behavior of this package between v6.0.1 and v6.1? But then again, v6.1 has other non-backward compatible changes too so maybe that is ok?
TBH it is going to take the scientific python world months to catch up anyway. At that time, v6.1 is due to release regardless.
@pllim - your comments make me waver in the opposite direction again. So far, it has certainly been the plan to just have numpy 2.0 be supported in 6.0.1... Let me put this on top as an alternative.
A bit less than a month actually, v6.1 Feature Freeze was planed for 2024-04-19.
Given the potentially breaking change that Numpy 2.0 will be for people in general, I agree with the initial proposal, adding support in Astropy 6.1 and moving forward a bit the release. I think that will make things more clear for downstream packages and users.
We have another reason to go straight to 6.1 for numpy 2 compat: string arrays now support addition, which implies a change in table aggregation rules.
See #16211 (comment)
Given that it might be a bit soon to do feature freeze this week, would doing it Friday next week be soon enough? Could someone confirm when Numpy 2.0 is scheduled for?
Could someone confirm when Numpy 2.0 is scheduled for?
Here's the latest estimate numpy/numpy#24300 (comment)
Currently, the pessimistic timeline puts numpy 2.0.0rc1 in 3 days from now, but there was no public statement for 11 days so I supposed it could still be delayed.
I've also not seen much else either, but rc1 seems to be getting close. I suspect a final release in about a month from now. To me, the main benefit of having a separate astropy release at the same time is that people can adjust both their numpy and astropy pieces to deal with the numpy changes (probably mostly the copy keyword, but also the changes in representation of scalars for doctests, etc.). E.g., I don't think we'd want to backport the astropy doc changes (#15065) to 6.0.x, yet it would be nice if we had docs that used the newest numpy.
Anyway, maybe the right time to send out a message to astropy-dev to ask if a Friday 29th freeze would really inconvenience anyone? My sense so far is that this release has been mostly about refactoring - the what's-new only has that io.ascii now uses 64-bit by default (which is really a numpy 2.0 driven change). So, in that respect little to stop us from releasing a bit earlier.
That one is being fixed upstream; I don't think there is much doubt that numpy will release fairly soon, and we might as well get ready for that. Unless there are other things we really want to have in 6.1 -- among the 80 open ones, I don't see anything of great interest, but that large number suggests that a few weeks of triaging those would be good regardless!
I see that @saimn just opened #16230, so it looks like we're going for a bugfix release after all ?
In any case, if 6.0.x is supposed to have numpy 2.0 compat, please consider backporting #16229 first
Activity
astrofrog commentedon Mar 15, 2024
I'd be happy with this, but which PRs are important to include for Numpy 2.0 support?
neutrinoceros commentedon Mar 15, 2024
@astrofrog see #16167 (comment) for a list of PRs that triggered this discussion.
astrofrog commentedon Mar 15, 2024
Thanks!
astrofrog commentedon Mar 15, 2024
So just to be clear, should we release 6.1 just before or just after numpy 2.0 final is released ideally?
neutrinoceros commentedon Mar 15, 2024
ABI compatibility requires we make a release (not necessarily 6.1) at some point between numpy 2.0rc1 and 2.0 final.
pllim commentedon Mar 15, 2024
I haven't been following closely so I am confused. I thought #16142 already took care of "copy" in a backward compatible way and all the extra stuff are "nice to have"?
mhvk commentedon Mar 15, 2024
@pllim - I think the conclusion is dawning on us that numpy 2.0 is enough of a break (mostly because of the changed meaning of
copy) that making all code work consistently between numpy 1 and numpy 2 is not going to work. Instead, people have to adjust. Since that means that astrop 6.0.x will work somewhat differently with numpy 1.x and 2.0 (if one passes incopyarguments), it seems cleanest just to say that for 6.0, numpy 2.0 is not supported, and for 6.1 to explain that one has to be careful.But perhaps we are being too careful? After all, most code will just work fine, and most breakage gives a clear exception.
pllim commentedon Mar 15, 2024
That would push v6.1 ahead for 2-3 months. I am not sure if I am comfortable with that. 🤔 🤪
Is it an option to issue a memo that "yes you can use v6.0.1 with numpy 2.0 but..." and then they can wait a bit more for v6.1 without us changing our release schedule? Or is that bad because we are changing the behavior of this package between v6.0.1 and v6.1? But then again, v6.1 has other non-backward compatible changes too so maybe that is ok?
TBH it is going to take the scientific python world months to catch up anyway. At that time, v6.1 is due to release regardless.
mhvk commentedon Mar 15, 2024
@pllim - your comments make me waver in the opposite direction again. So far, it has certainly been the plan to just have numpy 2.0 be supported in 6.0.1... Let me put this on top as an alternative.
saimn commentedon Mar 15, 2024
A bit less than a month actually, v6.1 Feature Freeze was planed for 2024-04-19.
Given the potentially breaking change that Numpy 2.0 will be for people in general, I agree with the initial proposal, adding support in Astropy 6.1 and moving forward a bit the release. I think that will make things more clear for downstream packages and users.
neutrinoceros commentedon Mar 18, 2024
We have another reason to go straight to 6.1 for numpy 2 compat: string arrays now support addition, which implies a change in table aggregation rules.
See #16211 (comment)
mhvk commentedon Mar 20, 2024
FWIW, scipy is doing an off-cycle release for numpy 2.0 support. https://mail.python.org/archives/list/[email protected]/thread/YM2MVB5T2PCL7MPEQDT5OGFH6GZOXU6J/
astrofrog commentedon Mar 20, 2024
Given that it might be a bit soon to do feature freeze this week, would doing it Friday next week be soon enough? Could someone confirm when Numpy 2.0 is scheduled for?
pllim commentedon Mar 20, 2024
I still think v6.0.x is compatible enough as it is... 👀
neutrinoceros commentedon Mar 20, 2024
Here's the latest estimate numpy/numpy#24300 (comment)
Currently, the pessimistic timeline puts numpy 2.0.0rc1 in 3 days from now, but there was no public statement for 11 days so I supposed it could still be delayed.
mhvk commentedon Mar 20, 2024
I've also not seen much else either, but
rc1seems to be getting close. I suspect a final release in about a month from now. To me, the main benefit of having a separate astropy release at the same time is that people can adjust both their numpy and astropy pieces to deal with the numpy changes (probably mostly thecopykeyword, but also the changes in representation of scalars for doctests, etc.). E.g., I don't think we'd want to backport the astropy doc changes (#15065) to 6.0.x, yet it would be nice if we had docs that used the newest numpy.Anyway, maybe the right time to send out a message to astropy-dev to ask if a Friday 29th freeze would really inconvenience anyone? My sense so far is that this release has been mostly about refactoring - the what's-new only has that
io.asciinow uses 64-bit by default (which is really a numpy 2.0 driven change). So, in that respect little to stop us from releasing a bit earlier.pllim commentedon Mar 20, 2024
Well... Our devdeps is currently failing, so it won't make sense for any release until #16211 is resolved anyway.
mhvk commentedon Mar 20, 2024
That one is being fixed upstream; I don't think there is much doubt that numpy will release fairly soon, and we might as well get ready for that. Unless there are other things we really want to have in 6.1 -- among the 80 open ones, I don't see anything of great interest, but that large number suggests that a few weeks of triaging those would be good regardless!
pllim commentedon Mar 20, 2024
numpy/numpy#26079 was abandoned. I hope they still plan to fix it...
mhvk commentedon Mar 20, 2024
Indeed, because a better way was found in discussion!
neutrinoceros commentedon Mar 25, 2024
I see that @saimn just opened #16230, so it looks like we're going for a bugfix release after all ?
In any case, if 6.0.x is supposed to have numpy 2.0 compat, please consider backporting #16229 first
saimn commentedon Mar 25, 2024
#16231: Numpy 2.0 support will come in v6.1.0, as discussed here.
astrofrog commentedon Mar 25, 2024
I've emailed astropy-dev notifying people of the upcoming feature freeze this Friday.
pllim commentedon May 6, 2024
astropy v6.1.0 is released and pins to numpy 2.0rc for build.