Path: csiph.com!fu-berlin.de!bofh.it!news.nic.it!robomod From: Andreas Tille Newsgroups: linux.debian.maint.python Subject: Re: Remaining numpy 1.24 issues on 32bit architectures Date: Mon, 30 Jan 2023 13:50:01 +0100 Message-ID: References: X-Mailbox-Line: From debian-python-request@lists.debian.org Mon Jan 30 12:41:25 2023 Old-Return-Path: X-Amavis-Spam-Status: No, score=-4.98 tagged_above=-10000 required=5.3 tests=[BAYES_00=-2, LDO_WHITELIST=-5, MEDS2=2, MURPHY_DRUGS_REL8=0.02] autolearn=no autolearn_force=no X-Policyd-Weight: using cached result; rate: -3.5 MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline X-Mailing-List: archive/latest/20759 List-ID: List-URL: List-Archive: https://lists.debian.org/msgid-search/Y9e63xFuEDZfjm6s@an3as.eu Approved: robomod@news.nic.it Lines: 65 Organization: linux.* mail to news gateway Sender: robomod@news.nic.it X-Original-Cc: Debian Med Project List , Debian Python X-Original-Date: Mon, 30 Jan 2023 13:41:03 +0100 X-Original-Message-ID: X-Original-References: <20230130113737.3pi35wrnnlvhy32u@nileshpatra.info> Xref: csiph.com linux.debian.maint.python:14873 Hi Nilesh, Am Mon, Jan 30, 2023 at 05:07:37PM +0530 schrieb Nilesh Patra: > > On Mon, Jan 30, 2023 at 09:12:57AM +0100, Andreas Tille wrote: > > I made some tiny steps forward ("only" 84 failures instead of 89 when I > > wrote my first mail) in the numpy 1.24 migration for 32bit architectures > > but I'm facing issues I do not have a real clue for. In > > > > https://salsa.debian.org/med-team/python-skbio/-/blob/master/debian/patches/numpy-1.24.patch#L123-L126 > > Apologies for pointing the discussion into an orthogonal direction for > once. Ofcourse, we could try fixing these, but if you look closely, skbio > has never built on 32 bit archs ever since around 2016 on i386[3] and > it has never built on the rest of 32 bit ever since it entered debian[4] > and now this new upstream FTBFS that you point to, won't really block > migration in any way. Hmmm, I have checked https://buildd.debian.org/status/package.php?p=python-skbio before I started to investigate time into this and it says `uncompiled` for the 32bit architectures. > So my question is this: Why are we trying hard to fix this on 32-bit _now_ > given that the upstream support has never been solid for this package on > 32-bit archs? I admit the 0.5.8-2 has migrated which I did not expected since when I was looking excuses contained those build problems. > > ... > > which obviosly[2] failed. I wonder whether someone might give some > > hints how to get dtypes consistently to one integer representation which > > is the background of nearly all these test suite issues. > > I can think of two alternatives to fix this: > > 1. There are a few type conversions to "int" (.astype(int)) in the skbio source code. > This defaults to 32-bit integer type on 32-bit machines. Explicitly > casting them to 64-bit can fix this. I happened to write a similar patch > for another package recently, see[5] if it helps. > > 2. Just ignore datatypes while comparing pandas dataframes with > `check_dtype` parameter. An example/reference patch here[6] Thanks for the additional hints (and have a nice trip). Kind regards Andreas. > > > [1] https://buildd.debian.org/status/package.php?p=python-skbio&suite=experimental > > [2] https://salsa.debian.org/med-team/python-skbio/-/jobs/3868951 > [3]: https://buildd.debian.org/status/logs.php?pkg=python-skbio&arch=i386 > [4]:https://buildd.debian.org/status/logs.php?pkg=python-skbio&arch=armhf > [5]: https://salsa.debian.org/med-team/python-bioframe/-/blob/master/debian/patches/32-bits.patch > [6]: https://salsa.debian.org/python-team/packages/python-upsetplot/-/blob/master/debian/patches/ignore-dtype-while-asserting.patch > > -- > Best, > Nilesh -- http://fam-tille.de