patch setuptools to allow for parallel usage

Question

patch setuptools to allow for parallel usage

Closed this issue 12 years ago · 41 comments

Currently we have to force the spkgs that use setuptools to build serially because setuptools does not do any file locking with easy_install.pth. This is an inconvenience that has a fairly straightforward fix.

Depends on #11874
Depends on #12994

CC: @jdemeyer @kini @jasongrout

Component: build

Author: Volker Braun

Branch: b276471

Reviewer: R. Andrew Ohana

Issue created by migration from https://trac.sagemath.org/ticket/13201

Answer 1 · 2012-07-04T22:25:39.000Z

Description changed:

--- 
+++ 
@@ -3,4 +3,4 @@
 
 Installation Instructions:
 * Use the new spkg [http://wstein.org/home/ohanar/spkgs/setuptools-0.6.16.p1.spkg](http://wstein.org/home/ohanar/spkgs/setuptools-0.6.16.p1.spkg)
-* Apply ... to the root repo.
+* Apply [attachment: trac13201.patch](https://github.com/sagemath/sage-prod/files/10655899/trac13201.patch.gz) to the root repo.

Answer 2 · 2012-07-04T22:25:39.000Z

Author: R. Andrew Ohana

Answer 3 · 2012-08-13T12:39:39.000Z

comment:2

Why re-invent the wheel? Why not use http://docs.python.org/library/fcntl.html#fcntl.lockf?

Also, this should obviously be reported upstream.

Answer 4 · 2012-09-04T21:01:41.000Z

comment:3

Replying to @jdemeyer:

Why re-invent the wheel? Why not use http://docs.python.org/library/fcntl.html#fcntl.lockf?

Because I was being dumb and couldn't find that function :).

Also, this should obviously be reported upstream.

Yup, although upstream is basically dead. IMO we should switch to distribute (I'll see about making an SPKG).

Answer 5 · 2012-09-06T11:17:56.000Z

new patch applied to setuptools; for review purposes

Answer 6 · 2012-09-06T11:22:27.000Z

Attachment: easy_install_lock.patch.gz

Attachment: trac13201.patch.gz

apply to root repo

Answer 7 · 2012-09-06T11:23:54.000Z

Changed dependencies from #11874 to #11874, #12994

Answer 8 · 2012-09-06T11:35:01.000Z

comment:6

I unfortuntely don't understand setuptools well enough to review this. It seems there are some changes unrelated to the locking. What do they do?

Also, do we really need to do this every time Sage runs (could it be moved to sage-location?):

    # Hack around setuptools since --egg-path isn't fully respected 
    sed -i 's-.*sagenb.*-\.\./\.\./\.\./\.\./devel/sagenb-' \ 
        "$SAGE_LOCAL/lib/python/site-packages/easy-install.pth"

You also need to handle the case that $SAGE_ROOT isn't writeable.

Answer 9 · 2012-09-06T11:50:54.000Z

comment:7

Replying to @jdemeyer:

I unfortuntely don't understand setuptools well enough to review this. It seems there are some changes unrelated to the locking. What do they do?

Internally setuptools identifies packages based on a normalized path (no absolute paths without symlinks, and a normalized case for case-insensitive filesystems). It reads these from easy-install.pth, and then creates a brand new easy-install.pth if it detects that it needs to. The issue is that it loads the list of packages when it starts, and then writes it back when it finishes, if any changes were made to the file in the meantime, setuptools doesn't detect them, and just overrides them with its new easy-install.pth. What I did was add a bit of code for reloading the file right before writing a new one -- making sure things that were deleted don't pop back up, and making sure things that were added don't disappear.

Also, do we really need to do this every time Sage runs (could it be moved to sage-location?):
    # Hack around setuptools since --egg-path isn't fully respected 
    sed -i 's-.*sagenb.*-\.\./\.\./\.\./\.\./devel/sagenb-' \ 
        "$SAGE_LOCAL/lib/python/site-packages/easy-install.pth" 
You also need to handle the case that $SAGE_ROOT isn't writeable.

Well no you don't have to run it every time sage starts, just every time a package using setuptools is installed, since setuptools always inscribes the absolute path into easy-install.pth (even if you specify a relative path for --egg-path).

Answer 10 · 2012-09-06T11:57:36.000Z

comment:8

Since sage-location specifically deals with rewriting paths, I would do it there.

Answer 11 · 2012-09-06T23:06:45.000Z

comment:9

I thought that was for rewriting paths when SAGE_ROOT changed. The issue is that setuptools rewrites paths, regardless of what happens to SAGE_ROOT, so we have to overwrite their rewrites. An alternative (now that we have the flask notebook has been merged), is to treat sagenb like any other upstream package, and just install it in site-packages (I don't think the sagenb developers would be opposed to this, we are already planning on doing this in the transition to git).

Answer 12 · 2012-09-07T07:30:13.000Z

comment:10

Replying to @ohanar:

I thought that was for rewriting paths when SAGE_ROOT changed.

Well, the reason you need to rewrite the setuptools paths is to allow relocation, right? (or am I missing something?)

Answer 13 · 2012-09-07T07:45:12.000Z

comment:11

Replying to @jdemeyer:

Well, the reason you need to rewrite the setuptools paths is to allow relocation, right? (or am I missing something?)

Well we also support clones for sagenb, which will brake every time setuptools is run (it readlinks everything). If we don't care about that, then we could.

I'm CCing Keshav and Jason on this ticket, they might have an opinion on how we handle this (since currently it only affects sagenb).

Answer 14 · 2012-09-07T08:28:09.000Z

comment:12

Since sagenb is now distributed without a repository, it might make sense for it not to be installed in SAGE_ROOT/devel/. The reason for it to be in SAGE_ROOT/devel/sagenb is so that the barrier is way lower for people wanting to patch it and work with it. If we're making them clone the git repository anyway, the barrier isn't lower.

So what do you think, kini? Should we just install it like a normal spkg? If they want to develop, they need to git clone and do sage setup.py develop?

(personally, I'm mildly in favor of going back to installing the git repo in SAGE_ROOT/devel/sagenb...)

Answer 15 · 2012-09-07T08:35:40.000Z

comment:13

Absolutely. I 100% think that sagenb should be installed like any other python package, i.e. in the site packages directory.

Answer 16 · 2013-03-16T22:42:12.000Z

comment:14

In any case, this needs work because it assumes that $SAGE_ROOT is writable and it needs to be rebased.

Answer 17 · 2013-03-17T09:34:06.000Z

comment:15

Replying to @ohanar:

Replying to @jdemeyer:

Why re-invent the wheel? Why not use http://docs.python.org/library/fcntl.html#fcntl.lockf?

Because I was being dumb and couldn't find that function :).

Also, this should obviously be reported upstream.

Yup, although upstream is basically dead. IMO we should switch to distribute (I'll see about making an SPKG).

At this point I'd like to point you to SPKG.txt and actually read it:

= setuptools =

== Description ==

setuptools is a collection of enhancements to the Python distutils (for Python 2.3.5 and up on mo
st platforms; 64-bit platforms require a minimum of Python 2.4) that allow you to more easily bui
ld and distribute Python packages, especially ones that have dependencies on other packages.

Website: http://pypi.python.org/pypi/setuptools/

The present spkg is based on the fork of setuptools known as distribute.
Website: http://pypi.python.org/pypi/distribute

And

=== setuptools-0.6.16 (Francois Bissey, June 1, 2011) ===
 * Switch to the "distribute" fork of setuptools and update to 0.6.16, 
 * adopt a Gentoo patch to avoid warnings with python-2.7 (works with 2.6 too). 
 * Remove the two windows binaries and patch so setup.py doesn't
   try to install them.

This was a pre-requisite to the upgrade to python-2.7.3. Fell free to upgrade to a newer version of distribute and change the name if you want too.

Answer 18 · 2013-03-17T09:50:25.000Z

comment:16

Yeah, I made that remark about 5 minutes before completely reading through the SPKG.txt. I haven't really looked at this in awhile, but I probably won't be touching this soon, as I'm really trying to put all of my time into working on the transition to git.

Answer 19 · 2013-03-17T09:53:43.000Z

comment:17

I checked if you said that in the ticket but didn't see it so I assumed... too much. ;)

Answer 20 · 2014-02-28T15:07:45.000Z

Branch: u/vbraun/parallel_setuptools

Answer 21 · 2014-02-28T15:24:15.000Z

Commit: a899c2c

Answer 22 · 2014-02-28T15:24:15.000Z

Branch pushed to git repo; I updated commit sha1. New commits:

`a899c2c`	`remove fake deps`

Answer 23 · 2014-02-28T20:07:38.000Z

Branch pushed to git repo; I updated commit sha1. New commits:

`39481e1`	`avoid duplicated entries in pth file`

Answer 24 · 2014-02-28T20:07:38.000Z

Changed commit from a899c2c to 39481e1

Answer 25 · 2014-02-28T20:23:19.000Z

Changed author from R. Andrew Ohana to Volker Braun, R. Andrew Ohana

Answer 26 · 2014-02-28T20:23:19.000Z

Description changed:

--- 
+++ 
@@ -1,6 +1,2 @@
 Currently we have to force the spkgs that use setuptools to build serially because setuptools does not do any file locking with `easy_install.pth`. This is an inconvenience that has a fairly straightforward fix.
 
-
-Installation Instructions:
-* Use the new spkg [http://wstein.org/home/ohanar/spkgs/setuptools-0.6.16.p1.spkg](http://wstein.org/home/ohanar/spkgs/setuptools-0.6.16.p1.spkg)
-* Apply [attachment: trac13201.patch](https://github.com/sagemath/sage-prod/files/10655899/trac13201.patch.gz) to the root repo.

Answer 27 · 2014-02-28T20:23:19.000Z

comment:23

New attempt... The patch didn't apply, so I wrote a different version. It is now IMHO clearer though perhaps at the cost of rewriting a pth file even if it is not dirty. But then it is hardly a performance-critical step.

Answer 28 · 2014-02-28T20:24:15.000Z

comment:24

We can't push the changes upstream since fcntl is unix only but setuptools has to work on (non-cygwin) windows

Answer 29 · 2014-02-28T20:26:43.000Z

comment:25

Critical as we currently have some races as I noticed in parallel testing.

Answer 30 · 2014-02-28T21:29:29.000Z

Changed author from Volker Braun, R. Andrew Ohana to Volker Braun

Answer 31 · 2014-02-28T21:29:29.000Z

comment:26

Looks fine, although I'll have to test it out. It would be good to get some sort of proper locking in upstream -- maybe doing something along the lines of http://code.activestate.com/recipes/65203/.

Answer 32 · 2014-02-28T21:43:27.000Z

Reviewer: R. Andrew Ohana

Answer 33 · 2014-02-28T21:43:27.000Z

comment:27

The patch version of setuptools will need a version bump to force a rebuild. Also, it would be good to have the patch documented in the SPKG.txt. Pending those two things, positive review.

Answer 34 · 2014-03-11T12:53:13.000Z

Changed commit from 39481e1 to 25f77ee

Answer 35 · 2014-03-11T12:53:13.000Z

Branch pushed to git repo; I updated commit sha1. New commits:

`25f77ee`	`added documentation, patchlevel bump`

Answer 36 · 2014-03-11T12:53:41.000Z

comment:29

Done.

Answer 37 · 2014-03-11T12:56:48.000Z

Branch pushed to git repo; I updated commit sha1 and set ticket back to needs_review. New commits:

`b276471`	`resolved merge conflict`

Answer 38 · 2014-03-11T12:56:48.000Z

Changed commit from 25f77ee to b276471

Answer 39 · 2014-03-11T16:15:35.000Z

Changed branch from u/vbraun/parallel_setuptools to b276471

Answer 40 · 2014-10-31T13:39:09.000Z

Changed commit from b276471 to none

Answer 41 · 2014-10-31T13:39:09.000Z

comment:32

You just broke sagenb: #17268