Support installation of Python packages into separate venvs depending on the python version

[mkoeppe]

#27824 made $SAGE_LOCAL a venv if a suitable system python3 is found.

In this ticket, we add support for installing several python versions into the same SAGE_LOCAL:

• keeping install records for Python packages in $SAGE_VENV/var/lib/sage/{installed,scripts,wheels}, separate from those of non-Python packages ($SAGE_LOCAL/var/lib/sage/{installed,scripts})
• likewise for $SAGE_LOCAL/var/tmp/sage/build
• here $SAGE_VENV defaults to $SAGE_LOCAL, but can be overridden to an arbitrary directory that will be used as the wheel-building venv, for example $SAGE_LOCAL/var/lib/sage/venv/$PYTHON_TAG.

This is activated by running, for example, make PYTHON_FOR_VENV=/usr/bin/python3.8 SAGE_VENV=$SAGE_LOCAL/var/lib/sage/venv/$PYTHON_TAG`.

To implement this, an SPKG needs an indication that it is a Python package - for example #30024.

This is mainly for #29039; but also facilitates testing with several Python versions without having to rebuild the Sage distribution.

This depends on:

• #30534 Repackage pynac as a pip-installable package

[mkoeppe]

A simple change. Compiles from a fresh checkout and runs without problems on macOS.

---

New commits:

​1363425	In a python3 build, install all Python packages into a venv
​a394268	sage-pip-install: Use PYTHON=sage-python23

[mkoeppe]

In the branch #27824, the venv is created by slightly more complicated code in build/make/Makefile. This complexity is not necessary for the present ticket.

[dimpase]

Does this ticket make sense separately, or only as part of #27824 ?

[mkoeppe]

#27824 is the use case that I have in mind. This ticket is a less intrusive change and should go through the betas first.

[jhpalmieri]

I tried building with this branch on OS X, and I got doctest failures:

sage -t --long src/sage/plot/plot.py  # 1 doctest failed
sage -t --long src/sage/matrix/matrix2.pyx  # 2 doctests failed
sage -t --long src/sage/finance/time_series.pyx  # 5 doctests failed
sage -t --long src/sage/misc/trace.py  # 4 doctests failed
sage -t --long src/sage/env.py  # 1 doctest failed
sage -t --long src/sage/repl/configuration.py  # 1 doctest failed
sage -t --long src/sage/schemes/affine/affine_homset.py  # 1 doctest failed

For example

File "src/sage/plot/plot.py", line 521, in sage.plot.plot
Failed example:
    os.system("sage -c \"if 'matplotlib' in sys.modules: sys.exit(1)\"") # long time
Expected:
    0
Got:
    Traceback (most recent call last):
      File "/Users/jpalmier/Desktop/Sage/sage_builds/TESTING/sage-9.1.beta0/local/bin/sage-eval", line 4, in <module>
        from sage.all import *
    ModuleNotFoundError: No module named 'sage'
    256

and

**********************************************************************
File "src/sage/matrix/matrix2.pyx", line 1127, in sage.matrix.matrix2.Matrix.pseudoinverse
Failed example:
    M.pseudoinverse()  # tol 1e-15
Expected:
    [0.0620518477661335 0.0206839492553778 0.0124103695532267]
    [ 0.124103695532267 0.0413678985107557 0.0248207391064534]
    [ 0.186155543298400 0.0620518477661335 0.0372311086596801]
Got:
    [0.0620518477661334 0.0206839492553778 0.0124103695532267]
    [ 0.124103695532267 0.0413678985107556 0.0248207391064534]
    [ 0.186155543298400 0.0620518477661335 0.0372311086596801]
Tolerance exceeded in 2 of 9:
    0.0620518477661335 vs 0.0620518477661334, tolerance 2e-15 > 1e-15
    0.0413678985107557 vs 0.0413678985107556, tolerance 3e-15 > 1e-15
**********************************************************************
File "src/sage/matrix/matrix2.pyx", line 1131, in sage.matrix.matrix2.Matrix.pseudoinverse
Failed example:
    M.pseudoinverse(algorithm="numpy")  # tol 1e-15
Expected:
    [0.0620518477661335 0.0206839492553778 0.0124103695532267]
    [ 0.124103695532267 0.0413678985107557 0.0248207391064534]
    [ 0.186155543298400 0.0620518477661335 0.0372311086596801]
Got:
    [0.0620518477661334 0.0206839492553778 0.0124103695532267]
    [ 0.124103695532267 0.0413678985107556 0.0248207391064534]
    [ 0.186155543298400 0.0620518477661335 0.0372311086596801]
Tolerance exceeded in 2 of 9:
    0.0620518477661335 vs 0.0620518477661334, tolerance 2e-15 > 1e-15
    0.0413678985107557 vs 0.0413678985107556, tolerance 3e-15 > 1e-15
**********************************************************************
1 item had failures:
   2 of  25 in sage.matrix.matrix2.Matrix.pseudoinverse
    [2472 tests, 2 failures, 9.06 s]

[mkoeppe]

Thanks for testing!

[mkoeppe]

Replying to jhpalmieri:

sage -t --long src/sage/matrix/matrix2.pyx # 2 doctests failed

Not sure what's causing the numerical failures.

sage -t --long src/sage/env.py # 1 doctest failed

For this one, I have created #29022.

[git]

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

​d4fbe40

_get_shared_lib_filename: Do not assume Python sysconfig paths are in SAGE_LOCAL

[git]

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

​bba08ec	Create module src/sage/env_config.py from src/sage/env_config.py.in, defining SAGE_LOCAL
​d1779cc	Merge branch 't/29022/create_module_src_sage_env_config_py_from_src_sage_env_config_py_in__defining_variables_for_use_in_sage_env' into t/29013/public/29013-use-py3-venv

[mkoeppe]

Replying to mkoeppe:

sage -t --long src/sage/env.py # 1 doctest failed

For this one, I have created #29022.

Fixed.

[mkoeppe]

The numerical problems seems to happen because numpy does not find BLAS.

[mkoeppe]

Actually, it does not find openblas or atlas, but it does find the "Accelerate framework":

accelerate_info:
  FOUND:
    extra_compile_args = ['-msse3', '-I/System/Library/Frameworks/vecLib.framework/Headers']
    extra_link_args = ['-Wl,-framework', '-Wl,Accelerate']
    define_macros = [('NO_ATLAS_INFO', 3), ('HAVE_CBLAS', None)]

  FOUND:
    extra_compile_args = ['-msse3', '-I/System/Library/Frameworks/vecLib.framework/Headers']
    extra_link_args = ['-Wl,-framework', '-Wl,Accelerate']
    define_macros = [('NO_ATLAS_INFO', 3), ('HAVE_CBLAS', None)]

[git]

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

​85147f9	build/pkgs/numpy/lapack_conf.py: Add a [DEFAULT] section to site.cfg
​93bae6f	Merge branch 't/29025/numpy__site_cfg_needs_a__default__section' into t/29013/public/29013-use-py3-venv

[mkoeppe]

Merged #29025, fixes the numerical doctests.

[dimpase]

I see a problem with cryptominisat spkg on this branch (which otherwise builds and passes most tests (not related to this packages) on Gentoo)

sage -t --warn-long 88.1 src/sage/sat/solvers/cryptominisat.py
**********************************************************************
File "src/sage/sat/solvers/cryptominisat.py", line 49, in sage.sat.solvers.cryptominisat.CryptoMiniSat
Failed example:
    solver = CryptoMiniSat()                                  # optional - cryptominisat
Exception raised:
    Traceback (most recent call last):
      File "/mnt/opt/Sage/sage-dev/local/lib/sage/venv/sage/lib/python3.7/site-packages/sage/sat/solvers/cryptominisat.py", line 69, in __init__
        from pycryptosat import Solver
    ModuleNotFoundError: No module named 'pycryptosat'

I wonder whether it uses a convoluted way to create python extensions, which got broken.

[dimpase]

oops, comment:18 belongs to #27824

[embray]

Minor bikeshed, but why the long path "lib/sage/venv/sage"? What about something like just "lib/sage-X.Y"; analogously to "lib/python-X.Y"? We've never had a "lib/sage" before in the first place, so I don't see any immediate need to make the path so much more complicated.

[mkoeppe]

Replying to embray:

Minor bikeshed, but why the long path "lib/sage/venv/sage"? What about something like just "lib/sage-X.Y"; analogously to "lib/python-X.Y"? We've never had a "lib/sage" before in the first place, so I don't see any immediate need to make the path so much more complicated.

1) "venv" has to be in the path to indicate that it's a venv in the technical sense and not a random place with invitation to install stuff in it.

2) Definitely no versioning.

3) I suppose I could drop the final "sage", making it lib/sage/venv

[mkoeppe]

Replying to dimpase:

I see a problem with cryptominisat spkg on this branch (which otherwise builds and passes most tests (not related to this packages) on Gentoo)

sage -t --warn-long 88.1 src/sage/sat/solvers/cryptominisat.py
**********************************************************************
File "src/sage/sat/solvers/cryptominisat.py", line 49, in sage.sat.solvers.cryptominisat.CryptoMiniSat
Failed example:
    solver = CryptoMiniSat()                                  # optional - cryptominisat
Exception raised:
    Traceback (most recent call last):
      File "/mnt/opt/Sage/sage-dev/local/lib/sage/venv/sage/lib/python3.7/site-packages/sage/sat/solvers/cryptominisat.py", line 69, in __init__
        from pycryptosat import Solver
    ModuleNotFoundError: No module named 'pycryptosat'

I wonder whether it uses a convoluted way to create python extensions, which got broken.

Upstream merged pull request ​https://github.com/msoos/cryptominisat/pull/551 to take care of a virtualenv installation.

[mkoeppe]

I haven't tested whether this patch helps for our venv. The patch is not in the current release 5.6.8 (which we have).

[git]

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

​2ed5d07

Add cryptominisat patch for virtual envs

[mkoeppe]

Try with this patch please. I can't compile cryptominisat because of an unrelated reason.

[dimpase]

patch landed into a wrong place, it should be one level down, in patches/

testing now, after correction of this.

[mkoeppe]

Sorry, thanks.

[git]

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

​27d945a

moved patch to the right place

[dimpase]

the patch worked, it's all good with tests now.

[mkoeppe]

Thank you!

[embray]

I might have missed some discussion elsewhere, but I still don't see the advantage of this extra complication over having the sage environment just extend sys.path where necessary. I say this as a big fan of virtualenvs in general, and having suggested doing something like this in the past. But the more I think about it the less I see any advantage in this case over having an additional sys.path entry.

[mkoeppe]

Replying to embray:

I say this as a big fan of virtualenvs in general, and having suggested doing something like this in the past.

Yes, of course this ticket was inspired by various insights and discussions shared by you and others on trac tickets.

But the more I think about it the less I see any advantage in this case over having an additional sys.path entry.

Previously, with our python2 build, there was no clear way to do it. But now with python 3 there is a standard and clean tool: that's venv. Installing into it is supported by standard tools.

[embray]

Or, at the very least, what I've suggested in the past was to make $SAGE_LOCAL itself a virtualenv, as virtualenvs already have a sort-of root filesystem, including among other things lib/pythonX.Y/site-packages.

With this approach, there will now be (depending whether or not the system Python is used) both:

• SAGE_LOCAL/lib/python3.7/site-packages
• SAGE_LOCAL/lib/sage/venv/sage/lib/python3.7/site-packages

If SAGE_LOCAL itself (if it exists) were a virtualenv already that would simplify matters greatly.

[mkoeppe]

Replying to embray:

Or, at the very least, what I've suggested in the past was to make $SAGE_LOCAL itself a virtualenv, as virtualenvs already have a sort-of root filesystem, including among other things lib/pythonX.Y/site-packages.

No, the present ticket explicitly avoids that. The reason is that venv has no documented mechanism for adding non-Python stuff to it, in particular to its activation scripts.

[mkoeppe]

Replying to embray:

With this approach, there will now be (depending whether or not the system Python is used) both:

• SAGE_LOCAL/lib/python3.7/site-packages
• SAGE_LOCAL/lib/sage/venv/sage/lib/python3.7/site-packages

That's correct. In a fresh install, the first one will be empty.

In an install upgraded from earlier versions, there will be packages in both. When there's an update, the old version if removed from the first one and the new version is added to the second one.

[embray]

Replying to mkoeppe:

Replying to embray:

Or, at the very least, what I've suggested in the past was to make $SAGE_LOCAL itself a virtualenv, as virtualenvs already have a sort-of root filesystem, including among other things lib/pythonX.Y/site-packages.

No, the present ticket explicitly avoids that. The reason is that venv has no documented mechanism for adding non-Python stuff to it, in particular to its activation scripts.

There's no reason you can't--I've often install C libraries into a virtualenv by setting the path to it in ./configure --prefix.

[embray]

I really hate to be a pain in the ass because I know you put a lot of work into this, but I really don't get it... It really seems complicated and messy to both activate the Sage environment, and then within that activate a virtualenv. I don't think that's what I ever had in mind when I suggested using a virtualenv for Sage (or if I did, I take it back...)

I'd still like to know what problem this is solving. You say that it's required for #27824, but why? What specific problems related to #27824 is it solving? Because most likely there's another way.

[dimpase]

Well, the separation of #29013 and #27824 is artificial, but otherwise, they do the job. venv removes an extra complication of sagelib being very far from pip-installable.

[mkoeppe]

Replying to embray:

I don't think that's what I ever had in mind when I suggested using a virtualenv for Sage

Don't worry. I'm prepared to take full credit for this specific version of the idea, if necessary.

[embray]

Replying to dimpase:

Well, the separation of #29013 and #27824 is artificial, but otherwise, they do the job. venv removes an extra complication of sagelib being very far from pip-installable.

How? Why? There's really nothing that magical about virtualenvs. It's just an alternate install prefix plus (as of Python 3 and the venv module) some small support bits in the interpreter to report a different sys.prefix, etc.

[mkoeppe]

Replying to embray:

Replying to dimpase:

Well, the separation of #29013 and #27824 is artificial, but otherwise, they do the job. venv removes an extra complication of sagelib being very far from pip-installable.

How? Why? There's really nothing that magical about virtualenvs. It's just an alternate install prefix plus (as of Python 3 and the venv module) some small support bits in the interpreter to report a different sys.prefix, etc.

Part of the magic is to use the standard instead of project-specific tricks.

[embray]

Replying to mkoeppe:

Replying to embray:

Replying to dimpase:

Well, the separation of #29013 and #27824 is artificial, but otherwise, they do the job. venv removes an extra complication of sagelib being very far from pip-installable.

How? Why? There's really nothing that magical about virtualenvs. It's just an alternate install prefix plus (as of Python 3 and the venv module) some small support bits in the interpreter to report a different sys.prefix, etc.

Part of the magic is to use the standard instead of project-specific tricks.

What "tricks" are you avoiding?

[git]

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

​08ec551	build/make/deps: Make python3_venv a .PHONY target
​b2ab6f3	src/bin/sage-env: Simplify the venv path to $SAGE_LOCAL/lib/sage/venv
​7a43544	src/sage/env.py, envconfig.py.in: Add MAXIMA; src/sage/interfaces/maxima.py: Use it
​f2af478	Merge branch 't/29022/create_module_src_sage_env_config_py_from_src_sage_env_config_py_in__defining_variables_for_use_in_sage_env' into t/29013/public/29013-use-py3-venv

[mkoeppe]

Replying to mkoeppe:

Replying to embray:

Minor bikeshed, but why the long path "lib/sage/venv/sage"? ....

3) I suppose I could drop the final "sage", making it lib/sage/venv

Done.

[mkoeppe]

I've also merged in a few changes from #29032 (Erik's version of python spkg-configure / venv), and the current #29022.

[git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

​6b28ac1	In a python3 build, install all Python packages into a venv
​981f65e	sage-pip-install: Use PYTHON=sage-python23
​a6854bd	_get_shared_lib_filename: Do not assume Python sysconfig paths are in SAGE_LOCAL
​11ada01	build/pkgs/numpy/lapack_conf.py: Add a [DEFAULT] section to site.cfg
​3bd10e4	Add cryptominisat patch for virtual envs
​da13a9e	moved patch to the right place
​c9d067a	src/bin/sage-env: Simplify the venv path to $SAGE_LOCAL/lib/sage/venv
​199a72e	src/sage/env.py: Rework _get_shared_lib_filename, fix for Cygwin: it would privilege system DLLs over DLLs in SAGE_LOCAL

[embray]

I would be interested in your comments on #29032. By making $SAGE_LOCAL itself into a virtualenv--which it already essentially is in all other respects--no other weird hacks are needed.

Other than some things like the OSX fixes I don't see what this solves that my branch doesn't.

[git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

​980f444	In a python3 build, install all Python packages into a venv
​90bf090	sage-pip-install: Use PYTHON=sage-python23
​9edf068	_get_shared_lib_filename: Do not assume Python sysconfig paths are in SAGE_LOCAL
​ead2d0f	Add cryptominisat patch for virtual envs
​56e3daa	src/bin/sage-env: Simplify the venv path to $SAGE_LOCAL/lib/sage/venv
​208c7e6	src/sage/env.py: Rework _get_shared_lib_filename, fix for Cygwin: it would privilege system DLLs over DLLs in SAGE_LOCAL