Meta-ticket: Split sage-env into 5 to clean up sage configuration

[mkoeppe]

src/bin/sage-env (with generated configuration in src/bin/sage-env-config; installed in $SAGE_LOCAL/bin) is used in all of the following contexts:

1. Sage-the-distribution while building spkgs,
2. Sage-the-distribution for building sagelib,
3. for environment variables needed to start Sage,
4. as variables by Sage at runtime (through sage.env), and
5. for environment variables affecting subprocesses of Python invoked by sagelib modules.

This ticket proposes to split the configuration according to these 5 contexts. This will make the installation of Sage more modular and flexible.

1./2. Create build/bin/sage-build-env(with generated configuration in build/bin/sage-build-env-config) for the build-time environment variables for sage-the-distribution (spkg and sagelib). It is not installed in $SAGE_LOCAL.

2. In build/make/Makefile.in and the scripts generated by build/bin/sage-spkg, source build/bin/sage-build-env in addition to src/bin/sage-env. This is #29052.
3. Gradually, we will move settings from src/bin/sage-env[-config] to build/bin/sage-build-env[-config] that are known to be only needed for 1./2. (Note that not all compiler-related environment variables can be moved exclusively to 1./2. – some need to be added to 5 as well for some some calls to compilers needed at runtime to support things like %cython (?) and sage.calculus.desolvers.desolve_mintides.) As well as users' use of pip to install additional Python packages!
4. Eventually, we will remove the call to src/bin/sage-env from the build scripts.

2./3. Obtain sagelib's build-time configuration such as src/setup.py's library_dirs and runtime information (sage.env) from sage-config (introduced in #29038).

1. In src/Makefile, poison the SAGE_LOCAL variable.
   • #29411: make sagelib a script package
   • #29779: pkg-config installed from SPKG pkgconf should not depend on environment variable SAGE_LOCAL
2. Pass SAGE_PKGS (if still needed at all - see #28815 or #29705) via sage-config instead of from src/Makefile.

3./4./5. Using sage_conf (Python module and script sage-config, #29038), make sage.all fully functional when imported from a Python, without setting any environment variables (sage-env).

1. Initially, phase out src/bin/sage-env-config by using sage-config instead.
   • #29384: Clean up src/bin/sage-env-config.in: Move logic to src/bin/sage-env, move non-environment configuration variables to sage_conf.py
   • #29825: Clean-up for src/bin/sage-env
2. Set sage.env variables via the Python module sage_conf instead of relying on information from environment variables. Example: CYSIGNALS_CRASH_DAYS
   • #34236 sage.env: Give values from sage_conf precedence over environment variables
3. Set environment variables that are needed only by subprocesses invoked by sagelib in the environment of these subprocesses, rather than relying on them begin set in sage-env. This environment could be provided by a variable sage_conf.SAGE_SUBPROCESS_ENV and/or more invidual variables. Example: R_PROFILE
   • #30563: Use configuration variable MAXIMA instead of hardcoding "maxima"
   • #30818 Set environment for subprocesses invoked by Sage
4. Finally, remove use of sage-env for all purposes except for sage -sh.

Aspects of downstream sage packaging:

• The goal is to enable downstream sage packagers to use a whole unmodified src directory to build and install sagelib.
  • more precisely, to use the pip-installable source package created by build/pkgs/sagelib/spkg-src (#29950); see also #30036 (pip-installable sage)
• Downstream packagers could provide their own "implementation" (one Python module ​sage_conf.py) of the configuration module defined by #29038, instead of patching sources. They would not necessarily use the reference implementation in build/pkgs/sage_conf, which is intended for use with sage-the-distribution.

Binary packages:

• #31417 version of package sage_conf for relocatable binary distributions

pip-installable Sage

• #29039 pip-installable version of package sage_conf - installs non-Python bits of the Sage distribution in ~/.sage/
• #31396 relocatable wheel version of package sage_conf

More related tickets:

• cleaning of src/bin as described in #21570, #21559.
• #25486: Discover SAGE_SCRIPTS_DIR to make $SAGE_LOCAL/bin/sage work from any directory, in an environment without SAGE_* variables - this adds a SAGE_ROOT variable to sage-env-config
• #30724: src/bin/sage-env, src/bin/sage-env-config.in: Remove PYTHON_FOR_VENV
• #29133 META-META-TICKET: Build, packaging, testing improvements

[fbissey]

One comment. Currently sage_setup is installed because it needs to be installed to be doctested - like the rest of sage. But nothing in sage_setup is really needed at runtime and distro can choose not to install it. I don't in sage-on-gentoo. I tend to push for thing not needed at runtime to be moved in there. And that's where I would have put sage-build-env but I am not going to fight it going into build/bin although I'd like everything used by setup.py to be neatly under src.

[mkoeppe]

Thanks - I've updated the description

[embray]

I agree.

Consider a slight alternative: Rewrite sage-env as a Python script. This would give it much more flexibility and probably more legibility, including some command-line options for different environments (--build, especially).

The output would of course be the shell commands to actually set the variables. So rather than source sage-env one would eval $(sage-env). My one concern is that it would be slower, but ideally it's something that shouldn't happen frequently anyways, but rather only when not already in sage-env.

[fbissey]

I have a tendency to mix up sage-env the script and sage.env.py, my comment is really an example of that. However a lot is common between the two.

[embray]

Right--they could be one in the same.

[fbissey]

Replying to embray:

Right--they could be one in the same.

I have thought about it often. Of course that really means that you'll have to removing all the build time stuff. But in sage-on-gentoo where I ship a very simplified sage-env (in /etc as well) they are virtually doing the same thing, one in shell, the other in python.

[embray]

In either case--whether making a single script, or two separate shell scripts, this would be useful for gentoo (and probably other distros as well) then right?

[fbissey]

Probably, the current sage-env is a forest mostly dedicated to the build system. I just discard it and bring my own, it takes less space and maintenance than a mega patch.

[mkoeppe]

Ideally, the configuration for sagelib runtime should take place in a generated Python file and not involve environment variables at all.

[embray]

More likely parts of it would be static, and other parts could be generated, but yes, what I meant above in terms of rewriting sage-env in Python was that it would not (by default) involve environment variables at all.

But it's still useful for interacting with other tools, especially for build purposes, as well as dropping into ./sage -sh to be able to set environment variables, which is why I suggest a mode to output a list of variables that can be read from the shell.

[mkoeppe]

New commits:

​7405dd5	Move sage-dist-helpers from src/bin to build/bin
​9189d60	Split out build/bin/sage-build-env-config from sage-env-config

[embray]

As-is this is going to conflict with tickets like #27825 that are adding still more environment variables needed at build-time to sage-env-config.

I'm completely in favor of this separation in principle though. The question is whether we should do this first and then update tickets like #27825 on top of it, or the other way around?

[dimpase]

I really would like to first get our existing tickets (#27825, #27822) touching sage-env-config in.

[mkoeppe]

The tickets would be trivial to rebase on top of this one.

[mkoeppe]

Probably one should introduce a command sage --buildsh, which would provide the larger environment.

[embray]

For the sake of a bit more terseness I wonder if we could call it just sage-env-build.in and sage-env-build. The only reason for the -config in the other one is to distinguish it from plain sage-env.

Now that I look I'm not even entirely sure we need to separate sage-env-config from sage-env.

I think we could probably have a sage-env.in produce sage-env and get rid of the separate sage-env-config entirely, though it's possible there's some subtlety to that that I'm missing.

[embray]

Went ahead and rebased on current develop so it could incorporate all the additional variables we've since added.

I didn't change the name of the script yet; I would still prefer to shorten it but if anyone disagrees we can leave it as-is.

---

New commits:

​a34a9f1	Move sage-dist-helpers from src/bin to build/bin
​ff07121	Split out build/bin/sage-build-env-config from sage-env-config

[fbissey]

The only subtility to sage-env-config in my mind, is that you can touch it or even regenerate it without touching sage-env. But I am sure you don't need a separate file to achieve your objectives in either case.

[embray]

No, probably not. I might try it, though we'll leave that as a separate exercise from this ticket.

[embray]

Moving tickets from the Sage 8.8 milestone that have been actively worked on in the last six months to the next release milestone (optimistically).

[embray]

Ticket retargeted after milestone closed

[mkoeppe]

The run-time environment variables of sage could as well be configured by a Python module src/sage/env_config.py - see #29022.

[mkoeppe]

Replying to mkoeppe:

Probably one should introduce a command sage --buildsh, which would provide the larger environment.

As an extension of this, there could be sage --buildsh SPKG, which would set up the environment in which spkg-install runs. This could be useful for testing new versions of packages.

[mkoeppe]

Replying to mkoeppe:

The run-time environment variables of sage could as well be configured by a Python module src/sage/env_config.py - see #29022.

Or #29038 - Python package sage_conf: Provides optional configuration information for sagelib

[mkoeppe]

Hoping for input from people involved in packaging.

[gh-tobiasdiez]

Not sure if this is the right place, but I would suggest to use the refactoring process to also convert the environment variables to configuration files.

The idea would be to have a toml file (or yaml or whatever you prefer) in some predefined folder (say src for the start). This file could be generated e.g by the configure script. This config file is then parsed and wrapped by a python class (natural candidate would be the current sage_conf.py). The python class also provides reasonable default values in case the config file doesn't exist.

This has several advantages in my opinion:

• More transparent, as one sees on one glance which variables influence sage.
• Easily editable by the user/developer.
• Local (e.g. think about multiple virtual environments with different configs).
• Type-safe, i.e, it is ensured when parsing the toml file that the types of the config variables is correct.
• Easier for downstream package managers, since they only need to provide a proper config file.
• (Higher security as one cannot inject arbitrary python code, which would be the case if sage_conf.py should be edited.)

What do you think?

[mkoeppe]

sage_conf defines an interface already - and the plan is indeed to get rid of dependencies on environment variables using this interface.

How sage_conf obtains the configuration is orthogonal to the defined interface.

I am not interested in defining configuration files at the moment. There is just too much potential on wasting energy on discussions regarding the "best" config file format.

[gh-tobiasdiez]

Well, you are actually proposing to add a configuration file in the form of a python file (sage_conf.py). One could easily bypass an extensive discussion and follow the arguments of ​https://www.python.org/dev/peps/pep-0518/#overview-of-file-formats-considered.

But I agree that my suggestion concerning a config file is somewhat out of line with the goal of this ticket. I would still suggest to keep this in mind while designing the interface for sage_conf, so that it might be easy in the future to switch to a design based on a config file. In particular, I would replace the module wide variables (like VERSION) with methods in a config class (like Config.get_version) that can also provide suitable defaults and do slight post-processing.

[mkoeppe]

Replying to gh-tobiasdiez:

Well, you are actually proposing to add a configuration file in the form of a python file (sage_conf.py).

Note that sage_conf.py is already in use since Sage 9.1

[mkoeppe]

Replying to gh-tobiasdiez:

I would still suggest to keep this in mind while designing the interface for sage_conf, so that it might be easy in the future to switch to a design based on a config file. In particular, I would replace the module wide variables (like VERSION) with methods in a config class (like Config.get_version) that can also provide suitable defaults and do slight post-processing.

All of this discussion is certainly valid and I am sure that you (and others) have valuable insights there, but now is simply not the right time for it. To get the design right, one needs to know about objectives and constraints. We are in the process of expanding the ways of Sage installation/deployment into several directions, including pip-installability and modularization; and also binary packaging will need an overhaul.

[mkoeppe]

Setting new milestone based on a cursory review of ticket status, priority, and last modification date.

[mkoeppe]

Setting a new milestone for this ticket based on a cursory review.

[mkoeppe]

From the discussion in #32904 regarding sage_conf at build time (with pip's build isolation) vs. runtime:

• pip specifies "ignore-installed" while constructing the build environment and thus will build sage-conf twice. ​https://github.com/pypa/pip/blob/main/src/pip/_internal/build_env.py#L221
• The PyPI version sage_conf (from #29039) keys the SAGE_ROOT/SAGE_LOCAL that it creates in ~/.sage to the version of sage_conf
• So if the sage_conf version is not completely pinned in the sagemath-standard build and install requirements, then different versions (and thus SAGE_ROOTs) may end up being used at build and install of sagemath-standard

[mkoeppe]

From ​https://trac.sagemath.org/ticket/32904#comment:24:

• Split sage-conf into build and runtime configuration. There is still an overlap (and thus duplication) but at built time sage needs less dependencies than at runtime. In addition, maybe directly incorporate the build-time sage-conf in the sagemath package.

[mkoeppe]

Replying to mkoeppe:

From gh-tobiasdiez in ​https://trac.sagemath.org/ticket/32904#comment:24:

• Split sage-conf into build and runtime configuration. There is still an overlap (and thus duplication) but at build time sage needs less dependencies than at runtime. In addition, maybe directly incorporate the build-time sage-conf in the sagemath package.

I think this is an interesting direction