[Scons-users] SCons under virtualenv

[Paweł Tomulik]

Hi all,

I'd like to discuss some issues related to SCons and virtualenv.

Currently SCons uses hard-coded PATH to search for executables. Because
of this, when it gets executed from virtualenv, it still prefers system
executables over these from virtualenv. So, for example, when we try to
execute python interpreter from SConscript running in virtualenv, the
execution jumps out of virtualenv, leading to possible confusions
(another version of python gets called, some packages installed
purposely in virtualenv are not found by scons tools, etc.). This kind
of issue seems to be currently observed when testing SCons on travis-ci.
Travis uses virtualenv, and some scripts from SCons end-to-end tests get
accidentally executed by python2, despite the testsuite was started in
python3 (runtest.py is started in python3, but some script snippets get
passed by scons to non-virtualenv python2). This masks some regressions,
which would be revealed under python3.

My idea was to change SCons behaviour, such that it would look into
virtualenv in the first place, and then search through all the hardcoded
directories. The implementation would simply prepend to
env['ENV']['PATH'] the search path from virtualenv, when a virtualenv is
in use. In all cases I've seen, it's just about prepending a single
directory, like '/home/user/.local/share/virtualenvs/<random>/bin', so
final PATH is like
'/home/user/.local/share/virtualenvs/<random>/bin:/usr/bin:/opt/bin:/usr/local/bin'.

This approach breaks the rule of predictable and repeatable SCons
behaviour defined as "the search PATH is same on every run, on every
variant of given platform, except you change this explicitly". My
question is, if it would be reasonable to redefine this repeatability a
little bit, such that it would take into account virtualenvs?

There is a pull request, which implements the change. Currently it makes
the change by default, but this can be disabled with --ignore-virtualenv
flag (or SCONS_IGNORE_VIRTUALENV environment variable).

As William pointed out, this is still too risky to introduce this change
in next release, without some transition period. What others think? May
it have such a huge impact on users? What are your opinions?


PR: https://github.com/SCons/scons/pull/3216


-- 
Pawel Tomulik