[Avocado-devel] [RFC] Environment Variables

[Ademar Reis]

On Wed, Jun 01, 2016 at 04:02:54PM -0300, Cleber Rosa wrote:
> On 06/01/2016 03:07 PM, Ademar Reis wrote:
> > On Tue, May 31, 2016 at 07:30:43AM -0300, Cleber Rosa wrote:
> > > 
> > 
> > I'm replying on top of Cleber because he already said a few
> > things I was going to say.
> > 
> > > On 05/25/2016 05:31 AM, Amador Pahim wrote:
> > > > Hi folks,
> > > > 
> > > > We have requests to handle the environment variables that we can set to
> > > > the tests. This is the RFC in that regard, with a summary of the ideas
> > > > already exposed in the original request and some additional planning.
> > > > 
> > > > The original request is here:
> > > > https://trello.com/c/Ddcly0oG/312-mechanism-to-provide-environment-variables-to-tests-run-on-a-virtual-machine-remote
> > > > 
> > > > 
> > > > Motivation
> > > > ==========
> > > > Avocado tests are executed in a fork process or even in a remote
> > > > machine. Regardless the fact that Avocado is hard coded to set some
> > > > environment variables, they are for internal consumption and user is not
> > > > allowed to control/configure its behavior.
> > > 
> > > You mean this:
> > > 
> > > http://avocado-framework.readthedocs.io/en/latest/WritingTests.html#environment-variables-for-simple-tests
> > > 
> > > Right? Basically, the fact that Avocado sets some of the job/test state as
> > > environment variables, that can be used by SIMPLE tests.
> > > 
> > > > The motivation is the request to provide users an interface to set
> > > > and/or keep environment variables for test consumption.
> > 
> > I'm not sure if they're necessarily for test consumption. I think
> > the motivation for the original request was to provide the
> > standard Unix interface of environment variables for when tests
> > are run remotely.
> > 
> 
> If the motivation is basically about setting the env vars when running tests
> remotely, than this brings the discussion about the *local* behavior to:
> 
> 1. Should Avocado default to the standard UNIX behavior of cloning the
> environment?
> 
>  A: IMHO, yes.

That's the current behavior (see my example at the end of the
previous email). Except when one runs tests remotely, which is
precisely the use case this feature would "fix".

> 
> 2. Could Avocado have have a feature to start tests in a clean(er)
> environment?
> 
>  A: Possibly yes, but seems low priority.  The use case here could be seen
> as a plus in predictability, helping to achieve expected test results in
> spite of the runner environment.  A real world example could be a CI
> environment that sets a VERBOSE environment variable. This env var will be
> passed over to Avocado, to the test process and finally to a custom binary
> (say a benchmark tool) that will produce different output depending on that
> environment variable.  Doing that type of cleaning in the test code is
> possible, but the framework could help with that.
> 
> 2.1. If Avocado provides a "clean(er) test environment" feature, how to
> determine which environment variables are passed along?
> 
>  A: The "env-keep" approach seems like the obvious way to do it.  If the
> mechanism is enabled, which I believe should be disabled by default (see
> #1), its default list could contain the more or less standard UNIX
> environment variables (TERM, SHELL, LANG, etc).

Agree. But like you said such a feature would be low priority and
optional. The important thing is that the implementation of what
we're discussing in this RFC would not interfere with it.

> 
> > These environment variables can change the behavior of both
> > Avocado (the runner itself), the tests (after all nothing
> > prevents the test writer from using them) and all sub-processes
> > executed by the test.
> > 
> 
> Right.
> 
> > Locally, this is standard:
> > 
> >   $ TMPDIR=/whatever/tmp VAR=foo ./avocado run test1.py
> > 
> > But when running avocado remotely, there's no way to configure
> > the environment in the destination. The environment variables set
> > in the command line below will not be "forwarded" to the remote
> > environment:
> > 
> >   $ TMPDIR=/whatever/tmp VAR=foo ./avocado run test1.py \
> >                                  --remote...
> > 
> 
> Right.
> 
> > > > 
> > > > Use cases
> > > > =========
> > > > 1) Use the command line or the config file to set the environment
> > > > variables in tests processes environment; access those variables from
> > > > inside the test.
> > > > 2) Copy from current environment some environment variable(s) to the
> > > > tests processes environment; access those variables from inside the test.
> > 
> > I think we don't even have to go that far. We can simply say the
> > intention is to set the environment variables in the environment
> > where Avocado is run. The mechanism is quite standard and well
> > understood.
> > 
> > And here comes an important point: I don't think this should be a
> > mechanism to pass variables to tests. Although, again,
> > environment variables can be used for that purpose, Avocado
> > should have a proper interface to provide a dictionary of
> > configuration and variables to each test.
> > 
> 
> The only valid reason for having such a mechanism to pass *different*
> environment variables to tests, talking about local environment, would be
> *if and only if* the same environment variable to be set when running
> Avocado would change the behavior of Avocado itself.  Example:
> 
>  $ AVOCADO_LOG_EARLY=1 avocado run avocado-self-tests.py
> 
> This way, both the first level avocado process (our "real" runner) and other
> instances run by the "avocado-self-test.py" code would react to that
> variable. *BUT* this seems a corner case, and I wouldn't think it justifies
> the implementation of such a feature at this point.

I think we should KISS: environment variables are well understood
in the Unix world. This feature will simply guarantee they (or a
subset of them) are kept when tests are run in a different
environment (remotely, in a VM, in a container -- or whatever
else we implement in the future).

> 
> > Currently, this is erroneously provided by the multiplexer
> > (including --mux-inject), but it should be cleaned up in the near
> > future.
> > 
> 
> Right.
> 
> > > > 
> > > > Proposal
> > > > ========
> > > > - To create a command line option, under the `run` command, to set
> > > > environment variables that will be available in tests environment process:
> > > > 
> > > >  $ avocado run --test-env='FOO=BAR,FOO1=BAR1' passtest.py
> > > > 
> > > 
> > > I can relate to this use case...
> > 
> > This would be a simple way of doing it, but something like
> > "--env-keep=FOO,FOO1" could be a better approach (more about it
> > below). So one would write it this way:
> > 
> >   $ FOO=BAR FOO1=BAR1 avocado run --env-keep='FOO,FOO1' passtest.py
> > 
> 
> Right. Again, the only drawback of --env-keep is something along the lines
> of the "AVOCADO_LOG_EARLY" example I gave earlier.
> 
> > > 
> > > > - To create an option in config file with a dictionary of environment
> > > > variables to set in test process environment. It can be used as a
> > > > replacement or complement to the command line option (with lower priority):
> > > > 
> > > >  [tests.env]
> > > >  test_env_vars = {'FOO': 'BAR', 'FOO1': 'BAR1'}
> > > > 
> > > 
> > > ... while putting those in a config file does not seem like something one
> > > would do.
> > > 
> > > In all cases, and more explicitly in the config file example, this is only
> > > really necessary if/when the environment variable to pass to the test
> > > actually harms Avocado (considering a local execution, that is, in a forked
> > > process).
> > > 
> > > So, if Avocado and the test, share the use of environment variables by the
> > > same name, then this is a must.  Also in the case of execution in other
> > > "runners", such as remote/vm, this can be quite valuable.
> > 
> > I don't like this interface because I think this opens the door
> > for abuse. Like I said in a previous paragraph, this interface
> > should not be a mechanism for passing variables to tests.
> > 
> 
> I can also see it being abused very quickly.
> 
> > > 
> > > > - Create an option in config file with a list of environment variable
> > > > names to copy from avocado main process environment to the test process
> > > > environment (similar to env_keep in the /etc/sudoers file):
> > > > 
> > > >  [tests.env]
> > > >  env_keep = ['FOO', 'FOO1', 'FOO2']
> > > > 
> > 
> > I like this approach because it reinforces the message that we're
> > keeping (or forwarding) some of the environment variables from
> > the original environment where the test runner was run.
> > 
> 
> Then, we need a "env_reset" mentality, and a default list of environment
> variables that we forward by default (along the same lines of sudo):
> 
> Defaults    env_reset
> Defaults    env_keep =  "COLORS DISPLAY HOSTNAME HISTSIZE INPUTRC KDEDIR
> LS_COLORS"
> Defaults    env_keep += "MAIL PS1 PS2 QTDIR USERNAME LANG LC_ADDRESS
> LC_CTYPE"
> Defaults    env_keep += "LC_COLLATE LC_IDENTIFICATION LC_MEASUREMENT
> LC_MESSAGES"
> Defaults    env_keep += "LC_MONETARY LC_NAME LC_NUMERIC LC_PAPER
> LC_TELEPHONE"
> Defaults    env_keep += "LC_TIME LC_ALL LANGUAGE LINGUAS _XKB_CHARSET
> XAUTHORITY"
> 
> Like I briefly mentioned before.

That would be useful only for the use-case of resetting the
environment where *tests* are run.

As of today, there's no difference between the environment where
a test is run and the environment where the test runner is run.
They share the same variables (I like that).

The feature we're discussing would forward (keep, clone, copy,
whatever you want to call it) some of the env variables from the
env where the *original runner* is running to the env where the
*remote runner* will run. It says nothing about the env where the
actual *tests* are run. It also doesn't create any variables. It
simply forward them if they already exist in the environment.

So env_reset is not necessary. If we ever implement it, it'll
have no impact in this env-keep feature, because they'll be
working on different layers (test runner vs test).

> 
> > > > 
> > > 
> > > Right, this makes sense. But it also brings the point that we may actually
> > > change the default behavior of keeping environment variables from Avocado in
> > > the tests' process.  That is, they would get a much cleaner environment by
> > > default.  While this sounds cleaner, it may break a lot of expectations.
> > 
> > I wonder what the motivation would be to clean-up the environment
> > where tests are run. Can you please elaborate?  If we indeed
> > decide to implement this change, then I would say we should honor
> > whatever is set in env_keep.
> > 
> 
> I mentioned before a few points about test predictability, but the other
> points that came later also support this idea.
> 
> Again, if env-keep is applicable for tests running remotely *only*, this
> changes a lot of things.  If we're talking about both local and remote, than
> all what I said before applies.

There's no "local" and "remotely", because local already works.
So this feature is about extending the behavior we have locally
to remote use-cases.

> 
> > > 
> > > > For every configuration entry point, the setting have to be respected in
> > > > local and remote executions.
> > > > 
> > > > Drawbacks
> > > > =========
> > > > 
> > > > While setting an environment variable, user will be able to change the
> > > > behavior of a test and probably the behavior of Avocado itself. Maybe
> > > > even the OS behavior as well. We should:
> > > > - Warn users about the danger when using such options.
> > > 
> > > I fail to see where an environment variable, to be set by Avocado in the
> > > test process, can or should impact Avocado itself.  If it does, then we'd
> > > probably be doing something wrong.  I'm not sure we need warnings that
> > > exceed documenting the intended behavior.
> > 
> > I think the environment has to be set where Avocado is run, not
> > where tests are run. Which is why I prefer --env-keep and
> > [env-keep].
> > 
> > So in the case of:
> > 
> >   $ FOO=bla avocado run test.py --remote=...
> > 
> > $FOO is available inside the environment where Avocado and its
> > tests are run, both locally and remotely.
> > 
> 
> Agreed, but let's just keep in mind that one (very much) corner case (where
> the environment may affect Avocado itself).
> 
> > > 
> > > > - Protect Avocado environment variables from overwriting.
> > > 
> > > About protecting the Avocado's own environment variables: agreed.
> > 
> > This is something that won't need any change. There are variables
> > which are read by Avocado and others which are written by it.
> > 
> > For example:
> > 
> >  * TMPDIR will influence Avocado's behavior (standard Unix
> >    variable)
> >  * AVOCADO_VERSION is written by Avocado. Setting it externally
> >    won't make any difference.
> > 
> >   $ TMPDIR=/home/ademar/tmp avocado run examples/tests/env_variables.sh --show-job-log | grep Temporary
> >   Temporary dir: /home/ademar/tmp/avocado_L9YiE4
> > 
> >   $ AVOCADO_VERSION=0 ./scripts/avocado run examples/tests/env_variables.sh --show-job-log | grep Version
> >   [stdout] Avocado Version: 35.0
> > 
> 
> Exactly.  Now I think I missed that fact that Avocado's own environment
> variables are indeed already "protected", because they're set at when the
> SIMPLE tests' processes are created, which happens later and thus takes
> precedence.

Yes, that's why we don't have to worry too much about the
details. The way environment variables work is quite standard and
well understood.

> 
> > > 
> > > > 
> > > > Looking forward to read your comments.
> > > > 
> > > 
> > > Overall, this is definitely welcome.  Let's discuss possible implementation
> > > issues, such as remote/vm support, because it wouldn't be nice to introduce
> > > something like this with too many caveats.
> > > 
> > 
> > We appear to have different understandings about what this
> > feature should be about. IMO it should be about the standard unix
> > environment where Avocado and tests are run. The primary use-case
> > is for remote/vm support (in other words, whenever there's a
> > change in the environment).
> > 
> 
> Then we should make it clear what we want to tackle first.  If remote/vm
> support is the primary goal, then maybe we should think about
> --remote-env/--vm-env options to not poison the discussion and
> implementation with the env_reset/env_keep like features.
> 

If we think env_reset is a different use-case (and based on your
use case #2 from the beginning of this e-mail looks like we do),
then indeed --remote-keep-env and --vm-keep-env are the options
we're looking for (with respective configuration file entries).

I was hoping we could implement just one "--keep-env=" instead of
multiple "--<something-remote>-keep-env=" (--remote-keep-env=,
--vm-keep-env=, --docker-keep-env=, etc), but maybe that's not a
good idea.

Thanks.
   - Ademar

-- 
Ademar Reis
Red Hat

^[:wq!