Client-side app-less sandboxes

[erikbern]

Implements the basics of this. Demo:

>>> from modal import Sandbox
>>> s = Sandbox.create("/bin/bash", "-c", "echo hello world")
>>> for line in s.stdout:
...     print(line)
... 
hello world

>>> s.poll()
0

TODO for this PR:

• Support apps optionally
• Add tests
• Add tests for custom image and secrets

Will do in subsequent PRs:

• Update docstrings to use new syntax
• Add deprecation warnings for old syntax

[aksh-at]

LGTM! Should we add a deprecation message for the old code path?

[aksh-at]

Are we able to resolve all deps without an app ID?

[erikbern]

yeah should be fine – it's just images that need it and i fixed that in the backend

[erikbern]

actually it's possible secrets need it too, let me check

[aksh-at]

Ah looks like we still need to add the non-experimental scheduler param, now that it's out. Outside the scope of this PR.

[aksh-at]

Should we make the following args keyword-only?

[erikbern]

yeah good point

[erikbern]

actually believe they are keyword-only because of the * automatically (i just learned something new about python!)

[aksh-at]

oh cool, makes sense! (any args would be spread into entrypoint_args so it has to be this way)

[erikbern]

Ok slightly bad news: in order for this to work, we need to support app-less secrets and mounts as well.

The good news is that this is something I've been hoping to add support for anyway for a while. But it will take a bit of time to figure out how to do it, unfortunately.

[erikbern]

LGTM! Should we add a deprecation message for the old code path?

Was planning to do that in a subsequent PR (I want to update some integration tests and docs first)

[erikbern]

This works now, but lack of output is a bit confusing with custom images.

>>> image = Image.debian_slim().pip_install("numpy").run_commands("sleep 30")
>>> s = Sandbox.create("/bin/bash", "-c", "echo hello world", image=image)  # This blocks for >30s
>>> s.poll()
0
>>> [line for line in s.stdout]
['hello world\n']

[aksh-at]

This works now, but lack of output is a bit confusing with custom images.

Ah yeah, I think we used to be able to see output in the app logs, but decoupling makes that difficult 🤔

[erikbern]

Ah yeah, I think we used to be able to see output in the app logs, but decoupling makes that difficult 🤔

I think we can fix this in a different way. I've had it on my backlog for a long time to break out the output management into a singleton (separate from the resolver) that you can activate using a context manager (with the idea that all CLI commands enable this automatically). Something like

with modal.enable_logs():
    ...

I don't love singletons but in this case I think it's fine since stdout is already a singleton

[erikbern]

ok this FINALLY works

>>> from modal import Image, Secret, Mount, Sandbox
>>> secret = Secret.from_dict({"FOO": "bar"})
>>> s = Sandbox.create("/bin/bash", "-c", "echo $FOO", secrets=[secret], environment_name="main")
>>> s.poll()
0
>>> [line for line in s.stdout]
['bar\n']