cannot cache functions which embed a relative import

Summary

Caching does not appear to work on functions which, at call time, perform a relative import.

Steps to reproduce

This is a minimal reproducible which gives the same error message I see in my code.

First, set up the troublesome code in a new Python module called “spam” with a submodule “spam.eggs”:

% mkdir spam
  then create spam/__init__.py and spam/eggs.py with the following content
% cat spam/__init__.py

def load():
  from . import eggs

% cat spam/eggs.py
# nothing

Second, create the following streamlit program - note that load() returns None!

import streamlit as st
import spam

@st.cache
def load():
    return spam.load()

Expected behavior:

I expected to load() to a cached value of None.

Actual behavior:

streamlit fails at the cache step with the following:

ValueError: Empty module name

Streamlit encountered an error while caching the body of load(). This is likely due to a bug in /Users/dalke/cvses/chemfp_demo/spam/__init__.py near line 2:

from . import eggs

Please modify the code above to address this.

If you think this is actually a Streamlit bug, you may file a bug report here.
Traceback:

File "/Users/dalke/demo/show_problem.py", line 5, in <module>
    def load():
File "/Users/dalke/local/lib/python3.9/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1027, in _gcd_import
File "<frozen importlib._bootstrap>", line 961, in _sanity_check

Is this a regression?

This is the first day I’ve used streamlit. I have no experience to know this.

Debug info

• Streamlit version: Streamlit, version 0.76.0
• Python version: Python 3.9.1
• Using Conda? PipEnv? PyEnv? Pex? No. Straight pip.
• OS version: macOS 10.14.6
• Browser version: Firefox 84.0.2

Additional information

None.

[adalke]

For those with the same issue, I was able to work around it with:

def load():
    spam = eval("import spam")
    return spam.load()

[adalke]

Not a problem.

Turns out I couldn’t use cache anyway - my underlying system uses a mmap’ed file, which can’t be pickled. The mmap was designed for a fast startup, which should be a natural fit to streamlit. However, it looks like Firefox is taking up most of the memory on my laptop, so when streamlit restarts, the mapped pages are flushed, and I’m limited by how long it takes my laptop to re-read 2GB from mass storage. Which means that cache wouldn’t help as the pickle overhead is higher than mmap read.

In case it helps with the user story, the underlying “load” command in my system has its own gzip reader which is faster than Python’s built-in gzip library. This usually isn’t needed, so rather than import it eagerly when the top-level package is imported, it’s imported only when the input filename ends with “.gz”.