Replace custom build system with Poetry

[eecavanna]

Background

Currently, the Runtime uses a custom dependency management solution that works something like this:

1. Update a requirements/*.in file: main.in for production dependencies and dev.in for development dependencies
2. Run the command $ make update-deps, which — under the hood — runs these commands, which update the main.txt and dev.txt files
   

   nmdc-runtime/Makefile

   Lines 6 to 16 in f0e8096

   	update-deps:
   	# --allow-unsafe pins packages considered unsafe: distribute, pip, setuptools.
   	pip install --upgrade pip-tools pip setuptools
   	pip-compile --upgrade --build-isolation \
   	--allow-unsafe --resolver=backtracking --strip-extras \
   	--output-file requirements/main.txt \
   	requirements/main.in
   	pip-compile --allow-unsafe --upgrade --build-isolation \
   	--allow-unsafe --resolver=backtracking --strip-extras \
   	--output-file requirements/dev.txt \
   	requirements/dev.in

3. Run the command $ make init, which — under the hood — runs these commands, which install the packages listed in main.txt and dev.txt into the current Python environment

Pros:

• It is already implemented
• (Add yours here...)

Cons:

• People sometimes forget step 2
• (Add yours here...)

Proposal

Switch to using Poetry.

If the Runtime used Poetry, the above process would become:

1. Update the pyproject.toml file (there are different sections of the file for production versus development dependencies)
2. Run $ poetry install to install the packages listed in the file — this will generate a poetry.lock file if one doesn't already exist (commit this file to the repo once it exists)

[dwinston]

Sure, let's pursue this. The main wrinkle is reproducible Docker builds, which may just mean adding a

poetry export -f requirements.txt -o requirements.txt

step (https://pythonspeed.com/articles/pipenv-docker/)?

some more context that informed my initial decision wrt pip-tools:

• https://pythonspeed.com/articles/poetry-vs-docker-caching/
• https://hynek.me/articles/python-app-deps-2018/

[eecavanna]

OK, thanks.

I was under the impression having the poetry.lock file in the repo (and not having or using a requirements.txt file anywhere, at any time) would suffice, as long as dependencies are installed via poetry install. The docs for poetry install (source) say:

The install command reads the pyproject.toml file from the current project, resolves the dependencies, and installs them.

If there is a poetry.lock file in the current directory, it will use the exact versions from there instead of resolving them. This ensures that everyone using the library will get the same versions of the dependencies.

Based on that, I don't think a requirements.txt file is necessary for reproducibility, specifically.

I briefly skimmed the first link in your message and saw that the introduction of a requirements.txt file may be to enable us to benefit more from cacheing in order to speed up container image builds. I am in favor of speeding up container image builds from their current durations.

[eecavanna]

I'll take a crack at this tonight. I don't think I'll finish tonight, but I think I'll at least come away with more knowledge about what will be involved with making this switch.

[eecavanna]

I've run into my first hiccup: incompatible ("contradictory") requirements (you can ignore the yellow line about the broken pre-existing environment).

I don't know whether this also happens with the current build system and I just haven't noticed it, or it is unique to either Poetry or the version specifiers I ended up with in the pyproject.toml file, which are:

mkdocs-mermaid2-plugin = "^1.2.1"
# ...
nmdc-schema = "==11.1.0"

Anyway, I'll try to resolve it.

I see that nmdc-schema (in its own project.toml file) has declared mkdocs-mermaid2-plugin as one of its production dependencies. I think that imposes that constraint on every project that uses nmdc-schema.