Should I prefer granular per-dir build files, or one top-level one with big fat globbing. 🧵 1. First option is what

tailor

would have me use. One

BUILD

file in each directory that matters. ◦ Pros: Granular. Pants internals kind of assume this pattern (so target names are sane, etc...). ◦ Cons: Still boilerplate-y (MUCH less so than say, Bazel. But still have N files which just repeat

python_sources()

,

python_test_utils

,

python_tests

over and over). 2. Second is one

BUILD

file in the "root" which has nested globs for each (E.g.

sources=["***/**.py", "!**/conftest.py", ...]

) and then I can override when necessary at the directory level (assuming that works 🤔 ) ◦ Pros: Minimal boilerplate ◦ Cons: Less granular. Pants internals don't assume this (the target name of the generator is funny looking when printed) Q: Is pants smart enough for option 2 to not have double ownership? Can it be taught to be? Q: What considerations should be made about cache invalidation? If I edit the top-level file in #2, do I bust any caching? Q: For option 2, can I improve the names that get spit out? Does the "generators-are-not-targets" proposal help this case?

As it stands, we recommend one target per directory. What

tailor

does. See https://www.pantsbuild.org/docs/targets#target-generation and https://github.com/pantsbuild/pants/issues/13488 But we think this isn't flexible enough, you have to repeast yourself too much: https://github.com/pantsbuild/pants/issues/13767

(added my two cents)

It is not smart enough to handle double ownership, and I suspect you'd get a rule graph cycle if you tried querying what targets already exist inside a target generation rule. Since that's cyclical to compute

I'm afraid to query my top-level generator 😬

Generate-Generators from top level 😁 .. would prob still be cyclical though..

Related is that currently the only way to parametrize config is to create multiple targets for the same entity. Thus, multiple targets is currently a first-class concept But I'd love to replace that with something better. This parametrization problem has stymied many past attempts to simplify targets. Feedback welcomed in https://github.com/pantsbuild/pants/issues/13882

Should I be worried that Option 2 seemingly works without cycles? 😬