Avoid issues related with pins persistence #355
Conversation
conda-bot
added
the
cla-signed
| n_pins = 0 | ||
| tasks = self._specs_to_tasks(in_state, out_state) | ||
| for (task_name, task_type), specs in tasks.items(): | ||
| log.debug("Adding task %s with specs %s", task_name, specs) | ||
| if task_name == "ADD_PIN": | ||
| if task_name == "ADD_PIN" and attempt == 1: |
There was a problem hiding this comment.
Makes sense given the Matrix discussion linked in the OP.
If it doesn't complicate things too much, would it make more sense to add an is_initial_attempt: bool (or just attempt: int) parameter to _specs_to_tasks and have the ADD_PIN task simply not created?
To me, that would conceptually make more sense since creating a "task" that will be ignored otherwise is awkward ;).
There was a problem hiding this comment.
(Or the tasks could be created beforehand and split into initial and recurring ones; but that might need more restructuring which might be overkill for this single special-handling case for now.)
There was a problem hiding this comment.
Yea I get your point, but here's my rationale:
IMO the limitation that solver.add_pins() has persistent effects in the pool is an implementation detail of the libmamba library. specs_to_tasks aims to be as unaware of those implementation details as possible. It's not that we shouldn't pin in the 2nd iteration and following ones; we do want, but there's a technical limitation. I'd like specs_to_tasks to reflect that logic, hence why I ignore it here.
It could be cleaner, but we are going to remove a lot of code from the `state, and I plan to include some cleanup in the solver module.
I hope it makes sense.
There was a problem hiding this comment.
How many attempts would be typical?
There was a problem hiding this comment.
Depends on the environment. Hopefully the first one is enough. But with old ones with many war scars, it can take a few. We fast-track some cases, like changing Python version, because otherwise we'd need for the solver to realize all the Python-depending packages are conflicting.
|
Thanks for the checks @mbargull! |
Description
When the
libmambapy.Solvercalls.add_pin(), it injects some special packages in theinstalledpool / repo (see thread). We keep the same repo across attempts in our retry loop, so we only need to / should invoke that once.That seems to fix #354, but let's see if another problems arise in the test suite.
Checklist - did you ...
newsdirectory (using the template) for the next release's release notes?