cached runs of dmypy stop reporting "unused type: ignore" warnings

[rggjan]

Using mypy 0.782, I run
dmypy run .

on my codebase. The first time it runs, it correctly reports a few

error: unused 'type: ignore' comment

same as mypy. However, when rerunning dmypy run ., these errors disappear, while all other errors remain.

For some reason, this type of warning (warn_unused_ignores) seems to be missed sometimes when using the cache.

Unfortunately, it doesn't seem trivial to reproduce, so I can't give a reproducible project right now...

[jgillard]

I noticed this today with mypy 0.942, so this is still an issue.

> dmypy run
Success: no issues found in 604 source files

> mypy --cache-dir=/dev/null
apps/matchmaking/services.py:260: error: Unused "type: ignore" comment
Found 1 error in 1 file (checked 604 source files)

> mypy --version
mypy 0.942

[peterbygrave]

This is still a problem in 0.991

[AlexWaygood]

Has anybody managed to find a reproducible example yet?

[meshy]

I have recently encountered this on 1.0.1. I have not yet produced a minimal example.

In my case the bulk of the unused type: ignore warnings which disappear are on lines that have been typed by the django-stubs project. I wonder if it may have something to do with that project, or with plugins in general.

[meshy]

I have a new example that doesn't have any external requirements.

unused/__init__.py
(empty file)

unused/empty.py
(empty file)

example.py

from unused.empty import *
a = 1  # type: ignore

Instructions (in a clean virtualenv)

$ dmypy stop
Daemon stopped

$ dmypy run -- --package example --warn-unused-ignores
Daemon started
example.py:2: error: Unused "type: ignore" comment
Found 1 error in 1 file (checked 46 source files)

$ dmypy run -- --package example --warn-unused-ignores
Success: no issues found in 1 source file

---

From what I can tell this issue happens when we call fine_grained_manager.update([], to_delete) in dmypy_server.Server.fine_grained_increment_follow_imports.

In this case, to_delete is [('unused', '/.../mypy-9655/unused/__init__.py')].

For some reason, when this update is made, the Unused "type: ignore" comment also goes missing.

Next up: dig into mypy, and work out what's going on.

EDIT: I have created a failing test based on this example, and continue to investigate a fix.

[meshy]

My PR now contains a fix for this that I hope should get things working for most people.

[ilevkivskyi]

This should be fixed by #14835

[meshy]

While I think things have improved, I'm not sure that it was entirely fixed by that PR.

I've produced two more PRs (failing tests only) with similar cases, though I haven't yet been able to get to the bottom of them yet:

• Type ignore comments erroneously marked as unused by dmypy #15043
• More cases showing errors disappearing from dmypy reruns #15049

[ilevkivskyi]

OK, I see, let's then keep this open to track those.

[meshy]

Thanks :)

[m-honel-ywh2bt]

Still facing the issue with python 3.12.1 / mypy 1.7.1.

Using meshy's example in a clean virtualenv:

$ dmypy stop
Daemon stopped

$ dmypy --version
dmypy 1.7.1

$ python --version
Python 3.12.1

$ dmypy run -- --package example --warn-unused-ignores
Daemon started
example.py:2: error: Unused "type: ignore" comment  [unused-ignore]
Found 1 error in 1 file (checked 1 source file)

$ dmypy run -- --package example --warn-unused-ignores
Success: no issues found in 1 source file

[meshy]

@m-honel-ywh2bt we're making progress on this!

Just a couple of days ago we solved the examples in #15043. Testing on my local code base suggests that we may have missed some edge-cases there though, so I'm trying to produce some more minimal examples for those.

There are also some cases in #15049 that we haven't found the root cause of yet. I hold onto the hope that we'll get to the bottom of this!

[meshy]

If anyone following this thread is able to check if #15043 works for them, I'd appreciate the feedback.

[m-honel-ywh2bt]

@meshy Using #15043, I can confirm that Unused "type: ignore" comment errors seem to appear everytime dmypy is ran, without having to restart the daemon. 👌🏻

[JukkaL]

I think there is another case where type ignore comments still aren't being reported. Here are a few failing test cases:

[case testUnusedTypeIgnorePreservedAfterChangeInOtherModule1]
# flags: --warn-unused-ignores
[file main.py]
import a

def f() -> None:
    1()
    a = 1  # type: ignore

[file a.py]
[file a.py.2]
x = 1

[out]
main.py:4: error: "int" not callable
main.py:5: error: Unused "type: ignore" comment
==
main.py:4: error: "int" not callable
main.py:5: error: Unused "type: ignore" comment

[case testUnusedTypeIgnorePreservedAfterChangeInOtherModule2]
# flags: --warn-unused-ignores
[file main.py]
import a

def f() -> None:
    1()
    a = 1  # type: ignore

def g() -> None:
    1()  # type: ignore

[file a.py]
[file a.py.2]
x = 1

[out]
main.py:4: error: "int" not callable
main.py:5: error: Unused "type: ignore" comment
==
main.py:4: error: "int" not callable
main.py:5: error: Unused "type: ignore" comment

The error on line 5 isn't reported on the second step.

The root cause is that we don't report unused type ignores in nodes only processed using reprocess_nodes, i.e. when following fine-grained dependencies or reprocessing nodes that previously generated errors.

I think this is a potential way to fix this (but I haven't thought about this very carefully yet):

• Collect a set of all the triggered/processed nodes in propagate_changes_using_dependencies.
• At the end of a daemon update, perform the equivalent of generate_unused_ignore_notes, but only report errors within the previously triggered nodes. The filtering step seems necessary, as some parts of modules are often not processed during a fine-grained increment.

[meshy]

@JukkaL good catch.

Do you think this blocks #15043, or could we tackle that case separately?

[JukkaL]

Ah, I forgot to mention that the test cases I mentioned above aren't specifically covered by #15043 yet. It's okay to leave fixing these for another PR, but we shouldn't close this issue until all known cases are fixed.

[meshy]

OK great, thank you. I've removed the "fixes" comment from that PR so we avoid closing this issue.