Deprecate suspicious.py?

[JulienPalard]

BPO	42238
Nosy	@birkenfeld, @rhettinger, @ned-deily, @ambv, @willingc, @JulienPalard, @pablogsal, @erlend-aasland, @AlexWaygood
PRs	bpo-42238: Doc: Remove make suspicious from the CI and docs builds. #23313 bpo-42238: Check Misc/NEWS.d/next/ for reStructuredText issues. #23802 bpo-42238: Doc CI: Disable suspicious checks. #26575 bpo-42238: rstlint: Add two new checks. #26966 bpo-42238: Fix small rst issue in NEWS.d/. #27238 bpo-42238: [doc] Some more make suspicious false positives. #27945 bpo-42238: [doc] remove unused, and deduplicate, suspicious ignore rules. #28137 bpo-42238: [doc]: update susp-ignored.csv after a98b273c. #28827 bpo-42238: [doc]: Hide false positive in make suspicious. #29636 bpo-42238: [doc] Announce the future removal of make suspicous. #29652 bpo-42238: [doc]: Spotted errors while working on rstlint. #30879 bpo-42238: [doc] Avoid hardcoding fast-moving lines in susp-ignored.csv. #30981 bpo-42238: [doc] moving from rstlint.py to sphinx-lint. #31097 bpo-42238: [doc] Some more make suspicious false positives. #31977 bpo-42238: [doc] Some lines moved in rst, but had hardcoded lineno in… #32070 bpo-42238: [doc]: make suspicious: false positive. #32292 bpo-42238: [doc]: A make sucpicious false positive. #32329

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/JulienPalard'
closed_at = None
created_at = <Date 2020-11-01.23:49:23.937>
labels = ['docs']
title = 'Deprecate suspicious.py?'
updated_at = <Date 2022-04-06.11:53:50.624>
user = 'https://github.com/JulienPalard'

bugs.python.org fields:

activity = <Date 2022-04-06.11:53:50.624>
actor = 'mdk'
assignee = 'mdk'
closed = False
closed_date = None
closer = None
components = ['Documentation']
creation = <Date 2020-11-01.23:49:23.937>
creator = 'mdk'
dependencies = []
files = []
hgrepos = []
issue_num = 42238
keywords = ['patch']
message_count = 56.0
messages = ['380167', '380169', '380171', '381075', '381802', '382219', '383149', '383171', '383292', '390898', '390904', '391163', '391164', '391177', '391182', '391185', '391209', '391212', '392165', '392177', '395311', '396567', '396568', '396774', '396891', '397792', '397802', '400259', '400324', '401123', '403523', '404067', '406575', '406604', '406654', '406810', '406978', '408724', '411565', '411571', '411636', '411642', '411983', '412135', '412136', '412355', '412377', '412382', '412403', '412980', '414984', '415860', '416639', '416643', '416759', '416861']
nosy_count = 9.0
nosy_names = ['georg.brandl', 'rhettinger', 'ned.deily', 'lukasz.langa', 'willingc', 'mdk', 'pablogsal', 'erlendaasland', 'AlexWaygood']
pr_nums = ['23313', '23802', '26575', '26966', '27238', '27945', '28137', '28827', '29636', '29652', '30879', '30981', '31097', '31977', '32070', '32292', '32329']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = None
url = 'https://bugs.python.org/issue42238'
versions = []

[JulienPalard]

I was not here 21 years ago when it was introduced [1], but according to the commit message it was introduced to find leftover Latex mardown.

It tries to find 4 patterns in Sphinx node text (not in raw rst files):

::(?=[^=])| # two :: (but NOT ::=)

This one has ~100 false positive in susp-ignored.csv (pypi classifiers, slices, ipv6, ...)

:[a-zA-Z][a-zA-Z0-9]+| # :foo

This one has ~300 false positive in susp-ignored.csv (slices, C:\, ipv6, ...)

| # (seldom used by itself)

This one has ~20 false positive in susp-ignored.csv (mostly reStructuredText in code-blocks)

(?<!\.)\.\.[ \t]*\w+: # .. foo: (but NOT ... else:)

This one does not have false positives.

The script, on my laptop (with a core i9), is slow (4mn20s), and it's probably way slower on the CI.

I tried to search for suspicious is:pr in:comments on github to see if it's usefull:

• 2 contributor had an issue with the script (bpo-34831: Asyncio tutorial #9748, IDLE - Test WidgetRedirector #66139)
• 5 had to add false positive to susp-ignored.csv (bpo-4356: Add key function support to the bisect module #20556, listdir() doesn't work with non-trivial symlinks #57981, The copy module already uses copyreg #55690, Incorrect coverage file from trace test_pickle.py #53563, os.listdir inconsistenly releases the GIL on win32 #51164)
• 4 had to update susp-ignored.csv (bpo-30670: Add pp function to the pprint module #11769, With invalid FD, os.device_encoding() returns None under Linux but raises an error under Windows #49802, Undetected error in _struct.pack_into #47944, Backport next() #46971)
• 1 did not addedd to susp-ignored but changed to avoid a false positive (bpo-39939: Add str.removeprefix and str.removesuffix #18939)

Case where it actually helped:

• Finding an error: (bpo-36345: Include the code from Tools/scripts/serve.py for the wsgiref-base web server example #12562 .. literalinclude: instead of .. literalinclude::)
• Finding refs in code block (bpo-17045: Improve C-API doc for PyTypeObject. #7413)
• Writing plaintext in Misc/NEWS (bpo-30052: Always regenerate cross-references #1339)

I'd go for enhancing rstlint.py (which is fast, ~1s on my laptop) a bit to try to handle the `.. literalinclude:` missing a `:` errors, and dropping suspicious.

So I'd appreciate feedback on this script, did it helped you recently?

1: 700cf28

[rhettinger]

I would love to see this disappear. For me, it has been a recurring time waster.

[ned-deily]

If it still does useful checks, we could just limit it to running by release managers during the release manufacturing process; it is already run then. It would still allow for problems to be caught and fixed by the RM prior to release tagging. But I don't have a strong opinion about its overall usefulness. I recall it catching at least one problem in a release I was involved with and that was probably before we had the CI checks.

[JulienPalard]

Good idea Ned!

So proposed plan:

• Drop it from docs build and the CI to avoid time loss.
• Add it as a step of PEP-101, for a few release, for good measures
• I'll check it from time to time between releases, just to ensure it does not accumulate tons of things to do on the the release day.

If during a few release, we notice this tool is no loger usefull we'll be able to drop it.

If we spot some errors that can be migrated to the rstlint.py we'll do.

If this tool is in fact usefull, we'll have to think about it again.

[JulienPalard]

New changeset c9c6e9f by Julien Palard in branch 'master':
bpo-42238: Doc: Remove make suspicious from the CI and docs builds. (GH-23313)
c9c6e9f

[JulienPalard]

I'll try to track this closely, the drop of make suspicious has been merged in master, but not backported yet, one week ago.

Up to today, no commit has introduced suspicious warnings (0 false positive, 0 false negative).

I think I'll soon backport the PR, if no one objects, because the current situation is a bit risky: some PR could get merged to master but fail the tests during the backport.

[JulienPalard]

Today I spotted one error that suspicious could have spotted:

-both ``Callable``s no longer validate their ``argtypes``, in 
+both ``Callable``\ s no longer validate their ``argtypes``, in 

I'm fixing it in #23800

make suspicious found it as:

[whatsnew/changelog:23] "`" found in "Callable``s no longer validate their ``argtypes"

So I'll doodle around this to see if tools/rstlint.py or Sphinx itself can see the error.

[JulienPalard]

I created #23802 with a port of the "wrong backtick detector" (at least, one aspect from it) from suspicious to rstlint.

I dropped the suspicious check three weeks ago, and found a single committed issue since then, which was fixable (and is fixed) in rstlint, so I'll continue this experience and see where it goes.

[JulienPalard]

New changeset b973542 by Julien Palard in branch 'master':
bpo-42238: Check Misc/NEWS.d/next/ for reStructuredText issues. (GH-23802)
b973542

[vstinner]

Pablo reenabled the check:

commit 20ac347 (upstream/master, master)
Author: Pablo Galindo <[email protected]>
Date: Mon Apr 12 20:53:15 2021 +0100

Fix Sphinx errors in the documentation and re-activate the suspicious check (GH-25368)

The suspicious check is still executed as part of the release process and release managers have been
lately fixing some actual errors that the suspicious target can find. For this reason, reactivate the suspicious
until we decide what to do in a coordinated fashion.

[pablogsal]

Background on why I reactivated it:

I every single alpha release I had to fix actual documentation errors (no false positives) like the one fixed here:

20ac347

I don't like the check, but the release management team don't like to have to restart the release process just fo fix a bunch of docs issues that should have been detected by the CI.

If we remove the fix, we need to coordinate all together because currently the status quo is that the RM team now is responsable of fixing the documentation errors as there are several release tools (and other ones) that depend on this target passing.

[rhettinger]

If reactivated, the tool needs to be substantially improved. It is NOT smart. The false positives for slicing and logging examples are unnecessarily annoying. It creates a barrier for people submitting documentation patches. Each of the 367 entries in the susp-ignored file represents wasted time for contributors.

Also the CSV format is arcane, hard-to-read, and hard-to-edit. It looks like it was quickly thrown together by someone who didn't care about usability. Perhaps there should be a simpler tools that says, "take this current failure and mark it as a false positive" without trying to be over specific.

Mandatory checks with a high false positive rate are an anti-pattern for CI systems. Already we've had one case of a contributor (me) who abandoned a doc patch rather than fight this tooling.

[pablogsal]

I agree with all you said, but I am against forcing the release management team to absorb the responsibility of fixing all the errors that the CI didn't catch, real or not. Releasing is already a multi-hour process and having many docs errors and formating mistakes to fix is a toll that should be paid at release time.

[vstinner]

Pablo: "I every single alpha release I had to fix actual documentation errors (no false positives)"

I don't get it. Do you mean that the Python release process is blocked by typos in the documentation? Why did you catch them in the release process, and not in the regular Python development workflow?

Can't we get the same errors in the regular workflow?

I understand that the release process requires no warning from "make suspicious", but as explained in other comments, this tool is not reliable. While it catchs some real bugs, the cost of false positives is too high.

Doc/tools/susp-ignored.csv contains 368 lines, IMO it's a sign that this tool either requires a serious reworking, or must be disabled/removed.

Doc/tools/extensions/suspicious.py is a CPython specific extension, it's not part of Sphinx. Maybe someone should open a feature request on Sphinx to get a similar feature without false positives (or at least less false positives).

[pablogsal]

Why did you catch them in the release process, and not in the regular Python development workflow?

Because our release tools run suspicious.py and it fails and then we need to manually inspect every single failure and fix it as almost all times there are real positives among them

Can't we get the same errors in the regular workflow?

That's what I did reactivating suspicious.py

I understand that the release process requires no warning from "make suspicious", but as explained in other comments, this tool is not reliable. While it catchs some real bugs, the cost of false positives is too high

Sure, we all agree here. Where it seems that we don't agree is how to proceed. I would be super happy to remove the tool if is agreed everywhere, including with the rest of release managers.

[vstinner]

Julien Palard is working actively on the Python documentation. He is the one who removed suspicious "from the CI and docs builds". I trust his judgment here. I suggest to remove it again from there *and* from the release scripts.

[vstinner]

I wrote a PR to enhance -W option documentation. The suspicious job complains whereas my Sphinx format is valid:

The full form of argument is::

   action:message:category:module:lineno

Example of suspicious error:

[using/cmdline:438] ":message" found in "action:message:category:module:lineno"

I see that Doc/tools/suspicious.csv already contains an ignore rule for Doc/using/cmdline.rst.

[vstinner]

Doc/tools/extensions/suspicious.py is a CPython specific extension, it's not part of Sphinx. Maybe someone should open a feature request on Sphinx to get a similar feature without false positives (or at least less false positives).

This extension was written by Gabriel Genellina in 2009 when he discovered that:

Several documents contain invalid reST markup that "leaks"
into the html output (missing ``, incorrect indentation, etc.)

Extension added in bpo-4811 with:

commit 700cf28
Author: Georg Brandl <[email protected]>
Date: Sun Jan 4 10:23:49 2009 +0000

Add "suspicious" builder which finds leftover markup in the HTML files.

Patch by Gabriel Genellina.

The code didn't evolved much since it was added in 2009. Main changes.

commit 144c269
Author: Ezio Melotti <[email protected]>
Date: Thu Mar 28 18:01:11 2013 +0200

Update the suspicious builder to detect unused rules, and remove currently unusued rules.

bpo-22537:

commit 14b5a4d
Author: Georg Brandl <[email protected]>
Date: Thu Oct 2 08:26:26 2014 +0200

Closes bpo-22537: Make sphinx extensions compatible with Python 2 or 3, like in the 3.x branches

Filename renamed:

• master: Doc/tools/extensions/suspicious.py
• Commit 160cbce renames Doc/tools/suspicious.py to Doc/tools/extensions/suspicious.py
• Commit f16fbf9 renames Doc/tools/sphinxext/suspicious.py to Doc/tools/suspicious.py

[JulienPalard]

I should have monitored this more closely, I started monitoring it weekly, then life got over until today when I'm even surprised to see activity on the issue, sry!

(Surprise leading me to investigate why I had not received notifications from bpo, leading me to a bug in my sieve filter...)

I counted, during your releases make suspicious spotted:

• 1 true positive in 20ac347
• 1 false positive in 20ac347
• 1 true positive in 57f21db
• 1 false positive in de833b6

if I missed none, the success ratio is not good (which is already known).

What I'm aiming to do in this issue is to list the true positives over time, by passing the tool manually from time to time, to try to see what kind of rule is usefull in such a tool, and ultimately try to spot those errors in a reliable way.

I'd go for removing it from both release process and the CI, this would lower pressure on RMs and contributors, while easing the study of the tool in this issue.

I even consider removing it from the Makefile / tools hierarchy to ensure nobody runs it, because if someone run it locally during its contribution process it may hide true positives from me (leading me to think there's no true positives and the tool is useless).

[vstinner]

At least, I would suggest to remove it from the release process.

If some people working on the documentation want to keep the tool, maybe it can be an optional CI job? IMO the problem is that currently, it's part of a single "Documentation" job. I don't recall if it's mandatory or not. Also, we need maybe more explanation in the CI job result how to mark false positives.

[JulienPalard]

New changeset 227a093 by Julien Palard in branch 'main':
bpo-42238: Doc CI: Disable suspicious checks. (GH-26575)
227a093

[JulienPalard]

Spotted a true positive in b19f455:

:const:``None``

(I'm trying to build a true positive list, to have usefull cases where suspicious is usefull, so in the long term I can maybe implement those cases in rslint instead.)

[JulienPalard]

The previous one could probably be implemented in rstlint using (an equivalent of):

git grep ':[a-z]\+:``[^:` ]+``' Doc/

Maybe specialized to known roles, like the script specializes to known directives.

[JulienPalard]

Another true positive:

... versionchanged:: 3.11

from : 6cb145d / #26820

I also think it can be implemented in rstlint.

[JulienPalard]

New changeset 01331f1 by Julien Palard in branch 'main':
bpo-42238: rstlint: Add two new checks. (GH-26966)
01331f1

[JulienPalard]

Another one \o/

Fix in: #27238

It is:

:func:pdb.main

Detected by make suspicious as:

WARNING: [whatsnew/changelog:320] ":func" found in ": Refactor argument processing in :func:pdb.main to simplify"
WARNING: [whatsnew/changelog:320] ":pdb" found in ": Refactor argument processing in :func:pdb.main to simplify"

I added a test in the same PR to spot this one in rstlint.

[ambv]

New changeset fbf1008 by Julien Palard in branch 'main':
bpo-42238: Fix small rst issue in NEWS.d/. (bpo-27238)
fbf1008

[JulienPalard]

10 false positives today, see: https://github.com/python/cpython/pull/27945/files

I think I now need to put a deadline on this issue, let's say: one year without true positives and I drop it, so 2021-07-19 for now.

(In the meantime, I continue to convert each true positives to rstlint, I converted 3 of them so far.)

[vstinner]

Congrats @JulienPalard for sphinx-lint!