NEW: Better package splitting for multi-output recipes with negative glob in outputs/files #5216

carterbox · 2024-03-06T01:20:20Z

Description

Ressurects #4197. Closes #4196.

The purpose of this PR is to make it easier to split packages into multiple outputs using glob expressions. This is accomplished in two ways:

A negative pattern match is provided. This makes it easier to use fewer glob expressions if you need to include an entire directory tree except for a single file (type). For example, you might want to glob lib/**/libfoo*, but not lib/**/*.a
Only the files installed to the PREFIX in the top-level build are considered. This removes the need to craft your glob expressions to avoid the artifacts installed to the PREFIX by host dependencies. For example, you can now glob include/**/* without pulling in the headers of other packages. It is already the behavior of a single-output recipe to ignore files added to the prefix by host dependencies, so I'm not sure why this feature didn't make it into multi-output recipes.

These new behaviors are only used if the include/exclude keywords are used under the files key. The previous behavior is retained.

outputs:
  - name: foo
    files:
      - bin/*  # old behavior; includes artifacts from other packages
  - name: bar
    files:
      include:
        - bin/*  # new behavior; only matches artifacts from this recipe
      exclude:  # optional
        - bin/*.exe # new behavior
  - name: zee
    script: install.py  # old behavior

Checklist - did you ...

Add a file to the news directory (using the template) for the next release's release notes?
Add / update necessary tests?
Add / update outdated documentation?

closes conda#4196

as keys in `files`

conda-bot · 2024-03-06T01:20:42Z

We require contributors to sign our Contributor License Agreement and we don't have one on file for @carterbox.

In order for us to review and merge your code, please e-sign the Contributor License Agreement PDF. We then need to manually verify your signature, merge the PR (conda/infrastructure#891), and ping the bot to refresh the PR.

docs/source/resources/define-metadata.rst

codspeed-hq · 2024-03-20T17:10:49Z

CodSpeed Performance Report

Merging #5216 will not alter performance

_{Comparing carterbox:files-exclude (c92f1a1) with main (cdca0b4)}

Summary

✅ 3 untouched benchmarks

carterbox · 2024-03-20T19:49:28Z

I'm not sure that the failure for linux 3.8 23.5.0 serial is related to this PR.

----------------------------- Captured stderr call -----------------------------
Traceback (most recent call last):
  File "/usr/share/miniconda/envs/test/lib/python3.8/site-packages/conda/exception_handler.py", line 16, in __call__
  File "/usr/share/miniconda/envs/test/lib/python3.8/site-packages/conda/cli/main.py", line 66, in main_subshell
  File "/usr/share/miniconda/envs/test/lib/python3.8/site-packages/conda/cli/conda_argparse.py", line 31, in <module>
  File "/usr/share/miniconda/envs/test/lib/python3.8/site-packages/conda/base/context.py", line 24, in <module>
ModuleNotFoundError: No module named 'conda._vendor.appdirs'

jaimergp

I like this a lot. The PR is super clear and the documentation is on-point. I only added a couple comments to clarify my own doubts. The new positional argument could be considered API breaking, so let's see how the rest of the team feels about it. I don't have strong feelings, just applying an abundance of caution Just In Case ✨ .

Once we have discussed that, I'll be super happy to approve. 👍

jaimergp · 2024-03-21T14:17:50Z

conda_build/build.py

+    metadata: MetaData,
+    env,
+    stats,
+    new_prefix_files: set[str],


Since this is adding a new positional argument to a (let's assume) public signature, what about making it a None kwarg. This way downstream users won't run into number of arguments errors:

Suggested change

new_prefix_files: set[str],

new_prefix_files: set[str] = None,

This is a good point; it is API breaking assuming these functions are part of the public API. I need to mark the new argument as typing.Optional and/or set appropriate defaults.

conda_build/build.py

jaimergp · 2024-03-21T14:22:05Z

conda_build/build.py

+    metadata: MetaData,
+    env,
+    stats,
+    new_prefix_files: set[str],


Same comment here about the API breakage. Not sure how prominent the usage of wheel outputs is, though.

#5216 (comment)

jaimergp · 2024-03-21T14:23:27Z

docs/source/resources/define-metadata.rst

@@ -1299,9 +1299,10 @@ You can specify files to be included in the package in 1 of
 Explicit file lists are relative paths from the root of the
 build prefix. Explicit file lists support glob expressions.
 Directory names are also supported, and they recursively include
-contents.
+contents. Files installed to the prefix by host dependencies will
+be matched by glob expressions.


I would maybe even highlight this further with an admonition or something similar. I wasn't aware of that design issue!

with an admonition or something similar.

I don't know what you mean. You want stronger language? For example: "WARNING! Files installed to the prefix by host dependencies also will be matched by glob expressions."

I wasn't aware of that design issue!

Yeah, I'm not sure whether this is a bug/oversight or a feature. It's definitely unexpected to me. The most active contributor in this section is @msarahan; maybe they know?

I meant an admonition in the RST sense. Triple-backtick + warning kind of syntax. Equivalent to this:

Warning
Files installed to the prefix by host dependencies also will be matched by glob expressions.

Done in 642906b

news/files-exclude.rst

jaimergp · 2024-03-21T14:26:27Z

tests/test-recipes/split-packages/copying_files/meta.yaml

    files:
-      - subpackage_file1
-      - somedir
-      - "*.ext"
+      include:
+        - subpackage_file1
+        - somedir
+        - "*.ext"
+      exclude:
+        - "*3.ext"


Should we replace the old test with a new one instead of adding a new case? Maybe I would add a new output in this meta.yaml and then test the different behaviors differently, if possible (e.g. how the list-of-str behaviour does copy anything in PREFIX, but include/exclude doesn't?).

I have updated the test recipe to include multiple outputs: one uses the old file matching and one uses the new file matching.

mbargull · 2024-03-21T14:55:56Z

docs/source/resources/define-metadata.rst

+Files can be excluded by specifying `files` as a dictionary separating
+files to `include` from those to `exclude`. Files installed to the prefix
+by host dependencies are automatically excluded when the include/exclude
+syntax is used:


We need to mention the precedence here, i.e., exclude entries have higher priority as per the implementation.

This mentions the new behavior, i.e., "Files [...] excluded when the include/exclude syntax is used".
But since this is a deviation from the previous "include" logic (which would be more of a force-include since it allows packaging pre-existing files.), I think we should make the difference more prevalent/explain the behavior for the old non-dict case more explicitly.

We need to mention the precedence here, i.e., exclude entries have higher priority as per the implementation.

exclude has a higher priority than which operation? It's unclear to me what ambiguity we would be removing by mentioning priority.

This mentions the new behavior, i.e., "Files [...] excluded when the include/exclude syntax is used". But since this is a deviation from the previous "include" logic (which would be more of a force-include since it allows packaging pre-existing files.), I think we should make the difference more prevalent/explain the behavior for the old non-dict case more explicitly.

Suggested change

Files can be excluded by specifying `files` as a dictionary separating

files to `include` from those to `exclude`. Files installed to the prefix

by host dependencies are automatically excluded when the include/exclude

syntax is used:

When defining `outputs/files` as a list, any file in the prefix (including those

installed by host dependencies) matching one of the glob expressions is

included in the output. Greater control over file matching may be

achieved by defining `files` as a dictionary separating files to

`include` from those to `exclude`.

When using include/exclude, only files installed by

the current recipe are considered. i.e. files in the prefix installed

by host dependencies are not matched. include/exclude may not be used

simultaneously with glob expressions listed directly in `outputs/files`.

We need to mention the precedence here, i.e., exclude entries have higher priority as per the implementation.

exclude has a higher priority than which operation?

The kind of precedence where if you do

files: include: - lib/libfoo.so exclude: - lib/libfoo.so

the exclusion wins.

It is now explained that when files match both exclude and include, they are excluded. d1b12a1

jaimergp · 2024-03-21T17:31:08Z

conda_build/build.py

+            include = files.get("include", [])
+            exclude = files.get("exclude", [])


This syntax will be affected by the "key is set to an empty block, which means None in YAML" kind of issue. We should change it to:

Suggested change

include = files.get("include", [])

exclude = files.get("exclude", [])

include = files.get("include") or []

exclude = files.get("exclude") or []

It might also be a good opportunity to sneak @jakirkham's changes added by #4971.

See line 1818:

- files = output.get("files", []) + files = output.get("files") or []

My understanding is that files.get("include") may return None, but include must be a list. So we use the falsiness of None to replace it with []. Equivalent to:

include = files['include'] if ('include' in files and files['include'] != None) else []

You might have

- output: something files: include: - path exclude: - not-this-path # [linux]

and then you'll have files["exclude"] = None on macOS, for example (linux is false). Not sure at what point we are ensuring that include is a list.

Under that circumstance files.get("exclude", []) will be None instead of the expected []. Is that what you mean? I think we are saying the same 😬

Yes. We are saying the same thing. I'm just trying to figure out if we should be doing something stronger than converting Falsy values into the empty list. Perhaps convert any non-list into a list? Something like:

include = files.get("include") if isinstance(files.get("include"), list) else []

That might silence errors when users populate it with a dict or something accidentally. I think we can just play it simple with falsy promotions.

jaimergp · 2024-03-21T17:44:34Z

conda_build/build.py

@@ -2817,7 +2848,7 @@ def build(
                    # This is wrong, files has not been expanded at this time and could contain
                    # wildcards.  Also well, I just do not understand this, because when this
                    # does contain wildcards, the files in to_remove will slip back in.
-                    if "files" in output_d:
+                    if "files" in output_d and not isinstance(output_d["files"], dict):


Suggested change

if "files" in output_d and not isinstance(output_d["files"], dict):

if not isinstance(output_d.get("files") or (), dict):

I have updated this condition to:

if ( "files" in output_d and output_d["files"] is not None and not isinstance(output_d["files"], dict) ):

jaimergp · 2024-05-22T11:00:25Z

tests/test_subpackages.py::test_subpackage_recipes[copying_files] is failing on macOS and Windows @carterbox

For some reason, target_platform is not set.

docs/source/resources/define-metadata.rst

Co-authored-by: Daniel Ching <carterbox@users.noreply.github.com>

wolfv · 2024-06-01T12:45:19Z

@carterbox I've started to implement this idea in rattler-build, too: prefix-dev/rattler-build#819

Would love your feedback once it's in a testable state. This is also in preparation of a "top-level" cache build (that can be split up into multiple packages).

carterbox · 2024-06-02T22:31:28Z

Thanks, @isuruf for fixing my platform detection logic in the tests.

@jaimergp, Looks like the failed test on Linux is in a test unrelated to this PR. Maybe someone can restart the tests to see if it is ephemeral or not?

kenodegard

this looks good, just a minor suggestion and a question about deprecating the old behavior

conda_build/build.py

kenodegard · 2024-06-05T11:12:42Z

conda_build/build.py

+        else:
+            keep_files = {
+                os.path.normpath(pth)
+                for pth in utils.expand_globs(files, metadata.config.host_prefix)
+            }


Given the change in behavior above with only allowing new files to be included should the old behavior of including any file in the prefix be deprecated?

Suggested change

else:

keep_files = {

os.path.normpath(pth)

for pth in utils.expand_globs(files, metadata.config.host_prefix)

}

else:

keep_files = {

os.path.normpath(pth)

for pth in utils.expand_globs(files, metadata.config.host_prefix)

}

if keep_files - new_prefix_files:

deprecated.topic(...)

It's still useful for eg when building gcc_bootstrap packages. Deprecating the implicit behaviour and adding a new flag would be useful.

I feel that the current behavior of including all files from the prefix for multi-output and only installed files from the prefix for single-output recipes is inconsistent and thus non-intuitive. rattler-build is adding a section called "always_include_files" for recipes that want to repackage artifacts from their host dependencies.

I'm favor of deprecating the old behavior, but I'm uncertain about any removal or replacement schedule.

rattler-build is adding a section called "always_include_files" for recipes that want to repackage artifacts from their host dependencies.

This exists in conda already, and is in use in conda-forge. Main use-case for that from my POV is to be explicit about overwriting some CMake metadata if an incremental build adds more targets to an already-existing CMake file (example).

We can deprecate that behaviour in a separate PR too. This PR has been open for too long and I wouldn't want to test your patience further @carterbox <3

🚀 LGTM! Just wanted to understand the thoughts/plans for the old behavior.

I'll open an issue to capture this feedback, though I think with the documentation this PR added we are in a better place already.

docs/source/resources/define-metadata.rst

Co-authored-by: Bianca Henderson <beeankha@gmail.com>

Co-authored-by: Ken Odegard <kodegard@anaconda.com>

jaimergp · 2024-06-07T15:15:02Z

Thanks @carterbox and everyone involved! Three months and 64 comments later, this one is in!

refs: - conda/conda-build#5216 Signed-off-by: Marcel Bargull <marcel.bargull@udo.edu>

The expand_globs function from conda_build.utils logs an ERROR when a glob expression returns no matches, this is overly alarming because the user may now use negative glob expressions which they don't care if it returns empty or the user may want to use the same set of glob expressions for multiple platforms some of which may return empty on some platforms. conda#5216 conda#5455

duncanmmacleod and others added 7 commits November 9, 2021 15:15

conda_build.build: simple implementation of files exclusion

a545986

closes conda#4196

tests: add test for glob exclusion in files

14e3548

add docs for negative file matches

7fbc142

news: add negative file match news entry

67904ce

conda_build.build: support separate include/exclude file lists

e56c573

as keys in `files`

Merge branch 'main' into files-exclude

be19766

REF: Only consider new files to the PREFIX when file picking

b667724

conda-bot mentioned this pull request Mar 6, 2024

Adding CLA signee carterbox conda/infrastructure#891

Merged

carterbox commented Mar 6, 2024

View reviewed changes

docs/source/resources/define-metadata.rst Outdated Show resolved Hide resolved

carterbox added 2 commits March 5, 2024 20:30

Update docs/source/resources/define-metadata.rst

063ff89

Merge branch 'main' into files-exclude

1ab7638

carterbox marked this pull request as ready for review March 8, 2024 19:21

carterbox requested a review from a team as a code owner March 8, 2024 19:21

carterbox added 2 commits March 8, 2024 13:27

Update define-metadata.rst

de7715d

Update files-exclude.rst

f56e201

carterbox changed the title ~~NEW: Negative matching 'files' for outputs~~ NEW: Better package splitting for multi-output recipes with negative glob in outputs/files Mar 8, 2024

Merge branch 'main' into files-exclude

0537c19

conda-bot added the cla-signed [bot] added once the contributor has signed the CLA label Mar 20, 2024

carterbox added 3 commits March 20, 2024 12:47

STY: Add missing __future__ import

30a5e1b

STY: Address RUFF lints

a92964f

STY: Remove extra whitespace in news

a2e3963

Merge branch 'main' into files-exclude

bfac4e0

jaimergp reviewed Mar 21, 2024

View reviewed changes

mbargull reviewed Mar 21, 2024

View reviewed changes

jaimergp reviewed Mar 21, 2024

View reviewed changes

BUG: Use default value to make API backwards-compatible

3c0b090

h-vetinari mentioned this pull request May 9, 2024

add recipe extension cep conda/ceps#72

Open

isuruf added 2 commits May 31, 2024 08:51

Fix tests

256017a

For some reason, target_platform is not set.

Merge branch 'main' into files-exclude

e5fdcbf

isuruf reviewed May 31, 2024

View reviewed changes

docs/source/resources/define-metadata.rst Outdated Show resolved Hide resolved

isuruf and others added 2 commits May 31, 2024 08:59

Update docs/source/resources/define-metadata.rst

606ee08

Remove serial

6d33054

Co-authored-by: Daniel Ching <carterbox@users.noreply.github.com>

carterbox mentioned this pull request Jun 2, 2024

Add build.files and negative globs to rattler-build prefix-dev/rattler-build#819

Merged

1 task

isuruf previously approved these changes Jun 3, 2024

View reviewed changes

jaimergp previously approved these changes Jun 3, 2024

View reviewed changes

kenodegard previously approved these changes Jun 5, 2024

View reviewed changes

JeanChristopheMorinPerso reviewed Jun 5, 2024

View reviewed changes

docs/source/resources/define-metadata.rst Outdated Show resolved Hide resolved

carterbox dismissed stale reviews from kenodegard and jaimergp via 5d8efe9 June 5, 2024 17:55

Update docs/source/resources/define-metadata.rst

5d8efe9

Co-authored-by: Bianca Henderson <beeankha@gmail.com>

carterbox dismissed isuruf’s stale review via 5d8efe9 June 5, 2024 17:55

REF: Use walrus operator to reduce conditions

c92f1a1

Co-authored-by: Ken Odegard <kodegard@anaconda.com>

jaimergp approved these changes Jun 6, 2024

View reviewed changes

kenodegard approved these changes Jun 6, 2024

View reviewed changes

jaimergp merged commit 4ec4ac7 into conda:main Jun 7, 2024
29 checks passed

carterbox deleted the files-exclude branch June 7, 2024 18:12

mbargull added a commit to isuruf/r-base-feedstock that referenced this pull request Jun 21, 2024

Update comment for conda-build>=24.7's files/exclude

088187e

refs: - conda/conda-build#5216 Signed-off-by: Marcel Bargull <marcel.bargull@udo.edu>

beeankha mentioned this pull request Jul 16, 2024

Release 24.7.x #5381

Closed

55 tasks

This was referenced Aug 10, 2024

Switch to new files: include/exclude capability conda-forge/llvmdev-feedstock#283

Draft

BUG: problems new files: include/exclude functionality #5455

Open

carterbox mentioned this pull request Aug 30, 2024

DOC: Demote logging about non-matching glob to warning #5472

Draft

1 task

	new_prefix_files: set[str],
	new_prefix_files: set[str] = None,

-Files can be excluded by specifying `files` as a dictionary separating
-files to `include` from those to `exclude`. Files installed to the prefix
-by host dependencies are automatically excluded when the include/exclude
-syntax is used:
+When defining `outputs/files` as a list, any file in the prefix (including those
+installed by host dependencies) matching one of the glob expressions is
+included in the output. Greater control over file matching may be
+achieved by defining `files` as a dictionary separating files to
+`include` from those to `exclude`.
+When using include/exclude, only files installed by
+the current recipe are considered. i.e. files in the prefix installed
+by host dependencies are not matched. include/exclude may not be used
+simultaneously with glob expressions listed directly in `outputs/files`.

		include = files.get("include", [])
		exclude = files.get("exclude", [])

	if "files" in output_d and not isinstance(output_d["files"], dict):
	if not isinstance(output_d.get("files") or (), dict):

NEW: Better package splitting for multi-output recipes with negative glob in outputs/files #5216

NEW: Better package splitting for multi-output recipes with negative glob in outputs/files #5216

Conversation

carterbox commented Mar 6, 2024 • edited Loading

Description

Checklist - did you ...

conda-bot commented Mar 6, 2024

codspeed-hq bot commented Mar 20, 2024 • edited Loading

Merging #5216 will not alter performance

Summary

carterbox commented Mar 20, 2024

jaimergp left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

carterbox Mar 21, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jaimergp Mar 21, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jaimergp Mar 21, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jaimergp commented May 22, 2024

wolfv commented Jun 1, 2024

carterbox commented Jun 2, 2024

kenodegard left a comment

Choose a reason for hiding this comment

kenodegard Jun 5, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jaimergp commented Jun 7, 2024

carterbox commented Mar 6, 2024 •

edited

Loading

codspeed-hq bot commented Mar 20, 2024 •

edited

Loading

carterbox Mar 21, 2024 •

edited

Loading

jaimergp Mar 21, 2024 •

edited

Loading

jaimergp Mar 21, 2024 •

edited

Loading

kenodegard Jun 5, 2024 •

edited

Loading