`pkg_rpm` to `pkg_tar` contains duplicated rpm files

[k0walik]

I noticed inconsistent behaviour when we feed pkg_deb and pkg_rpm to pkg_tar:

• pkg_deb to pkg_tar contains only a single file that matches package_file_name attribute
• pkg_rpm to pkg_tar contains two files in tarball: one named after the rule, the other one that is named after package_file_name

While putting a single tarball may not have much sense, there are use-cases where one need to ship additional files along with deb/rpm package (i.e. instructions). My initial thought was the difference is lack of OutputGroupInfo in pkg_rpm, but I think the problem is DefaultInfo.files return value. pkg_deb returns only the output file, while pkg_rpm contains two output files: the default output file and output file.

Minimum reproducible example:

# //tests/deb_rpm_tar/BUILD
load("//pkg:deb.bzl", "pkg_deb")
load("//pkg:tar.bzl", "pkg_tar")
load("//pkg:rpm.bzl", "pkg_rpm")
load("//pkg:mappings.bzl", "pkg_files")

genrule(
    name = "gen_foo",
    outs = ["usr/bin/foo"],
    cmd = "touch $@",
)

pkg_tar(
    name = "deb_input",
    srcs = [":gen_foo"],
)

pkg_deb(
    name = "test_deb",
    data = ":deb_input",
    description = "Some package",
    maintainer = "john.doe@example.com",
    package = "debpackage",
    package_file_name = "debpackage-0.0.0.deb",
    version = "0.0.0",
)

pkg_tar(
    name = "deb_tar",
    srcs = [":test_deb"],
)

pkg_files(
    name = "rpm_input",
    srcs = [":gen_foo"],
)

pkg_rpm(
    name = "test_rpm",
    srcs = [":rpm_input"],
    description = "Description",
    license = "Apache 2.0",
    package_file_name = "rpmpackage-0.0.0-0.rpm",
    release = "1",
    summary = "Summary",
    version = "0.0.0",
)

pkg_tar(
    name = "rpm_tar",
    srcs = [":test_rpm"],
)

$ tar tvf bazel-out/k8-fastbuild/bin/tests/deb_rpm_tar/deb_tar.tar
-r-xr-xr-x 0/0           10662 2000-01-01 00:00 debpackage-0.0.0.deb
$ tar tvf bazel-out/k8-fastbuild/bin/tests/deb_rpm_tar/rpm_tar.tar 
-r-xr-xr-x 0/0            6254 2000-01-01 00:00 rpmpackage-0.0.0-1.rpm
-r-xr-xr-x 0/0            6254 2000-01-01 00:00 test_rpm.rpm

[aiuto]

I think the root cause is reallypackage_file_name = "rpmpackage-0.0.0-0.rpm".
That will give you an implicit output of a real file "rpmpackage-0.0.0-0.rpm"
and a default output of the symlink "test_rpm.rpm". This is an intentional behavior:

• the physical file has a name containing the right (possibly dynamically determined) version numbers
• rules depending on the rule name can pick up the rpm at a well known name.

The same kind of thing happens for pkg_deb(). Anyone depending on just the name gets all output files.

If you need to get a specific single file, you can do the file_group+output_group redirection as seen here.

rules_pkg/examples/where_is_my_output/BUILD

Line 45 in edd4d3d

filegroup(

So in #684 you have the core of a correct change, except the tests don't test the right thing - the number of files is variable, not fixed. What it should be is that the expected output groups are created in particular situations. That should be an analysis test or two. I can comment more about how do to that later. But now I ran out of time on rules_pkg for the week.

[k0walik]

I think this is still inconsistent behaviour between pkg_deb and pkg_rpm rules. If the output should contain both package_file_name.rpm and symlink to rule_name.rpm, then the pkg_deb should be fixed to include the symlink as well - not just the package_file_name.deb. See the return value implementation here: https://github.com/bazelbuild/rules_pkg/blob/main/pkg/private/deb/deb.bzl#L168
That being said: can I piggy back it on #684 or do we want another change to address this?

As for the testing part I'll update my change to test it correctly.

[aiuto]

I stand corrected. The DefaultInfo for the other rules is to have only the single file, not both.
Forget the tests, they should fold into the existing file naming tests and or be analysis tests.
Changing RPM to agree with the other rules should be a separate change.

[k0walik]

Fixed in #692