如何在自定义 Bazel 规则或 genrule 中 package 生成 python 文件？

Question

Let's say I have a scripts which generates Java and Python code. The files it generates depends on the input config, it can vary based on the config. Now, in the case of Java, I can write a simple genrule, point to invoke the script and then call jar command to generate a single source package of the generated Python files.

Eg:

genrule(
    name = "generated_java_srcjar",
    srcs = glob([
        ":input-files"
    ]),
    outs = [
        "xsd.srcjar"
    ],
    cmd = """
        mkdir -p $(GENDIR)/generated-code
        $(location path/to/custom/script) --output-directory $(GENDIR)/generated-code --input $(locations :input-files)
        $(JAVABASE)/bin/jar cMf $@ -C $(GENDIR)/generated-code .
    """,
)

java_library(
    name = "generated_java_jar",
    srcs = [
        ":generated_java_srcjar"
    ],
)

In the Java example, to declare an output for the genrule, I'm able to generate a single output (source package of the generated files) and then use the same rule as a source in the java_library rule.

Now, I want to achieve something similar in Python. How can I package the generated Python files after invoking the custom script so that I can then pass it as a source to py_library rule?

Answer 1

One way to do this is to write your own Starlark rule that returns a PyInfo provider with your generated files in the transitive_sources . You'd then add the rule to the deps attribute of your py_library rule.

The drawback is that you have to know at loading phase (so before the code generation has happened) what files will be generated by your tool, no cheating with zips/jars.

If that's not an option I could think of a hacky workaround where you produce a zip that you depend on in your py_libarary's data deps. You then unpack the file at runtime before you actually import it.

Depending on how you've set up your python rules you might also be able to build a python wheel in your code generator and depend on that.