Hi all,
A while ago we had pretty good consensus that we should vendor
("pre-shade") specific dependencies, and there's start on it here:
https://github.com/apache/beam/tree/master/vendor
IntelliJ notes:
- With shading, it is hard (impossible?) to step into dependency code in
IntelliJ's debugger, because the actual symbol at runtime does not match
what is in the external jars
- With vendoring, if the vendored dependencies are part of the project
then IntelliJ gets confused because it operates on source, not the produced
jars
The second one has a quick fix for most cases*: don't make the vendored dep
a subproject, but just separately build and publish it. Since a vendored
dependency should change much more infrequently (or if we bake the version
into the name, ~never) this means we publish once and save headache and
build time forever.
WDYT? Have I overlooked something? How about we set up vendored versions of
guava, protobuf, grpc, and publish them. We don't have to actually start
using them yet, and can do it incrementally.
(side note: what do other projects like Flink do?)
Kenn
*for generated proto classes, they need to be altered after being generated
so shading happens there, but actually only relocation and the shared
vendored dep should work elsewhere in the project
