Kenneth Knowles created BEAM-3204:
-------------------------------------
Summary: Coders only should have a FunctionSpec, not an
SdkFunctionSpec
Key: BEAM-3204
URL: https://issues.apache.org/jira/browse/BEAM-3204
Project: Beam
Issue Type: Improvement
Components: beam-model
Reporter: Kenneth Knowles
Assignee: Henning Rohde
We added environments to coders to account for "custom" coders where it is only
really possible for one SDK to understand them, like this:
{code}
Coder {
spec: SdkFunctionSpec {
environment: "java_sdk_docker_container",
spec: FunctionSpec {
urn: "beam:coder:java_custom_coder",
payload: <serialized java bytes>
}
}
}
{code}
But a coder must be understood by both the producer of a PCollection and its
consumers. A coder is not the same as other UDF, though these are user-defined.
A pipeline where either the producer or consumer cannot handle the coder is
invalid, and we will have to build our cross-language APIs to prevent
construction of such a pipeline. So we can drop the environment.
I think there are some folks who want to reserve the ability to add an
environment later, perhaps, to not pain ourselves into a corner. In this case,
we can just add a field to Coder.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
