Details
-
Improvement
-
Status: Open
-
P3
-
Resolution: Unresolved
-
None
-
None
Description
Problem/Status
The schema proto has an encoding_position field that is currently unused in every row coder implementation. The intention of this field is that it indicates an alternative order for the fields to be encoded in by beam:coder:row:v1 implementations. Currently all the implementations ignore this field, and always encode the fields in the order that they appear in the schema.
Motivation
The idea with the encoding position is that it will give runners a way to enforce schema compatibility (BEAM-9502), by re-ordering the way fields are encoded when the schema changes between two job submissions. Schema changes could be due to fields re-ordering, or field additions/deletions.
Code pointers
The Python beam:coder:row:v1 implementation lives in row_coder.py
The Java implementation is a little more complicated, distributed between SchemaCoder, RowCoder, and RowCoderGenerator. RowCoderGenerator contains the code relevant to this jira - it uses bytebuddy to generate Java code for the coder. We need it to generate code that puts fields in the order specified by encoding_position.
Testing
Python and Java implementations should be tested with unit tests (RowCoderTest, row_coder_test). We should also test them for compatibility by adding test cases that exercise the encoding_position in standard_coders.yaml. These tests will be executed by CommonCoderTest and standard_coders_test. There's some example code for generating a new test case here.
Attachments
Issue Links
- is related to
-
BEAM-12198 Support Dataflow update when schemas are used
-
- Triage Needed
-
-
BEAM-9502 SchemaCoder is not update compatible
-
- Resolved
-
- relates to
-
-
- Open
-
- links to
