Support sink schema sidecar

[jiangpengcheng]

Summary

This adds explicit sink schema sidecar support to the Go launcher SDK. Generic executable functions return raw bytes, so the runtime needs a sidecar schema definition to create schema-aware Pulsar or Kafka producers without asking the user function process to decode and re-encode records.

The new API exposes pf.SinkSchema, pf.WriteSinkSchemaSidecar, and pf.StartWithSinkSchema. StartWithSinkSchema writes sink.schema.json next to the executable before entering the normal Start loop, matching the generic runtime sidecar lookup path. The sidecar payload uses the shared fields consumed by the runtime: schemaType, name, schemaData, and properties. Empty or bytes schema types are skipped because those do not require a producer schema.

Validation

• go test -count=1 ./...
• go vet ./...
• go test -count=1 ./... in examples
• go vet ./... in examples

[freeznet]

Review: Support sink schema sidecar

Thanks for this — the approach is clean and lines up with how Start already derives the working directory (filepath.Dir(os.Executable()) in pf/function.go), so the sidecar lands next to context.sock exactly where the generic runtime looks. Decoupling the public SinkSchema from the private sinkSchemaSidecarPayload wire struct is a good call, and pinning the always-emit-{} properties behavior with a test is nice.

A few change requests / suggestions (details inline):

1. Missing Apache license header on both new files — every other file in pf/ (function.go, context.go, log.go) carries the ASF header. This is the main must-fix (license/RAT consistency).
2. none schema type isn't skipped — only "" and bytes are. Pulsar's SchemaType.NONE is also "no schema", so a none caller would still get a sidecar written.
3. No normalization / exported constants for schemaType — the value is written verbatim, so correctness depends entirely on the caller matching the runtime's expected vocabulary.
4. Stale sidecar isn't removed on the skip path (low severity if work dirs are always ephemeral).
5. Test coverage — add the empty-string skip case and an uppercase BYTES case to lock in the EqualFold boundary.

None of these are happy-path correctness bugs; #1 is the primary blocker.

[freeznet]

Re-review: commit 70f20aa "Address schema sidecar review feedback"

Thanks for the quick turnaround — all five items from the previous pass are addressed:

• ✅ Apache license header on both files
• ✅ none treated as no-schema (via isNoSchemaType)
• ✅ Stale sidecar removed on the skip path, with the correct !os.IsNotExist guard
• ✅ Exported SchemaType* constants + doc on SinkSchema.SchemaType
• ✅ Table-driven tests for empty / bytes / BYTES / none / NONE + stale removal + normalization

Two follow-ups from the new commit (details inline):

1. Casing of the emitted schemaType (please confirm). normalizeSinkSchemaType now lowercases the value that is written to the sidecar, and TestWriteSinkSchemaSidecarToDirNormalizesSchemaType locks "JSON" → "json" in. Pulsar's canonical SchemaType enum names are UPPERCASE and SchemaType.valueOf(...) is case-sensitive. If the generic runtime parses the string that way, lowercasing breaks every type — and it's also a regression vs. the first version, which wrote the value verbatim (a caller passing "JSON" used to work). If the runtime is case-insensitive / expects lowercase, all good — I just want it confirmed, since the blast radius is every schema.

2. SchemaTypeBool = "bool" looks wrong. Pulsar's boolean type is SchemaType.BOOLEAN → "boolean", independent of the casing question above.

Everything else looks good.