Building JVM RPC tooling around Bloomberg's needs

The JVM community at Bloomberg had tried adapting entire open-source solutions to our internal technologies. That brought incompatibilities and performance issues. Some of the foundational RPC libraries, including codecs, were unmaintained, with bugs that had gone unfixed for years.

I developed a Kotlin-based JVM code generator and codec around Bloomberg’s requirements. Better type safety, performance, and cross-language compatibility were the goals. There were concrete problems to work on: the old XJC-based tooling could not represent xs:choice as sealed type hierarchies, and the old BER codec had subtle behavioral differences from the C++ and Python implementations.

Choosing what to reuse

Bloomberg’s JSON, XML, and BER behavior did not fit the approach of reusing an entire stack. That did not mean every useful piece of open-source software had to go.

The new generator supports Jackson and kotlinx.serialization. KotlinPoet handles Kotlin source generation. Jackson gives me a framework for implementing Bloomberg data formats, while kotlinx.serialization provides a serialization framework with custom encoders and decoders. Those are useful foundations on which to implement the behavior our environment needs.

The tooling is in limited beta and moving toward production.

A better boundary

This project expanded my understanding of code generation, schemaful and schemaless formats, serialization, wire formats, and performance optimization at both the language and hardware level.

What I took away is a more deliberate approach to reuse. Keep the frameworks that help, implement the behavior specific to the environment, and be willing to inspect or adapt open-source code with appropriate attribution and licensing. Blind reuse and blind rewrites both miss that boundary.