JVM Optimization: A VarInt Tale of Unexpected Results

While optimizing Java code in a massive distributed data processing platform, the author identified VarInt encoding as a potential optimization target. He wrote a highly optimized VarInt encoder using SIMD instructions, achieving a 4x speedup in benchmarks. However, surprisingly, the optimization yielded no improvement in production. The culprit? The benchmark used random numbers, while real-world numbers tend to be much smaller, rendering the algorithm's worst-case performance irrelevant in practice. The change was ultimately reverted, but the experience served as valuable proof-of-concept for developing and productionizing custom JIT optimizations.