Incrementally deploying ranking code at Glean

This post details the implementation of a dynamic ranking code deployment system for Glean's search engine. The core innovation is the use of Java's dynamic class loading to load JAR archives from Cloud Storage without restarting the Index Server. To handle multiple versions of ranking code, class names are remapped to include release tags (e.g., `com.glean.ranking.release123.TermInfo`). This allows for more frequent releases (nightly for internal, weekly for customers), faster developer experimentation, and instant rollback via configuration changes. The system addresses challenges like initial JAR decoding latency (around 1 second) by using warmup requests and caching subsequent loads.