Worked on stabilizing vector rescoring workflows in the elastic/elasticsearch repository, focusing on serverless deployment environments. Addressed a critical bug by changing the default configuration to disable direct IO for vector rescoring, preventing BBQ shard read failures when direct IO is unavailable. Used Java and backend development skills to implement and validate the change across various deployment contexts, ensuring compatibility and maintaining performance expectations. Documented the update and prepared it for mainline integration to support more reliable serverless operations. The work emphasized risk mitigation and operational stability, contributing to improved reliability for vector rescoring in serverless Elasticsearch deployments.