Apache Flink 2.2.0 has launched, bringing significant advancements in real-time data processing and artificial intelligence integration. The update highlights include the introduction of ML_PREDICT for large language model inference and VECTOR_SEARCH for streaming vector similarity searches, enhancing Flink’s capabilities for AI-driven applications. This version includes contributions from 73 global contributors, implements nine Flink Improvement Proposals (FLIPs), and resolves over 220 issues. Flink 2.2 expands the Table API to support machine learning model inference. Users can create models from providers like OpenAI and integrate them within data processing pipelines. The VECTOR_SEARCH function allows for real-time querying and similarity analysis, addressing previous limitations in handling unstructured data. The introduction of Materialized Tables aims to streamline both batch and stream data processes, allowing users to specify data freshness and automatically generate corresponding refresh pipelines. Other notable improvements include an optimized SinkUpsertMaterializer operator for handling out-of-order changelog events, enhanced delta join capabilities for better performance with change data capture sources, and new SQL type handling that respects NOT NULL constraints. Balanced task scheduling has been introduced to improve resource management during processing, aiming for greater reliability and performance in data pipelines.