TechPrufer
Get started
Confluent
Confluent logoConfluent · Feature trail

Flink AI Model Inference

22 posts · Sep 2016 – Aug 2026

What this trail covers

Confluent Cloud for Apache Flink®️ now supports AI model inference, allowing models to be used as resources in Flink SQL. This enables real-time data enrichment by invoking generative AI (GenAI) large language model (LLM) endpoints (e.g., OpenAI, Azure OpenAI, Amazon Bedrock, Google Vertex AI) and custom ML/deep learning models directly from Flink SQL statements for tasks like classification, text generation, and clustering. This integration allows for real-time ML predictions to be applied to streaming data, enhancing applications with context and enabling smarter decision-making. The feature is currently in early access.

All Confluent trails
  1. Flink: Programmatic Capabilities for Developers and dbt/SQL-Native for Data PractitionersThis post announces the General Availability of the Flink Table API in Java and Process Table Functions (PTFs) in Java for developers, enabling code-first development of sophisticated Flink applications with serverless operational simplicity. It also expands User-Defined Function (UDF) support to Python on AWS and Java on GCP. For analytics engineers, it highlights the General Availability of the dbt adapter and Materialized Tables for managing Flink pipelines via dbt, and introduces Snapshot Queries for unified batch and streaming analysis. For AI, it details Flink AI Model Inference for calling remote model endpoints and the introduction of Streaming Agents for event-driven agents on Flink. The post emphasizes the co-design of Kafka and Flink, sharing unified governance, security, and the Kora storage engine. It also introduces Flink Advantage Packs for predictable pricing.
  2. Add Your First Flink Streaming TransformationThis post introduces the concept and practical steps for adding simple, stateless streaming transformations to existing Kafka streams using Apache Flink SQL. It details how to identify source Kafka streams, define input schemas, apply basic transformations like filtering and field projection, and observe the output. The post emphasizes an incremental approach, contrasting streaming with batch processing and highlighting the benefits of real-time data processing for reducing latency and simplifying architectures. It also outlines common first transformations such as event filtering, field projection/reshaping, real-time enrichment, and introductory event aggregation.
  3. Focal Systems: Boosting Store Performance with an AI Retail Operating System and Real-Time DataThis post details how Focal Systems uses Confluent Cloud's managed Kafka and serverless Flink for real-time stream processing in their AI retail operating system. They leverage Flink SQL and table abstractions mapped to Kafka topics to compute per-store, per-camera metrics from computer vision and retail data. This processing transforms raw detections into actionable insights like shelf availability, out-of-stock events, and planogram compliance signals. The Flink pipeline also handles pre-AI data readiness steps such as PII removal, masking, OCR, and feature extraction, ensuring data is privacy-safe and model-ready before reaching downstream systems. This architecture replaces custom, unscalable frameworks with a managed, cloud-native solution, accelerating development and enabling real-time decisioning.
  4. Detecting the Unexpected: Built-in Real-Time Anomaly Detection in Confluent Cloud for Apache Flink®Introduces the `ML_DETECT_ANOMALIES` function in Confluent Cloud for Apache Flink®, which uses the ARIMA algorithm for real-time anomaly detection. Details the ARIMA components (AR, I, MA) and the auto-ARIMA implementation. Explains configuration options like Seasonal Pattern Detection (STL), Adaptive Training Windows (`minTrainingSize`, `maxTrainingSize`, `updateInterval`), and Confidence-Based Detection (`confidencePercentage`). Provides real-world use cases in financial services, retail, IoT, SaaS, and supply chain. Includes a SQL example for using the function and discusses its billing model (CFUs). Mentions future plans for multi-variate anomaly detection.
  5. Why Flink Agents Are the Future of Enterprise AIThis post introduces Flink Agents, a new sub-project within Apache Flink, developed collaboratively by Confluent and Alibaba. It outlines the motivation for integrating AI agents into stream processing, highlighting the limitations of traditional chatbot-style AI and the need for event-driven, autonomous systems in enterprise contexts. The post details four key areas of focus for Flink Agents: agent semantics, dynamic topology support, enhanced observability for agent decision-making, and Model Context Protocol (MCP) support. It provides a code example demonstrating how developers can define and apply agents using Flink's Table API. The contribution emphasizes the benefits of this integrated approach, including unified infrastructure, end-to-end consistency, built-in fault tolerance, seamless integration, and replayability, positioning Flink Agents as a solution for production-ready AI agent development.
  7. Confluent Cloud Q2 '25 unifies stream and batch processingThis post introduces Snapshot Queries for Apache Flink® in Confluent Cloud, enabling batch-style queries on Kafka topics and Tableflow data (Iceberg/Delta Lake) using Flink SQL. This feature allows for ad hoc, point-in-time queries and accelerates pipeline development and debugging by running queries in batch mode (SET 'sql.snapshot.mode' = 'now'), delivering results up to 50-100x faster than streaming jobs over historical data. It also announces UDF support in Java for Flink on Azure, enhanced Confluent for VS Code extension for Flink SQL development, Tableflow support for Delta Lake tables, CCN routing and IP Filtering for Flink and Schema Registry, AWS PrivateLink for Schema Registry, Enterprise clusters on Google Cloud, and cross-cloud Cluster Linking between AWS and Azure.
  8. Using Apache Flink for Model Inference: A Guide for Real-Time AI Applications | ConfluentThis post details the implementation of remote model inference with Apache Flink on Confluent Cloud. It introduces the concept of 'model' as a first-class citizen in Flink SQL and Table APIs, enabling registration of remote AI models (e.g., Azure OpenAI, AWS SageMaker) via connection configurations and SQL DDL statements. It demonstrates the use of the `ML_PREDICT` function for tasks like generating vector embeddings for RAG. The post outlines key steps: data ingestion into Kafka, configuring API calls to model servers using Flink's asynchronous I/O, optimizing network efficiency through batching and efficient data encoding, and independent scaling and monitoring of model servers. It also provides best practices for error handling, model drift monitoring, and cloud resource optimization.
  9. Confluent Platform 7.8: Confluent Platform for Apache Flink® (GA), mTLS Identity for RBAC Authorization, and MoreThis post announces the general availability of Confluent Platform for Apache Flink, providing an enterprise-grade Flink distribution and control plane for on-prem deployments. It highlights Confluent Manager for Apache Flink (CMF) for Kubernetes-native job lifecycle management, RBAC authorization with mTLS identity for enhanced security, and improvements to Confluent Control Center for message browsing and schematized message production. Additional platform support for Ubuntu and Alma Linux, KRaft mode enhancements, and Confluent for Kubernetes updates are also detailed.
  10. Explore the Flink Table API: Streamline Data Processing with EaseThis post introduces and explains the Apache Flink Table API, highlighting its unified relational approach for stream and batch processing. It details how the Table API integrates with other Flink APIs (ProcessFunction, DataStream API, Flink SQL) and provides practical Java and Python examples for creating tables, performing filtering, grouping, and aggregation operations. The post also covers unit testing strategies for Table API applications and emphasizes its suitability for serverless deployments on Confluent Cloud.
  11. Accelerating Insurance Claims With a Stream Processing ArchitectureThis post details a practical application of Confluent's data streaming platform, specifically Apache Flink on Confluent Cloud, for accelerating insurance claims processing. It outlines a 5-step architecture for real-time claims, emphasizing the use of Kafka topics for event ingestion (FNOL), data enrichment via temporal joins with customer and policy data streamed from Oracle CDC, and Flink for data filtering and routing. The post highlights the benefits of stream processing over batch jobs for improved customer experience, reduced risk of inaccurate assessments, and enabling parallel processes like fraud detection and ML model training. It also showcases Confluent's managed connectors and Flink SQL capabilities.
  12. Flink 1.20 Adds Significant Improvements Along w/DeprecationsFlink 1.20 introduces the DISTRIBUTED BY clause (FLIP-376) for unified bucketing strategies in Flink SQL, replacing connector-specific syntax. It also adds MATERIALIZED tables (FLIP-435) to manage data freshness requirements directly within Flink, deciding between streaming or batch refresh modes based on configuration. The DataStream API now supports parallel aggregations on non-keyed streams in batch mode (FLIP-380). Operational improvements include unified file merging for checkpoints (FLIP-306) to reduce blob storage load and costs, and enhanced recovery for batch jobs from JobMaster failures (FLIP-383). Legacy SinkFunction API and other configuration options are deprecated in preparation for Flink 2.0.
  13. Flink SQL on Confluent Cloud Invokes ML Model & AI InferenceThis post details the integration of AI model inference into Flink SQL on Confluent Cloud, specifically demonstrating how to invoke an OpenAI GPT-4o model for sentiment classification of product reviews. It outlines the steps involved: creating a Kafka topic, producing Avro-serialized review data, defining a Flink SQL remote model resource pointing to the OpenAI API with a specific system prompt and task ('classification'), and then using the `ML_PREDICT` function within a Flink SQL `INSERT INTO` statement to populate a downstream `product_ratings` table with sentiment ratings. The post provides code excerpts for the Python producer and Flink SQL statements, and discusses the benefits of real-time ML enrichment for streaming data.
  14. Confluent Cloud for Apache Flink Now Generally AvailableThis post announces the general availability of Confluent Cloud for Apache Flink. It highlights the serverless nature of the offering, including elastic autoscaling, always-updated runtime, and usage-based billing. It details the various ways to interact with Flink (CLI, SQL Workspaces, Data Portal, Flink Actions) and mentions early access to Java UDFs. The post emphasizes cloud-native reliability with exactly-once processing and consistent checkpoints. It also covers production operations through programmatic deployments (REST API, Terraform) and integration with monitoring tools. The seamless integration with Kafka and Schema Registry, along with Stream Lineage, is also a key focus.
  15. Flink Updates: Topic Actions, Terraform support, GCP and Azure AvailabilityIntroduces 'Topic Actions' for Apache Flink, a feature that simplifies common stream processing tasks like topic deduplication and field masking by generating Flink SQL templates. Details the internal mechanism of Actions, which are templatized SQL statements submitted to Flink. Also announces Terraform support for deploying Flink statements and expanded availability of the managed Flink service to GCP, in addition to existing AWS and Azure support.
  16. Serverless Stream Processing With Apache Flink®This post details the serverless architecture of Confluent Cloud for Apache Flink, focusing on elastic autoscaling with scale-to-zero, evergreen runtime and APIs, and usage-based billing. It explains the use of compute pools for resource isolation and dynamic scaling, and introduces a proprietary algorithm based on the DS2 algorithm for fine-grained scaling of Flink SQL statements. The post also discusses fast rescaling techniques leveraging Flink's Adaptive Scheduler and Declarative Resource Management, and highlights cost-effectiveness through usage-based billing and compute pool pricing.
  17. Data Streaming and AI are Better When They’re TogetherThis post elaborates on the synergy between data streaming and AI/ML, highlighting how Confluent Cloud, particularly with its managed Apache Flink service, streamlines the integration of streaming data into AI/ML models. It details how Confluent Cloud enables embedding ML models in event streams for reduced latency, performing simulations and analysis with real-time data, and supporting asynchronous communication for AI/ML applications. It also mentions the deployment of AI/ML models at the edge using librdkafka and on mobile applications. The post emphasizes Confluent Cloud's capabilities in autoscaling, observability, and automatic model updates, and how Flink optimizes resource consumption for low latency and fault tolerance, facilitating real-time AI/ML pipelines.
  18. Using Apache Flink SQL to Build Real-Time Streaming AppsThis post introduces SQL Workspaces, a new browser-based interface for Apache Flink SQL on Confluent Cloud. It details how Confluent Cloud environments, Kafka clusters, and topics are mapped to Flink's catalog, database, and table abstractions, respectively. The post highlights the ability to write ANSI-standard SQL against streaming data, the support for perpetual SQL statements with real-time output updates, and the UX improvements of multiple editors within a single workspace for development. It also describes the inline runtime metrics panel for monitoring streaming statement performance and contextual entry points from Kafka topic views and Data Portal to launch SQL Workspaces.
  19. Choosing a Data Streaming Platform and Stream Processing EngineThis post discusses the importance of data streaming platforms and stream processing engines, highlighting Apache Flink and Kafka Streams as major players. It details Flink's capabilities for large-scale stream processing with fault tolerance and its use by enterprises for critical use cases. It also contrasts Flink with Kafka Streams, noting Flink's richer API, lower latency, and support for CEP/ML, while acknowledging Kafka Streams' advantage of not requiring extra infrastructure when tightly coupled with Kafka. The post emphasizes the complexity of building robust streaming frameworks and the need for specialized expertise, advocating for managed services to reduce complexity and accelerate time to insight and market. It introduces Confluent Cloud's Kora Engine as a cloud-native data streaming platform that abstracts infrastructure and operations, and mentions the acquisition of Immerok to enhance Flink capabilities within Confluent Cloud.
  20. Fully Managed Apache Flink and Cost Effective, Secure Data Pipelines With ConfluentThis post announces the open preview of Confluent Cloud for Apache Flink, a fully managed, serverless stream processing service. It highlights the benefits of using Flink for filtering, joining, and enriching data streams without the complexities of self-management. The integration with Kafka provides a unified platform for monitoring, security, and governance. The post also mentions improvements to Kora Engine, cost savings through Enterprise clusters with private networking (AWS PrivateLink), and updates to the Confluent Terraform provider including HashiCorp Sentinel integration, Data Catalog support, and a Resource Importer. Data Portal for Stream Governance is announced as coming soon. Bidirectional Cluster Linking is also enhanced.
  21. Introducing Confluent Cloud for Apache FlinkThis post introduces Confluent Cloud for Apache Flink, a fully managed, serverless stream processing service. It details the cloud-native architecture, focusing on serverless principles: elastic autoscaling with scale-to-zero, evergreen runtime and APIs, and usage-based billing. It introduces Flink compute pools as a mechanism for managing elastic compute resources. The post highlights the simplification of Flink SQL integration with Kafka and Schema Registry, and the ability to scale Flink independently from Kafka, enabling cross-cluster data joining.
  22. Dataflow Programming: Building a Java Pipeline With Flink and KafkaThis post details building a Java pipeline with Apache Flink and Kafka to process clickstream data. It covers setting up a Maven project with Flink dependencies, creating a Flink uber JAR, defining POJOs for data serialization (ClickStreamRecord and ClickStreamAnalytics), creating a Flink entry point, and configuring the stream execution environment. The pipeline consumes JSON records from a Kafka topic, deserializes them, groups records by key and time window (1 minute), and computes success/failure counts for each request type, producing the results to a separate Kafka topic. It highlights the use of Flink's POJO serializer and the importance of efficient data structures for performance.
  23. Getting Started with Flink SQL: In-Depth GuideThis post introduces Flink SQL as a declarative way to process and analyze data streams in real time, supporting ANSI SQL standards, user-defined functions, and a rich ecosystem of tools like JDBC Driver and SQL Gateway. It details dynamic tables, continuous queries, changelog types, and provides examples of Flink SQL syntax for connecting to Kafka, performing joins, aggregations, and windowing operations. The post emphasizes the interoperability of Flink APIs and the benefits of using Flink SQL for building real-time data products and pipelines.
  24. Apache Flink: Stream Processing for All Real-Time Use CasesThis post details how Apache Flink supports three key categories of use cases: event-driven applications, real-time analytics, and streaming data pipelines. For event-driven applications, it highlights fraud detection, anomaly detection, alerting, real-time routing, business process monitoring, and IoT applications, emphasizing Flink's customizable window logic, per-event stateful processing, and Complex Event Processing (CEP) library with MATCH_RECOGNIZE. For real-time analytics, it covers ad/campaign performance, content performance, telco network quality monitoring, and pattern analysis, showcasing Flink's low-latency processing, Flink ML library, time-series analysis capabilities via CEP, and materialized views. For streaming data pipelines, it discusses ETL/ELT, data transformation, and data enrichment, emphasizing Flink's ability to handle large-scale data movement and transformation. The post also links to related resources on Flink SQL and Confluent Cloud for Flink.
  25. Stream Processing with Apache FlinkThis post introduces Apache Flink as a stream processing framework that enhances Kafka use cases. It details Flink's powerful runtime, its layered APIs (DataStream, Table API, Flink SQL, Stateful Functions) with support for Python, Java, and SQL, and its unified stream and batch processing capabilities. The post highlights Flink's production-readiness features like flexible metrics and observability. It also announces the 'Inside Flink' blog series and an early access program for Confluent Cloud for Apache Flink, emphasizing its integration with AI model inference.
  26. Uniting Machine Learning and Data Streaming Technologies Part 2This post explores the integration of machine learning (ML) with data streaming technologies, focusing on bridging the JVM and Python ecosystems. It discusses the role of SQL (ksqlDB, Flink SQL) as a declarative language for building data pipelines and the limitations of SQL's composability. The post details strategies for inter-language communication, such as PyFlink and PySpark using Py4J, and Flink's process and thread modes. It also examines current integration patterns with ML experiment management platforms and provides a Python client code example for building a real-time ML-powered microservice using Confluent Python Client.
  27. Cloud Native Kafka + Flink with Confluent and ImmerokThis post announces the acquisition of Immerok, a startup with a fully managed Apache Flink service, to be integrated into Confluent Cloud. It outlines Confluent's strategy to make stream processing as easy as batch processing by offering a managed Flink service alongside Kafka. The initial launch will focus on SQL support, with plans to expand to the full Flink platform. The goal is to create a seamless integration between Kafka and Flink, simplifying their operation and management for users.
  28. Flink vs Kafka Streams: A Complete ComparisonThis post provides a detailed comparison between Apache Flink and Kafka Streams, focusing on their architectural differences and ideal use cases. It elaborates on Flink's cluster-based deployment, job lifecycle management, and its capabilities for high throughput, low latency, and exactly-once semantics. It contrasts this with Kafka Streams, which is presented as an embeddable library that leverages Kafka for coordination, load balancing, and fault tolerance, simplifying stream processing for application developers. Key differences in deployment, lifecycle, coordination, data sources/sinks, and semantic guarantees are outlined.