Training-serving skew — the divergence between features used during model training and those seen at inference time — silently degrades ML accuracy and doubles infrastructure costs. The solution is a unified kappa architecture: compute features once in Apache Flink, dual-write to an offline store (Apache Iceberg or Delta Lake) for training and an online key-value cache for serving. DoorDash measured a 35.7% feature-value mismatch in their dual-pipeline setup; Netflix replaced a $93M/year dual-pipeline backfill with a $2M/year kappa replay. The reference architecture covers Kafka ingestion via Confluent's Kora engine, serverless Flink with event-time watermarks and exactly-once semantics, Tableflow for automated Iceberg/Delta materialization, and Stream Governance for schema enforcement and lineage. A tooling comparison covers Databricks, SageMaker+Kinesis, Tecton, Feast, and Confluent, with a decision framework based on latency requirements, existing stack investment, and pipeline fragmentation. The post is authored by a Confluent employee and promotes the Confluent Data Streaming Platform throughout.
