Describe how you would design a Kafka + Spark streaming pipeline to process events with exactly-once semantics, tolerate broker or executor failures, and allow safe schema evolution. What components, configuration choices, and operational checks would you include?

I’d build the pipeline on Spark Structured Streaming reading from Kafka, use durable checkpointing on S3/HDFS, and push outputs via transactional or idempotent sinks. Practically I used this pattern for a system that handled ~50k events/s: Spark reads with Kafka source, stores offsets in the checkpoint, and I use foreachBatch to write atomically — either Kafka transactions (producer.enable.idempotence=true, transactional.id) or a database sink that accepts dedup keys. Cluster resilience comes from Kafka replication factor=3, min.insync.replicas=2, and Spark checkpointing plus dynamic allocation with retries so executor failures simply restart tasks. For schema evolution I used Avro + Confluent Schema Registry, enforcing backward/forward compatibility and adding fields with defaults. Operational checks: consumer lag, under-replicated partitions, end-to-end latency (SLO: 99th percentile <200ms), and automated alerts via Prometheus/Grafana. CI in Git, containerized with Docker, and orchestrated with Airflow.