Replacing a Kafka and Avro pipeline with direct Vault reads in Apache Arrow format took throughput from about 50K to about 600K rows per second.