About iQIYI
iQIYI is an online entertainment platform in China that produces and distributes video content through subscriptions, advertising, and content licensing.
Moving state out of Flink jobs
Large Flink state can slow recovery and make data difficult to inspect. iQIYI uses Fluss primary-key tables as a schema-enforced state store. In its reported production workload, primary-key lookups achieve P99 latency below one millisecond at 100,000 queries per second.
Replaying historical features
Fluss log tables provide the streaming source for long-term ML feature backfills. Longer retention and column pruning let jobs replay the fields they need. Combined with the disaggregated state architecture, iQIYI reports that a ten-tablet-server cluster handles up to one million queries per second for backfill workloads and can backfill one day of data in about thirty minutes.
Serving hot and cold feature data
For real-time feature enrichment, iQIYI combines Fluss primary-key tables with Paimon lake tiering and fallback lookups for cold data. The team reports ingestion of up to 20 GB/s and approximately 50% lower cost than its previous HBase-based solution.