Netflix 使用 RAW 空心内存对象存储对 Tudum 的 CQRS 架构进行了改造。

发布于 8 月 15 日

Netflix Tudum Architecture Changes: Netflix replaced its CQRS implementation using Kafka and Cassandra with a new solution leveraging RAW Hollow. Tudum is Netflix's official fan website launched in late 2021.
- Initial Architecture: Based on CQRS pattern with write part around 3rd-party CMS product and dedicated ingestion service. Read-optimized data published to Kafka topic and consumed by page data service in Cassandra. Had a near cache for performance.
- Issues: Caching refresh cycle led to CMS updates taking many seconds to show on website, causing problems for content editors to preview modifications.
RAW Hollow: An in-memory object store developed internally by Netflix. Designed to handle small to medium datasets with strong read-after-write consistency. Allows entire dataset to be in memory of each application process, offering low latency and high availability. Default is eventual consistency but supports strong consistency at individual request level.
New Architecture: Tudum engineers replaced Kafka and Cassandra with the RAW Hollow cluster. Concluded CQRS wasn't optimal and distributed, in-memory object store suited better. New solution eliminated cache invalidation problems and reduced data propagation and page construction times due to reduced request I/O and round-trip times.

阅读 10