Скачать или смотреть Netflix Tudum Architecture: From CQRS + Kafka to RAW Hollow

This episode breaks down how Netflix’s Tudum team rethought the read path for 20M+ users. The original CQRS design used an ingestion pipeline and a read store behind a near‑cache, which was great for scale but painful for editor previews due to refresh cycles and I/O hops. The redesign replaces network reads with RAW Hollow—a compressed, distributed, in‑memory object store embedded in each service—so page construction, search, and personalization fetch data in O(1) from local RAM. Learn how this cut homepage construction from ~1.4s to ~0.4s, enabled near‑instant previews with per‑request read‑after‑write consistency, and simplified the stack by removing the read‑path cache layer, external KV store, and message fan‑out.

What you’ll learn:

Why CQRS helped scale, but cache refreshes hurt preview latency

The bottleneck chain: CMS → ingest → Kafka → read DB → near‑cache timers

RAW Hollow fundamentals: compressed, co‑located, embedded clients, O(1) reads

New flow: “load once into memory → serve instantly → propagate updates”

Practical takeaways: minimize sequential I/O, co‑locate hot data, use selective consistency

Chapters:
0:00 Intro
0:22 The original CQRS + Kafka + read store + near‑cache
1:05 Why previews lagged (per‑key refresh, fragmented pages)
1:38 RAW Hollow overview and how it removes I/O hops
2:10 New architecture: embedded clients, strong preview reads
2:40 Results and lessons for similar content platforms