The numbers
| Metric | Cassandra (before) | ScyllaDB (after) |
|---|---|---|
| Cluster size | 177 machines (9:41) | 72 machines (17:56) |
| Tail latency (slow reads) | swinging 40 → 120+ ms (11:36) | steady 15 ms (18:06) |
| Writes | — | calm, predictable 5 ms (18:06) |
| Load handled | — | more data and more traffic on fewer than half the machines (18:06) |
- The recurring words are steady and predictable — precisely the wishlist item from years earlier: they did not just want fast, they wanted consistent, and they finally got it (
18:06,18:28).
The result that no graph shows
The system went quiet (18:28):
- No more weekend firefights.
- No more manually nursing sick nodes back to life.
- No more endless tuning sessions just to keep things upright.
The real test: 2022 Football World Cup final
- Traffic spiked visibly every time a goal went in — huge, sudden surges of people reacting at once (
18:28,18:53). - The old system would have buckled, with engineers watching dashboards instead of the match (
18:57). - The new system just handled it — the video's definition of what good infrastructure looks like (
18:57).
Closing thought
Behind one simple screen of instantly-appearing messages sit trillions of messages working quietly (18:57). The author offers a follow-up walkthrough covering message fan-out, notifications, presence, voice infrastructure, caching and scaling if requested in the comments (19:20).