Monster Scale Summit

Maciej Zimnoch shows how tablets, Raft, and the Scylla Operator give ScyllaDB elasticity on Kubernetes. Earlier, scaling requires one node at a time, gossip coordination, and full data streaming, taking hours or days. Raft now reaches consensus in seconds, and nodes start serving before the background stream finishes. A demo doubles cluster size in about a minute and rebalances within three. A second demo auto‑scales when disk hits 90 %, holding utilization high while cutting waste.

Topics discussed

• What single‑node bootstrap with Gossip and full streaming looks like in pre‑tablet ScyllaDB clusters
• How tablets, Raft, and consistent topology changes let nodes join in seconds and scale in parallel
• How the Scylla Operator on Kubernetes doubles cluster size in one minute and finishes load balancing in three
• When a Horizontal Pod Autoscaler adds one node per rack as soon as disk exceeds 90 % utilization
• Why tablets allow safe 90 % SSD usage, lowering hardware costs while keeping clusters responsive

Takeaways

• Tablets and Raft replace gossip and per‑node streaming, so new nodes reach quorum in seconds and start coordinating traffic even while data backfills, slashing bootstrap overhead.
• A Kubernetes‑managed cluster scales horizontally in parallel; the demo adds three nodes, doubles capacity in ~60 s, and rebalances traffic by the three‑minute mark with only minor latency upticks.
• Disk‑aware autoscaling uses tablets’ 90 % utilization headroom plus HPA to add capacity without manual action, cutting scaling lag from hours to minutes and avoiding under‑used SSDs.
• Combining rapid node readiness with background file‑based streaming keeps the cluster online during expansions, making maintenance windows shorter and protecting P95/P99 latency.

Top takeaway:  Tablets and Raft let ScyllaDB clusters on Kubernetes double capacity in under three minutes while keeping latency in check.