Diagonal scaling, LSMs and TigerBeetle
A recent recap promoted 'Diagonal Scaling'—mixing object storage, LSM trees and replicated state machines—as a survivable pattern at scale, with TigerBeetle cited as an example shared. The approach reframes storage and consensus choices for high-throughput services.
TigerBeetletigerbeetle.com published a newsletter dated March 11, 2026 that says the project demoed processing 1,000,000,000,000 (one trillion) transactions using their new “diagonal scaling” approach. Their engineering notes show the storage engine now exposes per-tree metrics and tracks logical vs physical bytes during compaction for an approximately 30-tree LSM forest, enabling direct estimation of write-amplification per tree.tigerbeetle.com The system is implemented as a replicated state machine using Viewstamped Replication and is documented to run as a six‑replica cluster with each replica’s state kept in a single on‑disk data file.github.com The Diagonal Scaling paper’s DIAGONALSCALE method reports experimental p95 latency reductions up to 40%, cost‑per‑query decreases up to 37%, and rebalancing reductions of 2–5× in benchmark surfaces.arxiv.org TigerBeetle’s website emphasizes a “scale up on 6 machines — replicated, never partitioned” design point as their scalability posture.tigerbeetle.com
