SYSTEM Cited by 1 source
Softstore (Databricks distributed KV cache)¶
Softstore is Databricks' distributed remote KV cache for services that aren't themselves sharded. It is built on systems/dicer and is the canonical example of Dicer's state-transfer feature in action.
Why Softstore exists¶
Not every service can or should restructure into a sharded stateful model. Softstore offers a cache tier those services can call into while benefiting from Dicer's auto-sharding capabilities (load-balancing, hot-key handling, graceful restart handling) under the hood.
State transfer during rolling restarts¶
- 99.9% of restarts in the Databricks fleet are planned rolling restarts — the full keyspace churns.
- Without state transfer, a rolling restart drops cache hit rate by ~30% (every evicted pod throws away its cache, successor pod starts cold).
- With state transfer: Dicer migrates per-slice state between pods across resharding, preserving a steady ~85% hit rate through the restart for a representative use case.
The architectural pattern is named as patterns/state-transfer-on-reshard — a cache (or any per-key-state service) that reshards across restarts without re-warming.
Seen in¶
- sources/2026-01-13-databricks-open-sourcing-dicer-auto-sharder — Softstore as the state-transfer case study; Figure 5 showing hit-rate preservation.