silk · research · also SILK LSM scheduler

SILK

SILK is a research extension of RocksDB that coordinates client I/O with background LSM maintenance.

Scheduling: SILK combines load-aware bandwidth allocation, flush and L0-compaction priority, and preemption of less urgent compactions. [1]

Code boundary: the pinned prototype checks its high-priority scheduling branch before returning for paused background compaction. That branch examines the default column family's L0 file count; it is not a general multi-family fairness guarantee. [2]

Evidence boundary: the reported experiments disable compression and commit logging. Sustained overload can still degrade performance. [1] The author repository identifies this as the ATC 2019 prototype, not a feature contract for current RocksDB. [3]

Relations

related

RocksDB write stall, RocksDB pending compaction bytes, LSM compaction, Log-structured merge-tree, Tail latency

Citations

1. Balmau et al. SILK: Preventing Latency Spikes in Log-Structured Merge Key-Value Stores. USENIX ATC 2019, sections 5, 6.1 and 6.4. (accessed 2026-11-10)
2. SILK author prototype, commit 35385f14b6d3948c75914c1571e92897280a483e: high-priority scheduling before the background-compaction pause check. (accessed 2026-11-10)
3. SILK author README at 35385f14b6d3948c75914c1571e92897280a483e: prototype identity and benchmark entry points. (accessed 2026-11-10)

Further reading

4. Hesela: SILK and the distinction between scheduling, backlog and durability. (accessed 2026-11-10)

JSON: /data/terms/silk.json