Dual-bucket throttling
Dual-bucket throttling protects steady storage traffic with a reserved token bucket alongside a shared main bucket.
In the NSDI 2026 design, all segments use the main bucket; only steady segments can use the cheque bucket. Classification uses recent segment traffic. Borrowing lasts one interval, and subsequent main-bucket credits replenish the reserve. [1]
This is a particular burst-isolation design, not a generic name for any two rate limits. Its capacity estimate and classification require workload-specific validation. The linked analysis checks the printed model equations separately from reported deployment results. [1] [2]
Relations
- related
- Token bucket, Tail latency, I/O scheduler
Citations
- Hu et al. Come Hell or Still Water: Alleviating Tail Latency in Cloud Block Store. NSDI 2026, section 4.2 and Appendix A.
- Hesela: burst isolation and a reproducible audit of the printed queueing equations.
Sources & caveats
- The paper describes an operator-specific production design. No implementation reproduction or independent latency measurement is claimed.