Skip to content

PATTERN Cited by 1 source

Batched consensus proposals

Pattern

Batch multiple client writes arriving within a short window into a single consensus proposal, amortizing the multi-round-trip cost of consensus across many operations and improving throughput without changing the consistency guarantee.

Mechanism

If a replica receives 10 writes in a span of 10ms, it places all of them in a single log event and proposes that event for the next slot. One consensus round decides all 10 writes atomically.

When to apply

  • Consensus round-trip cost dominates per-write latency
  • Write arrival rate is bursty (many writes in short windows)
  • Operations are independent enough to be batched without ordering constraints within the batch
  • The system is throughput-limited by consensus rounds, not by compute

Trade-offs

  • Gain: higher throughput (more operations decided per consensus round)
  • Cost: slightly higher tail latency for individual writes that wait for the batch window

Meerkat instance

Meerkat uses batching as one of several optimizations to squeeze better performance from the fundamental consensus latency. Particularly effective because Meerkat targets control-plane data written in bursts (e.g., configuration updates, leadership changes).

Seen in

Last updated · 573 distilled / 1,747 read