about things notes.zzstoatzz.io
notes systems write-batching.md
1.3 kB · 22 lines
Markdown
at main

write batching #

A maximum batch size is a limit, not the batch size a pipeline actually achieves. Even with no aggregation delay, a single writer can coalesce requests already waiting when it finishes the previous group. The number of producers and their work before and after each write determine how often that opportunity exists.

A tight index benchmark can therefore produce different groups from the same index inside a relay that also verifies signatures, queries metadata and waits for durable event publication. Increasing in-flight work can improve throughput by amortizing sync costs while increasing individual write latency. It can also move contention to another stage; more concurrency is not an unconditional win.

Measure actual groups or changed commits alongside submitted logical operations, producer concurrency, queueing, durability and remaining maintenance work. Keep input equality and acceptance checks separate from throughput. Do not infer pipeline capacity from the inverse of one operation's latency or compare engines under different batching opportunities without naming that difference.

sources #

  • zat.dev/zlay, offline pilot docs/zkv-relay-replay.md and paired raw measurements under docs/zkv-load/relay-replay, 2026-11-04.