Speculative Partial Writes in Erasure-Coded Storage Systems (PBS Design and Evaluation)

This article presents a Chinese translation of the USENIX'17 paper "Speculative Partial Writes in Erasure-Coded Systems". It explains that erasure coding (EC) improves storage efficiency but incurs significant write overhead, especially when write offsets are not aligned with EC groups, leading to costly partial writes.

Traditional solutions read the untouched data, recombine, and rewrite the whole EC group, or use incremental encoding based on data differences, both of which still require costly read‑modify‑write cycles on HDDs.

The authors propose a speculative partial‑write mechanism called PBS (Parity‑Based Speculation). PBS logs changes on parity blocks, allowing speculative writes that can later be confirmed or rolled back, reducing unnecessary reads and converting random writes into sequential appends.

Implementation details are built on Meituan’s distributed block storage system Ursa, providing strong consistency. The paper includes experimental results showing EC encoding/decoding throughput of 5–13 GB/s per CPU core, and PBS achieving write IOPS and latency comparable to three‑replica systems, outperforming prior EC‑PLog methods.

Failure‑recovery tests demonstrate that log size has limited impact on recovery speed thanks to in‑memory indexing.