How Meituan Cut Elasticsearch Search Latency by 84% with an RLE‑Based Inverted Index

This article details Meituan's search‑engine team optimization of Elasticsearch for a high‑traffic LBS scenario, describing the performance bottlenecks in term‑posting retrieval and merging, the design of a run‑length‑encoding (RLE) inverted index, its integration as a plugin, extensive benchmarking, and the resulting 84% reduction in TP99 query latency.

Backend SearchElasticsearchRoaringBitmap

0 likes · 25 min read

How Meituan Cut Elasticsearch Search Latency by 84% with an RLE‑Based Inverted Index

Meituan Technology Team

Nov 17, 2022 · Backend Development

Elasticsearch Query and Merge Optimization Using Run-Length Encoding for Meituan Takeaway Search

Meituan's food‑delivery search team identified heavy CPU and latency hotspots in Elasticsearch's posting‑list query and merge phases, then redesigned the inverted index using Run‑Length Encoding, hash‑based term lookup, index sorting and a custom SparseRoaringDocIdSet, ultimately reducing TP99 search latency by 84% and cutting CPU usage dramatically.

ElasticsearchIndex SortingJava

0 likes · 26 min read