Practical Elasticsearch Performance and Stability Tuning Guide

Elasticsearch performance and stability often raise concerns, so this guide consolidates practical tuning tips based on daily experience.

1. Configuration file tuning – enable JVM memory lock ( bootstrap.memory_lock: true ), use unicast discovery ( discovery.zen.ping.unicast.hosts ) and adjust related zen settings ( discovery.zen.ping_timeout , discovery.zen.join_timeout , discovery.zen.minimum_master_nodes ), configure fault detection intervals, limit thread‑pool queues, and set circuit‑breaker limits ( indices.breaker.total.limit , indices.breaker.request.limit , indices.breaker.fielddata.limit ).

2. System‑level tuning – choose a compatible JDK, set identical -Xms and -Xmx values (no more than 32 GB heap), disable swap ( swapoff -a ), raise file‑descriptor limits ( ulimit -n 65536 ), increase virtual memory map count ( vm.max_map_count=262144 ), and use an appropriate I/O scheduler (e.g., noop ) for SSDs.

3. Elasticsearch usage tuning – monitor hot threads ( GET /_nodes/hot_threads ) and pending tasks ( GET /_cluster/pending_tasks ), choose suitable field storage (prefer doc_values , disable unnecessary fielddata and _source ), adjust translog durability ( "index.translog.durability":"async" ) and refresh interval, disable dynamic mapping when possible, size bulk requests (5‑15 MB) and control concurrency, manage index and shard sizes (≤50 GB for logs, ≤20 GB for business), limit segment merging threads, use routing keys, index aliases, and avoid overly wide mappings ( index.mapping.total_fields.limit ).

Following these settings helps improve query latency, write throughput, and overall cluster reliability.