应为不存储数据的Elasticsearch节点分配多少CPU资源？

Hey folks, let's walk through how to figure out the right CPU core count for your Elasticsearch setup, based on the shard-thread relationship you're referencing.

Key Background to Start With #

First, a quick recap of the critical info we have:

• Elasticsearch's official docs recommend using multi-core processors but don't spell out exact core counts for peak indexing or search performance.
• From community deep dives into CPU utilization during searches, each index shard consumes one CPU thread when processing search requests—this is the core rule of thumb for sizing your CPU resources.

Applying This to Your Cluster Architecture #

You mentioned a cluster with:

• 3 master-only nodes (node.master: true, node.data: false): These nodes only handle cluster management tasks (like leader elections, metadata updates) and don't participate in actual indexing or search work. For these, you don't need heavy-duty CPU—4 cores (8 threads) is usually more than enough to keep cluster control operations running smoothly without bottlenecks.
• 3 data nodes (I’m assuming these are node.data: true, node.master: false since your description cut off): These are the workhorses that handle all search and indexing loads. Here’s how to calculate their CPU needs:
  • First, tally up your total shards (primary + replica). For example, if you have an index with 6 primary shards and 1 replica, that’s 12 total shards.
  • These shards will distribute evenly across your 3 data nodes, so each node ends up with ~4 shards in this example.
  • During peak search traffic, each shard uses one CPU thread. So each data node needs at least as many cores as the number of shards it hosts to avoid thread contention. If you’re also running ongoing indexing workloads, add an extra 20-30% of cores to reserve for indexing thread pools.
  • Sticking with the example: 4 shards per node means starting with 4 cores, plus 20% gives you 5 cores—so a 6-core CPU would be a safe, performant choice.

Quick Optimization Tips #

• Avoid over-sharding: Too many shards leads to unnecessary CPU thread competition that slows down performance. Aim for no more than 2-3 shards per CPU core on data nodes for search-heavy workloads.
• Monitor CPU usage regularly: Use the GET _cat/nodes?v command to check the cpu column, or leverage Kibana's monitoring dashboards. You want peak CPU usage to stay around 70-80%—if it’s consistently higher, either add more cores or adjust your shard layout to spread the load better.
• Keep master nodes isolated: Never assign data roles to master-only nodes—this ensures their CPU resources stay dedicated to cluster management, preventing control-plane bottlenecks that can break your whole cluster.

内容的提问来源于stack exchange，提问作者Shanim