Elasticsearch OpenSearch Threadpool

Opster Team

Last updated: Oct 30, 2022

| 2 min read

Overview

OpenSearch uses threadpools to manage how requests are processed and to optimize the use of resources on each node in the cluster.

What it’s used for

The main threadpools are for search, get and write, but there are a number of others which you can see by running: 

GET /_cat/thread_pool/?v&h=id,name,active,rejected,completed,size,type&pretty

You can see by running the above command that each node has a number of different thread pools, what the size and type of the thread pool are, and you can see which nodes have rejected operations. OpenSearch automatically configures the threadpool management parameters based on the number of processors detected in each node.  

Threadpool types

Fixed- a fixed number of threads, with a fixed queue size

thread_pool:
    write:
        size: 30
        queue_size: 1000

Scaling- a variable number of threads that OpenSearch scales automatically according to workload.

thread_pool:
    warmer:
        core: 1
        max: 8

fixed_autoqueue_size- a fixed number of threads with a variable queue size which changes dynamically in order to maintain a target response time

thread_pool:
    search:
        size: 30
        queue_size: 500
        min_queue_size: 10
        max_queue_size: 1000
        auto_queue_frame_size: 2000
        target_response_time: 1s

Examples

To see which threads are using the highest CPU or taking the longest time you can use the following query. This may help find operations that are causing your cluster to underperform.

GET /_nodes/hot_threads

Notes and good things to know

In general it is not recommended to tweak threadpool settings. However, it is worth noting that the threadpools are set based upon the number of processors that OpenSearch has detected on the underlying hardware. If that detection fails, then you should explicitly set the number of processors available in your hardware in OpenSearch.yml like this:

processors: 4

Most threadpools also have queues associated with them to enable OpenSearch to store requests in memory while waiting for resources to become available to process the request.  However the queues are usually of a finite size, and if that size becomes exceeded, then OpenSearch will reject the request.

Sometimes you may be tempted to increase the queue size to prevent requests being rejected, but this will only treat the symptom and not the underlying cause of the problem.  Indeed, it may even be counter productive, since by allowing a larger queue size, the node will need to use more memory to store the queue, and will have less space to actually manage requests.  Furthermore increasing the queue size will also increase the length of time that operations are kept in the queue, resulting in client applications facing time out issues.

Usually, the only case where it can be justified to increase the queue size is where requests are received in uneven surges and you are unable to manage this process client-side.

You can monitor thread pools to better understand the performance of your OpenSearch cluster. The OpenSearch monitoring panel in Kibana shows your graphs of the search, get, and write thread queues and any queue rejections.  Growing queues indicate that OpenSearch is having difficulty keeping up with requests, and rejections indicate that queues have grown to the point that OpenSearch rejects calls to the server.

Check the underlying causes of increases in queues. Try to balance activity across the nodes in the cluster and try to balance the demands on the cluster thread pool by taking actions on the client-side.