Give memory to the filesystem cache

Elasticsearch heavily relies on the filesystem cache in order to make search fast. In general, you should make sure that at least half the available memory goes to the filesystem cache so that Elasticsearch can keep hot regions of the index in physical memory.

Use faster hardware

If your search is I/O bound, you should investigate giving more memory to the filesystem cache (see above) or buying faster drives. In particular SSD drives are known to perform better than spinning disks. Always use local storage, remote filesystems such as NFS or SMB should be avoided. Also beware of virtualized storage such as Amazon’s Elastic Block Storage. Virtualized storage works very well with Elasticsearch, and it is appealing since it is so fast and simple to set up, but it is also unfortunately inherently slower on an ongoing basis when compared to dedicated local storage. If you put an index on EBS, be sure to use provisioned IOPS otherwise operations could be quickly throttled.

If your search is CPU-bound, you should investigate buying faster CPUs.

Document modeling

Documents should be modeled so that search-time operations are as cheap as possible.

In particular, joins should be avoided. nested can make queries several times slower and parent-child relations can make queries hundreds of times slower. So if the same questions can be answered without joins by denormalizing documents, significant speedups can be expected.

Search as few fields as possible

The more fields a query_string or multi_match query targets, the slower it is. A common technique to improve search speed over multiple fields is to copy their values into a single field at index time, and then use this field at search time. This can be automated with the copy-to directive of mappings without having to change the source of documents. Here is an example of an index containing movies that optimizes queries that search over both the name and the plot of the movie by indexing both values into the name_and_plot field.

PUT movies
{
  "mappings": {
    "properties": {
      "name_and_plot": {
        "type": "text"
      },
      "name": {
        "type": "text",
        "copy_to": "name_and_plot"
      },
      "plot": {
        "type": "text",
        "copy_to": "name_and_plot"
      }
    }
  }
}

Pre-index data

You should leverage patterns in your queries to optimize the way data is indexed. For instance, if all your documents have a price field and most queries run range aggregations on a fixed list of ranges, you could make this aggregation faster by pre-indexing the ranges into the index and using a terms aggregations.

For instance, if documents look like:

PUT index/_doc/1
{
  "designation": "spoon",
  "price": 13
}

and search requests look like:

GET index/_search
{
  "aggs": {
    "price_ranges": {
      "range": {
        "field": "price",
        "ranges": [
          { "to": 10 },
          { "from": 10, "to": 100 },
          { "from": 100 }
        ]
      }
    }
  }
}

Then documents could be enriched by a price_range field at index time, which should be mapped as a keyword:

PUT index
{
  "mappings": {
    "properties": {
      "price_range": {
        "type": "keyword"
      }
    }
  }
}

PUT index/_doc/1
{
  "designation": "spoon",
  "price": 13,
  "price_range": "10-100"
}

And then search requests could aggregate this new field rather than running a range aggregation on the price field.

GET index/_search
{
  "aggs": {
    "price_ranges": {
      "terms": {
        "field": "price_range"
      }
    }
  }
}

Consider mapping identifiers as keyword

The fact that some data is numeric does not mean it should always be mapped as a numeric field. The way that Elasticsearch indexes numbers optimizes for range queries while keyword fields are better at term queries. Typically, fields storing identifiers such as an ISBN or any number identifying a record from another database are rarely used in range queries or aggregations. This is why they might benefit from being mapped as keyword rather than as integer or long.

Avoid scripts

In general, scripts should be avoided. If they are absolutely needed, you should prefer the painless and expressions engines.

Search rounded dates

Queries on date fields that use now are typically not cacheable since the range that is being matched changes all the time. However switching to a rounded date is often acceptable in terms of user experience, and has the benefit of making better use of the query cache.

For instance the below query:

PUT index/_doc/1
{
  "my_date": "2016-05-11T16:30:55.328Z"
}

GET index/_search
{
  "query": {
    "constant_score": {
      "filter": {
        "range": {
          "my_date": {
            "gte": "now-1h",
            "lte": "now"
          }
        }
      }
    }
  }
}

could be replaced with the following query:

GET index/_search
{
  "query": {
    "constant_score": {
      "filter": {
        "range": {
          "my_date": {
            "gte": "now-1h/m",
            "lte": "now/m"
          }
        }
      }
    }
  }
}

In that case we rounded to the minute, so if the current time is 16:31:29, the range query will match everything whose value of the my_date field is between 15:31:00 and 16:31:59. And if several users run a query that contains this range in the same minute, the query cache could help speed things up a bit. The longer the interval that is used for rounding, the more the query cache can help, but beware that too aggressive rounding might also hurt user experience.

It might be tempting to split ranges into a large cacheable part and smaller not cacheable parts in order to be able to leverage the query cache, as shown below:

GET index/_search
{
  "query": {
    "constant_score": {
      "filter": {
        "bool": {
          "should": [
            {
              "range": {
                "my_date": {
                  "gte": "now-1h",
                  "lte": "now-1h/m"
                }
              }
            },
            {
              "range": {
                "my_date": {
                  "gt": "now-1h/m",
                  "lt": "now/m"
                }
              }
            },
            {
              "range": {
                "my_date": {
                  "gte": "now/m",
                  "lte": "now"
                }
              }
            }
          ]
        }
      }
    }
  }
}

However such practice might make the query run slower in some cases since the overhead introduced by the bool query may defeat the savings from better leveraging the query cache.

Force-merge read-only indices

Indices that are read-only would benefit from being merged down to a single segment. This is typically the case with time-based indices: only the index for the current time frame is getting new documents while older indices are read-only.

Don’t force-merge indices that are still being written to — leave merging to the background merge process.

Warm up global ordinals

Global ordinals are a data-structure that is used in order to run terms aggregations on keyword fields. They are loaded lazily in memory because Elasticsearch does not know which fields will be used in terms aggregations and which fields won’t. You can tell Elasticsearch to load global ordinals eagerly at refresh-time by configuring mappings as described below:

PUT index
{
  "mappings": {
    "properties": {
      "foo": {
        "type": "keyword",
        "eager_global_ordinals": true
      }
    }
  }
}

Warm up the filesystem cache

If the machine running Elasticsearch is restarted, the filesystem cache will be empty, so it will take some time before the operating system loads hot regions of the index into memory so that search operations are fast. You can explicitly tell the operating system which files should be loaded into memory eagerly depending on the file extension using the index.store.preload setting.

Loading data into the filesystem cache eagerly on too many indices or too many files will make search slower if the filesystem cache is not large enough to hold all the data. Use with caution.

Use index sorting to speed up conjunctions

Index sorting can be useful in order to make conjunctions faster at the cost of slightly slower indexing. Read more about it in the index sorting documentation.

Use preference to optimize cache utilization

There are multiple caches that can help with search performance, such as the filesystem cache, the request cache or the query cache. Yet all these caches are maintained at the node level, meaning that if you run the same request twice in a row, have 1 replica or more and use round-robin, the default routing algorithm, then those two requests will go to different shard copies, preventing node-level caches from helping.

Since it is common for users of a search application to run similar requests one after another, for instance in order to analyze a narrower subset of the index, using a preference value that identifies the current user or session could help optimize usage of the caches.

Replicas might help with throughput, but not always

In addition to improving resiliency, replicas can help improve throughput. For instance if you have a single-shard index and three nodes, you will need to set the number of replicas to 2 in order to have 3 copies of your shard in total so that all nodes are utilized.

Now imagine that you have a 2-shards index and two nodes. In one case, the number of replicas is 0, meaning that each node holds a single shard. In the second case the number of replicas is 1, meaning that each node has two shards. Which setup is going to perform best in terms of search performance? Usually, the setup that has fewer shards per node in total will perform better. The reason for that is that it gives a greater share of the available filesystem cache to each shard, and the filesystem cache is probably Elasticsearch’s number 1 performance factor. At the same time, beware that a setup that does not have replicas is subject to failure in case of a single node failure, so there is a trade-off between throughput and availability.

So what is the right number of replicas? If you have a cluster that has num_nodes nodes, num_primaries primary shards in total and if you want to be able to cope with max_failures node failures at once at most, then the right number of replicas for you is max(max_failures, ceil(num_nodes / num_primaries) - 1).