What's the best performance, to execute two different queries or a single?

Elastic Stack Elasticsearch
1

Hey, guys.
I'm planning a new feature and I would like to know what performs better.
Explaining a bit the scenario:

I have two different filters to be executed and generate different aggregations.

I can either:

  • Have a single search with no query filters and two filter aggregations that will split the data and perform inner aggregations.
  • Have two different searches, each with its query filter and aggregation to generate results.

Example:
Suppose that I have the following structure:

PUT /x
PUT /x/x/_mapping
{
  "properties": {
    "x":{
      "type": "integer"
    }
  }
}

PUT /x/x/1
{
  "x": 1
}

PUT /x/x/2
{
  "x": 2
}

PUT /x/x/3
{
  "x": 3
}

I could do:

GET /x/_search
{
  "size": 0,
  "aggs": {
    "f1": {
      "filter": {
        "range": {
          "x": {
            "gt": 2
          }
        }
      },
      "aggs": {
        "max": {
          "max": {
            "field": "x"
          }
        }
      }
    },
    "f2": {
      "filter": {
        "range": {
          "x": {
            "lt": 2
          }
        }
      },
      "aggs": {
        "max": {
          "max": {
            "field": "x"
          }
        }
      }
    }
  }
}

Or two different searches:

GET /x/_search
{
  "query": {
    "range": {
      "x": {
        "gt": 2
      }
    }
  }, 
  "size": 0,
  "aggs": {
    "max": {
      "max": {
        "field": "x"
      }
    }
  }
}

GET /x/_search
{
  "query": {
    "range": {
      "x": {
        "lt": 2
      }
    }
  }, 
  "size": 0,
  "aggs": {
    "max": {
      "max": {
        "field": "x"
      }
    }
  }
}

Considering a large data universe and complex filters, which one would perform better?

2

A faster query would be, using the 1st one, but changing "gt:2" and "lt:2" to `must_not: {term:2}}" in the body.

Makes sense ? So "x!=2" is a faster filter compared to 2 range filters.

Try that.

3

The "< 2" and "> 2" filters were only an example. My actual filters will be way more complex.
Can I ask how did you infer the first query would be more efficient (remember that I'm talking about large databases with complex filters)?