Why is ascending geo distance sorting faster than descending geo distance sorting

gemo1011 · March 21, 2019, 2:17pm

I'm using Elasticsearch 6.6 and have an index (1 shard, 1 replica) with the geonames (https://www.geonames.org/) dataset indexed (indexsize =1.3 gb, 11.8 mio geopoints).
I was playing around a bit with the geo distance sorting query, sorting the whole index for some origin points. So after some testing I saw that sorting ascending is always faster than sorting descending. here is an example query:

POST /geonames/_search?request_cache=false
{   
    "size":1,
    "sort" : [
        {
            "_geo_distance" : {
                "location" : [8, 49],
                "order" : "asc",
                "unit" : "m",
                "mode" : "min",
                "distance_type" : "arc",
                "ignore_unmapped": true
            }
        }
    ]
}

Here is the answer for ascending sorting (with explain and profile True):

{
  "took" : 1374,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : 11858060,
    "max_score" : null,
    "hits" : [
      {
        "_shard" : "[geonames][0]",
        "_node" : "qXTymyB9QLmxhPtGEtA_mA",
        "_index" : "geonames",
        "_type" : "doc",
        "_id" : "L781LmkBrQo0YN4qP48D",
        "_score" : null,
        "_source" : {
          "id" : "3034701",
          "name" : "Forêt de Wissembourg",
          "location" : {
            "lat" : "49.00924",
            "lon" : "8.01542"
          }
        },
        "sort" : [
          1523.4121312414704
        ],
        "_explanation" : {
          "value" : 1.0,
          "description" : "*:*",
          "details" : [ ]
        }
      }
    ]
  },
  "profile" : {
    "shards" : [
      {
        "id" : "[qXTymyB9QLmxhPtGEtA_mA][geonames][0]",
        "searches" : [
          {
            "query" : [
              {
                "type" : "MatchAllDocsQuery",
                "description" : "*:*",
                "time_in_nanos" : 265223567,
                "breakdown" : {
                  "score" : 0,
                  "build_scorer_count" : 54,
                  "match_count" : 0,
                  "create_weight" : 10209,
                  "next_doc" : 253091268,
                  "match" : 0,
                  "create_weight_count" : 1,
                  "next_doc_count" : 11858087,
                  "score_count" : 0,
                  "build_scorer" : 263948,
                  "advance" : 0,
                  "advance_count" : 0
                }
              }
            ],
            "rewrite_time" : 1097,
            "collector" : [
              {
                "name" : "CancellableCollector",
                "reason" : "search_cancelled",
                "time_in_nanos" : 1044167746,
                "children" : [
                  {
                    "name" : "SimpleFieldCollector",
                    "reason" : "search_top_hits",
                    "time_in_nanos" : 508296683
                  }
                ]
              }
            ]
          }
        ],
        "aggregations" : [ ]
      }
    ]
  }
}

and here for descending, just switched the parameter from asc to desc (also with profile and explain):

{
  "took" : 2226,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : 11858060,
    "max_score" : null,
    "hits" : [
      {
        "_shard" : "[geonames][0]",
        "_node" : "qXTymyB9QLmxhPtGEtA_mA",
        "_index" : "geonames",
        "_type" : "doc",
        "_id" : "Mq80LmkBrQo0YN4q11bA",
        "_score" : null,
        "_source" : {
          "id" : "4036351",
          "name" : "Bollons Seamount",
          "location" : {
            "lat" : "-49.66667",
            "lon" : "-176.16667"
          }
        },
        "sort" : [
          1.970427111052182E7
        ],
        "_explanation" : {
          "value" : 1.0,
          "description" : "*:*",
          "details" : [ ]
        }
      }
    ]
  },
  "profile" : {
    "shards" : [
      {
        "id" : "[qXTymyB9QLmxhPtGEtA_mA][geonames][0]",
        "searches" : [
          {
            "query" : [
              {
                "type" : "MatchAllDocsQuery",
                "description" : "*:*",
                "time_in_nanos" : 268521404,
                "breakdown" : {
                  "score" : 0,
                  "build_scorer_count" : 54,
                  "match_count" : 0,
                  "create_weight" : 9333,
                  "next_doc" : 256458664,
                  "match" : 0,
                  "create_weight_count" : 1,
                  "next_doc_count" : 11858087,
                  "score_count" : 0,
                  "build_scorer" : 195265,
                  "advance" : 0,
                  "advance_count" : 0
                }
              }
            ],
            "rewrite_time" : 1142,
            "collector" : [
              {
                "name" : "CancellableCollector",
                "reason" : "search_cancelled",
                "time_in_nanos" : 1898324618,
                "children" : [
                  {
                    "name" : "SimpleFieldCollector",
                    "reason" : "search_top_hits",
                    "time_in_nanos" : 1368306442
                  }
                ]
              }
            ]
          }
        ],
        "aggregations" : [ ]
      }
    ]
  }
}

So my question is, why is it like this ? As I understood Es calculates the distance from the origin point to every other point and then sorts them. So why is the descending sorting so much slower ?

Ignacio_Vera · March 21, 2019, 4:21pm

In the case of sorting ascending it uses a specialise sorting algorithm/strategy (LatLonDocValuesField.newDistanceSort). When sorting descending it is doing what you said.

github.com

elastic/elasticsearch/blob/c178e1bd7317dfb0c2126876915ae3e09bfa6fb6/server/src/main/java/org/elasticsearch/search/sort/GeoDistanceSortBuilder.java#L630


    if (nestedSort.getNestedSort() != null && nestedSort.getMaxChildren() != Integer.MAX_VALUE)  {
        throw new QueryShardException(context,
            "max_children is only supported on last level of nested sort");
    }
    // new nested sorts takes priority
    nested = resolveNested(context, nestedSort);
} else {
    nested = resolveNested(context, nestedPath, nestedFilter);
}


if (geoIndexFieldData.getClass() == LatLonPointDVIndexFieldData.class // only works with 5.x geo_point
        && nested == null
        && finalSortMode == MultiValueMode.MIN // LatLonDocValuesField internally picks the closest point
        && unit == DistanceUnit.METERS
        && reverse == false
        && localPoints.length == 1) {
    return new SortFieldAndFormat(
            LatLonDocValuesField.newDistanceSort(fieldName, localPoints[0].lat(), localPoints[0].lon()),
            DocValueFormat.RAW);
}

gemo1011 · March 21, 2019, 6:42pm

Thank you very much for your fast answer.
The different sorting strategies explain the differences in the query time.
So I'm trying to understand how the ascending sorting works. My Java knowledge isn't very good, so I only see that the LatLonDocValuesField.newDistanceSort returns a LatLonPointSortField that somehow calls SortField ?

Ignacio_Vera · March 22, 2019, 6:30am

LatLonDocValuesField.newDistanceSort is a primitive implemented at Lucene level:

github.com

iverase/lucene-solr/blob/master/lucene/core/src/java/org/apache/lucene/document/LatLonPointSortField.java

/*
 * Licensed to the Apache Software Foundation (ASF) under one or more
 * contributor license agreements.  See the NOTICE file distributed with
 * this work for additional information regarding copyright ownership.
 * The ASF licenses this file to You under the Apache License, Version 2.0
 * (the "License"); you may not use this file except in compliance with
 * the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
package org.apache.lucene.document;

import org.apache.lucene.geo.GeoUtils;
import org.apache.lucene.search.FieldComparator;

This file has been truncated. show original

The key of the algorithm is on LatLonPointDistanceComparator, the javadocs explain the strategy used. A bounding box from the min competitive distance is built and therefore we can reject points based on this bounding box instead of calculating the distance for every element which is expensive.

github.com

iverase/lucene-solr/blob/master/lucene/core/src/java/org/apache/lucene/document/LatLonPointDistanceComparator.java

/*
 * Licensed to the Apache Software Foundation (ASF) under one or more
 * contributor license agreements.  See the NOTICE file distributed with
 * this work for additional information regarding copyright ownership.
 * The ASF licenses this file to You under the Apache License, Version 2.0
 * (the "License"); you may not use this file except in compliance with
 * the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
package org.apache.lucene.document;

import java.io.IOException;

This file has been truncated. show original

gemo1011 · March 22, 2019, 9:45am

The filtering of the points beforehand with a bbox for ascending search makes sense.
But I don't understand how you get the min competitive distance for calculating the bounding box.
So looking at my query there are 11858060 hits that should be ordered. How do you choose the distance for the bbox ?
Looking at the code it seems like it using some kind of iterative bounding box creation and calculating the distances for the points in the bbox.

So for me it looks like that it takes (randomly ?) some points from the hits, calculates a bbox for the min value of this points.
If they are enough points for the output they get ordered by distance.
If they are not enough a bigger bbox gets created.
Until enough points are gathered, or a specific number of iterations was done:
if (setBottomCounter < 1024 || (setBottomCounter & 0x3F) == 0x3F)

Thank you for taking the time explaining this issue to me.

Ignacio_Vera · March 22, 2019, 10:41am

The process starts the same way when you calculate the distance for all hits on your query. You get those values from the docValues in your index that I believe they come in the order they were writing in the index.

Not sure if setBottom(int slot) is called every time you have a min competitive document or it is sampled. In your case because your query has a size=1, every new document that has a distance lower than the current min competitive document will replace the previous one.

At the beginning we create a new bounding box for each call to the method setBottomCounter < 1024 and after that we start sampling (setBottomCounter & 0x3F) == 0x3F.

gemo1011 · March 22, 2019, 1:22pm

Ok, I think I'm now roughly understanding the algorithm.
Thank you again

system · April 19, 2019, 1:22pm

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.

Topic		Replies	Views
Usual sorting with elasticsearch (boost/sort based on other documents) Elasticsearch	2	924	July 12, 2019
Trouble Sorting by Geodistance in Elasticsearch with Nested Geo_location Field Elasticsearch	2	109	May 13, 2024
Geo_point sort by distance Elasticsearch	8	15763	October 29, 2019
Clarification on geo distance sorting versus geopoint field distance method Elasticsearch	3	504	July 6, 2017
_geo_distance sort always returning Infinity Elasticsearch	5	13	September 13, 2024

Why is ascending geo distance sorting faster than descending geo distance sorting

Related topics