PostgreSQL 12: Implementing K-Nearest Neighbor Space Partitioned Generalized Search Tree Indexes

November 05, 2019
PostgreSQL 12: Implementing K-Nearest Neighbor Space Partitioned Generalized Search Tree Indexes

The value of indexing

PostgreSQL provides a simple linear distance operator <-> (linear distance). We will use this to find points that are closest to a given location.

PostgreSQL provides a simple linear distance operator the data, and performing no optimizations and having no indexes, we see the following execution plan:

time psql -qtAc "
EXPLAIN (ANALYZE ON, BUFFERS ON)
SELECT name, location
FROM geonames
ORDER BY location <-> '(29.9691,-95.6972)'
LIMIT 5;
"  <-- closing quote
                                      QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Limit  (cost=418749.15..418749.73 rows=5 width=38) 
        (actual time=2553.970..2555.673 rows=5 loops=1)
  Buffers: shared hit=100 read=272836
  ->  Gather Merge  (cost=418749.15..1580358.21 rows=9955954 width=38) 
                    (actual time=2553.969..2555.669 rows=5 loops=1)
        Workers Planned: 2
        Workers Launched: 2
        Buffers: shared hit=100 read=272836
        ->  Sort  (cost=417749.12..430194.06 rows=4977977 width=38)
                 (actual time=2548.220..2548.221 rows=4 loops=3)
              Sort Key: ((location <-> '(29.9691,-95.6972)'::point))
              Sort Method: top-N heapsort  Memory: 25kB
              Worker 0:  Sort Method: top-N heapsort  Memory: 26kB
              Worker 1:  Sort Method: top-N heapsort  Memory: 25kB
              Buffers: shared hit=100 read=272836
              ->  Parallel Seq Scan on geonames  (cost=0.00..335066.71 rows=4977977 width=38) 
                                        (actual time=0.040..1637.884 rows=3982382 loops=3)
                    Buffers: shared hit=6 read=272836
Planning Time: 0.493 ms
Execution Time: 2555.737 ms

real    0m2.595s
user    0m0.011s
sys    0m0.015s

and here are the results: (the same results for all requests, so we’ll omit them later.)

name location
Cypress (29.96911,-95.69717)
Cypress Pointe Baptist Church (29.9732,-95.6873)
Cypress Post Office (29.9743,-95.67953)
Hot Wells (29.95689,-95.68189)
Dry Creek Airport (29.98571,-95.68597)

So, 418749.73 is the OPTIMIZER cost to beat, and it took two and a half seconds (2555.673) to execute that query. This is actually a very good result, using PostgreSQL with no optimizations at all against an 11 million row table. This is also why we selected a larger data set, as there would be very minimal difference using indexes against less than 10 millions rows. Parallel sequential scans are fantastic, but that’s another article.

Adding GiST index

We begin the optimization process by adding a GiST index. Because our example query has a

LIMIT

clause of 5 items, we have a very high selectivity. This will encourage the planner to use an index, so we’ll provide one that works fairly well with geometry data.

time psql -qtAc "CREATE INDEX idx_gist_geonames_location ON geonames USING gist(location);"

The act of creating the index has a bit of an expense.

CREATE INDEX
real    3m1.988s
user    0m0.011s
sys     0m0.014s

And then run the same query again.

time psql -qtAc "
EXPLAIN (ANALYZE ON, BUFFERS ON)
SELECT name, location
FROM geonames
ORDER BY location <-> '(29.9691,-95.6972)'
LIMIT 5;
"
                                      QUERY PLAN
----------------------------------------------------------------------------------
Limit  (cost=0.42..1.16 rows=5 width=38) (actual time=0.797..0.881 rows=5 loops=1)
  Buffers: shared hit=5 read=15
  ->  Index Scan using idx_gist_geonames_location on geonames  
            (cost=0.42..1773715.32 rows=11947145 width=38) 
            (actual time=0.796..0.879 rows=5 loops=1)
        Order By: (location <-> '(29.9691,-95.6972)'::point)
        Buffers: shared hit=5 read=15
Planning Time: 0.768 ms
Execution Time: 0.939 ms

real    0m0.033s
user    0m0.011s
sys     0m0.013s

In this case, we see some pretty dramatic improvement. The estimated cost of the query is only 1.16! Compare that to the original cost of the unoptimized query at 418749.73. The actual time taken was .939 milliseconds (nine tenths of a millisecond), which compares to the 2.5 seconds of the original query. This result took less time to plan, got a dramatically better estimate, and took about 3 orders of magnitude less runtime.

Let’s see if we can do better.

Adding an SP-GiST index

time psql -qtAc "CREATE INDEX idx_spgist_geonames_location ON geonames USING spgist(location);"
CREATE INDEX 

real    1m25.205s
user    0m0.010s
sys        0m0.015s

And then we run the same query again.

time psql -qtAc "
EXPLAIN (ANALYZE ON, BUFFERS ON)
SELECT name, location
FROM geonames
ORDER BY location <-> '(29.9691,-95.6972)'
LIMIT 5;
"
                                      QUERY PLAN
-----------------------------------------------------------------------------------
 Limit  (cost=0.42..1.09 rows=5 width=38) (actual time=0.066..0.323 rows=5 loops=1)
   Buffers: shared hit=47
   ->  Index Scan using idx_spgist_geonames_location on geonames  
            (cost=0.42..1598071.32 rows=11947145 width=38) 
            (actual time=0.065..0.320 rows=5 loops=1)
         Order By: (location <-> '(29.9691,-95.6972)'::point)
         Buffers: shared hit=47
 Planning Time: 0.122 ms
 Execution Time: 0.358 ms
(7 rows)

real    0m0.040s
user    0m0.011s
sys        0m0.015s

Wow! Now using an SP-GiST index, the query cost only 1.09, and executed in .358 milliseconds (a third of a millisecond).

Let’s examine some things about the indexes themselves, and see how they stack up to each other on disk.

Index comparisons

indexname creation time estimate query time indexsize plan time
unindexed 0S 418749.73 2555.673 0 .493
idx_gist_geonames_location 3M 1S 1.16 .939 ms 868 MB .786
idx_spgist_geonames_location 1M 25S 1.09 .358 ms 523 MB .122

Conclusions

So, we see that SP-GiST is twice the speed of GiST in execution, 8x faster to plan, and about 60% of the size on disk. And (relevant to this article) it also supports KNN index searching as of PostgreSQL 12. For this type of operation, we have a clear winner.

Appendices

Setting up the data

For this article, we are going to use the data provided by the GeoNames Gazetteer.
This work is licensed under a Creative Commons Attribution 4.0 License
The Data is provided “as is” without warranty or any representation of accuracy, timeliness or completeness.

Create the structure

We start the process by creating a working directory and a little bit of ETL.

# change to our home directory
cd
mkdir spgist
cd spgist
# get the base data.  
# This file is 350MB.  It will unpack to 1.5GB
# It will expand to 2GB in PostgreSQL,
#    and then you will still need some room for indexes
#  All together, you will need about 
#  3GB of space for this exercise
#  for about 12M rows of data.

psql -qtAc "
CREATE TABLE IF NOT EXISTS geonames (
geonameid           integer primary key
,name               text 
,asciiname          text 
,alternatenames     text 
,latitude           numeric(13,5) 
,longitude          numeric(13,5)
,feature_class      text 
,feature_code       text 
,country            text 
,cc2                text 
,admin1             text 
,admin2             bigint 
,admin3             bigint 
,admin4             bigint 
,population         bigint 
,elevation          bigint 
,dem                bigint 
,timezone           text 
,modification date  );

COMMENT ON COLUMN geonames.geonameid          
 IS ' integer id of record in geonames database';
COMMENT ON COLUMN geonames.name               
 IS ' name of geographical point (utf8) varchar(200)';
COMMENT ON COLUMN geonames.asciiname          
 IS ' name of geographical point in plain ascii characters, varchar(200)';
COMMENT ON COLUMN geonames.alternatenames     
 IS ' alternatenames, comma separated, ascii names automatically transliterated, 
    convenience attribute from alternatename table, varchar(10000)';
COMMENT ON COLUMN geonames.latitude           
 IS ' latitude in decimal degrees (wgs84)';
COMMENT ON COLUMN geonames.longitude          
 IS ' longitude in decimal degrees (wgs84)';
COMMENT ON COLUMN geonames.feature_class      
 IS ' http://www.geonames.org/export/codes.html, char(1)';
COMMENT ON COLUMN geonames.feature_code       
 IS ' http://www.geonames.org/export/codes.html, varchar(10)';
COMMENT ON COLUMN geonames.country            
 IS ' ISO-3166 2-letter country code, 2 characters';
COMMENT ON COLUMN geonames.cc2                
 IS ' alternate country codes, comma separated, ISO-3166 2-letter country code, 
    200 characters';
COMMENT ON COLUMN geonames.admin1             
 IS ' fipscode (subject to change to iso code), see exceptions below, 
    see file admin1Codes.txt for display names of this code; varchar(20)';
COMMENT ON COLUMN geonames.admin2             
 IS ' code for the second administrative division, a county in the US, 
    see file admin2Codes.txt; varchar(80) ';
COMMENT ON COLUMN geonames.admin3             
 IS ' code for third level administrative division, varchar(20)';
COMMENT ON COLUMN geonames.admin4             
 IS ' code for fourth level administrative division, varchar(20)';
COMMENT ON COLUMN geonames.population         
 IS ' bigint (8 byte int) ';
COMMENT ON COLUMN geonames.elevation          
 IS ' in meters, integer';
COMMENT ON COLUMN geonames.dem                
 IS ' digital elevation model, srtm3 or gtopo30, average elevation of 3''x3'' 
    (ca 90mx90m) or 30''x30'' (ca 900mx900m) area in meters, integer. 
    srtm processed by cgiar/ciat.';
COMMENT ON COLUMN geonames.timezone           
 IS ' the iana timezone id (see file timeZone.txt) varchar(40)';
COMMENT ON COLUMN geonames.modification       
 IS ' date of last modification in yyyy-MM-dd format';
"  #<-- Don't forget the closing quote

ETL

wget http://download.geonames.org/export/dump/allCountries.zip
unzip allCountries.zip

# do this, and go get a coffee.  This took nearly an hour
#   there will be a few lines that fail, they don't really matter much
IFS=$'\n'

for line in $(<allCountries.txt)
do

    echo -n "$line" | 
        psql -qtAc
    "COPY geonames FROM STDIN WITH CSV DELIMITER E'\t';"
2> errors.txt
done

Clean up and Set up

Everything else we do from inside psql:

psql
-- This command requires the installation
--  of postgis2 from your OS package manager.
-- For OS/X that was `port install postgresql12-postgis2`
-- it will be something similar on most platforms.
-- (e.g. apt-get install postgresql12-postgis2, 
--  yum -y install postgresql12-postgis2, etc.)
CREATE EXTENSION postgis;
CREATE EXTENSION postgis_topology;

ALTER TABLE geonames ADD COLUMN location point;

-- Go get another cup of coffee, this is going to rewrite the entire table with the new geo column.
UPDATE geonames SET location = ('(' || latitude || ', ' || longitude || ')')::point;

DELETE FROM geonames WHERE latitude IS NULL or longitude IS NULL;
-- DELETE 32   -- In my case, this ETL anomoly was too small
--  to bother fixing the records

-- Bloat removal from the update and delete operations
CLUSTER geonames USING geonames_pkey;
Share this

More Blogs