PostgreSQL 生成空间热力图

6 minute read

背景

结合空间数据，计算基于地理位置信息的热力图，在空间数据可视化场景中是一个非常常见的需求。

结合流计算，可以实现实时的热力图计算。

结合并行计算，可以高效率的对海量数据进行热力图计算。

计算热力图中bucket的方法

https://www.postgresql.org/docs/devel/static/functions-math.html

width_bucket(operand dp, b1 dp, b2 dp, count int)	  
int	  
return the bucket number to which operand would be assigned in a histogram having count equal-width buckets spanning the range b1 to b2;   
returns 0 or count+1 for an input outside the range	  
width_bucket(5.35, 0.024, 10.06, 5)	  
3  
  
width_bucket(operand numeric, b1 numeric, b2 numeric, count int)	  
int	  
return the bucket number to which operand would be assigned in a histogram having count equal-width buckets spanning the range b1 to b2;   
returns 0 or count+1 for an input outside the range	  
width_bucket(5.35, 0.024, 10.06, 5)	  
3  

例如

postgres=# select width_bucket(1,1,10,10);  
 width_bucket   
--------------  
            1  
(1 row)  
  
postgres=# select width_bucket(0,1,10,10);  
 width_bucket   
--------------  
            0  
(1 row)  
  
postgres=# select width_bucket(10,1,10,10);  
 width_bucket   
--------------  
           11  
(1 row)  
  
postgres=# select width_bucket(9.9,1,10,10);  
 width_bucket   
--------------  
           10  
(1 row)  

width_bucket(  
  p1 -- 输入值  
  p2 -- 边界值（最小，包含）  
  p3 -- 边界值（最大，不包含）  
  p4 -- 切割份数  
)  
  
当小于最小边界值时，返回0  
当大于等于最大边界值时，返回p4+1  

例如x轴的边界是1,10000，y轴的边界是1,10000。

x,y两个方向分别切割为50个bucket，一共2500个bucket，求一个点落在哪个bucket：

width_bucket(pos[0], 1, 10001, 50),  -- x轴落在哪列bucket  
width_bucket(pos[1], 1, 10001, 50),  -- y轴落在哪列bucket  

例子

1、建表

create table tbl_pos(  
  id int,    
  info text,   -- 信息  
  val float8,  -- 取值  
  pos point    -- 位置  
);  

2、写入1亿个点

vi test.sql  
insert into tbl_pos values ( random()*100000, md5(random()::text), random()*1000, point((random()*10000::int), (random()*10000::int)) );  
  
pgbench -M prepared -n -r -P 1 -f ./test.sql -c 50 -j 50 -t 2000000

3、热力图计算

强制并行计算

postgres=# set min_parallel_table_scan_size =0;  
SET  
postgres=# set min_parallel_index_scan_size =0;  
SET  
postgres=# set parallel_setup_cost =0;  
SET  
postgres=# set parallel_tuple_cost =0;  
SET  
postgres=# set max_parallel_workers_per_gather =28;  
SET  
postgres=# alter table tbl_pos set (parallel_workers =28);  
ALTER TABLE  

热力图计算SQL，效率还不错：

select   
  width_bucket(pos[0], 0, 10001, 50),  -- x轴落在哪列bucket  
  width_bucket(pos[1], 0, 10001, 50),  -- y轴落在哪列bucket  
  avg(val),  
  min(val),  
  max(val),  
  stddev(val),  
  count(*)  
from tbl_pos  
  group by 1,2;  
  
  
                                                                                      QUERY PLAN                                                                                         
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------  
 Finalize GroupAggregate  (cost=1252812.00..1252928.00 rows=200 width=48) (actual time=2632.324..2672.909 rows=2500 loops=1)  
   Group Key: (width_bucket(pos[0], '0'::double precision, '10001'::double precision, 50)), (width_bucket(pos[1], '0'::double precision, '10001'::double precision, 50))  
   ->  Sort  (cost=1252812.00..1252826.00 rows=5600 width=96) (actual time=2632.290..2648.544 rows=72500 loops=1)  
         Sort Key: (width_bucket(pos[0], '0'::double precision, '10001'::double precision, 50)), (width_bucket(pos[1], '0'::double precision, '10001'::double precision, 50))  
         Sort Method: external merge  Disk: 9824kB  
         ->  Gather  (cost=1252460.37..1252463.37 rows=5600 width=96) (actual time=2532.132..2564.905 rows=72500 loops=1)  
               Workers Planned: 28  
               Workers Launched: 28  
               ->  Partial HashAggregate  (cost=1252460.37..1252463.37 rows=200 width=96) (actual time=2522.428..2523.559 rows=2500 loops=29)  
                     Group Key: width_bucket(pos[0], '0'::double precision, '10001'::double precision, 50), width_bucket(pos[1], '0'::double precision, '10001'::double precision, 50)  
                     ->  Parallel Seq Scan on tbl_pos  (cost=0.00..1189951.79 rows=3571919 width=16) (actual time=0.030..1302.462 rows=3448276 loops=29)  
 Planning time: 0.154 ms  
 Execution time: 2676.288 ms  
(13 rows)  

样本

postgres=# select                   
  width_bucket(pos[0], 0, 10001, 10),  -- x轴落在哪列bucket  
  width_bucket(pos[1], 0, 10001, 10),  -- y轴落在哪列bucket  
  avg(val),  
  min(val),  
  max(val),  
  stddev(val),  
  count(*)  
from tbl_pos  
  group by 1,2;  
 width_bucket | width_bucket |       avg        |         min          |       max        |      stddev      |  count    
--------------+--------------+------------------+----------------------+------------------+------------------+---------  
            1 |            1 | 499.638668709335 | 0.000637955963611603 | 999.998900108039 | 288.562996477433 | 1002686  
            1 |            2 | 499.772206697849 |  0.00113388523459435 | 999.999452847987 | 288.505295714968 | 1000891  
            1 |            3 |  500.44455454312 |  0.00135181471705437 | 999.997937120497 |  288.45102360668 |  999911  
            1 |            4 | 500.234164866407 |  0.00214902684092522 | 999.999100342393 | 288.707167816157 | 1000473  
            1 |            5 | 499.793710464008 | 0.000125262886285782 | 999.999575316906 | 288.672382834812 |  999036  
            1 |            6 | 500.366854944369 |  0.00212574377655983 | 999.999585561454 | 288.558891852102 |  998866  
            1 |            7 | 499.825623783545 | 0.000547617673873901 | 999.999700114131 | 288.582317248892 | 1000902  
            1 |            8 | 499.393569281915 |  0.00330200418829918 | 999.999083112925 | 288.561094278074 | 1000193  
            1 |            9 | 499.713056248083 |  0.00243959948420525 | 999.999618623406 | 288.709997455837 | 1000017  
            1 |           10 | 500.312448499828 |  0.00238511711359024 | 999.999850522727 | 288.865560266629 |  998469  
            2 |            1 | 499.848655048635 |  0.00146497040987015 | 999.999508261681 | 288.639402346948 | 1000917  
            2 |            2 | 500.084846394446 |   0.0005294568836689 | 999.999178107828 | 288.704696698903 |  997594  
            2 |            3 |  499.99258346144 |  0.00163912773132324 |  999.99839020893 | 288.507497234907 | 1001310  
            2 |            4 | 499.817295558208 |  0.00184541568160057 | 999.997940845788 | 288.767308817191 | 1000607  
            2 |            5 |  499.87314410326 |  0.00135786831378937 | 999.999302905053 | 288.593077096809 |  998588  
            2 |            6 | 499.825467223571 | 0.000847037881612778 | 999.998526647687 | 288.789326889728 | 1000426  
            2 |            7 |  499.50907809986 |  7.4971467256546e-05 |   999.9989871867 | 288.535982009648 | 1001179  
            2 |            8 | 499.850422744194 | 0.000966247171163559 | 999.999921303242 | 288.516738657089 | 1000745  
            2 |            9 | 500.110417044655 | 0.000320374965667725 | 999.999660998583 |  288.77420504779 |  999978  
            2 |           10 | 500.135548004555 | 0.000233296304941177 | 999.999852851033 | 288.520964728395 |  998363  
........  

取出数据，即可渲染。

结合流计算，可以在FEED数据写入时，实时的进行计算。而不是QUERY发起时计算。参考本文末尾的文档。

小结

PostgreSQL非常适合于时空数据的分析，包括本文提到的热力图分析。

使用并行计算，即查即算，1亿个点，差不多耗时2.7秒。

如果使用流式计算，写入时即算，查询时查的是结果，效率更高。

参考

1、求bucket值

https://www.postgresql.org/docs/devel/static/functions-math.html

width_bucket(operand dp, b1 dp, b2 dp, count int)	  
int	  
return the bucket number to which operand would be assigned in a histogram having count equal-width buckets spanning the range b1 to b2;   
returns 0 or count+1 for an input outside the range	  
width_bucket(5.35, 0.024, 10.06, 5)	  
3  
  
width_bucket(operand numeric, b1 numeric, b2 numeric, count int)	  
int	  
return the bucket number to which operand would be assigned in a histogram having count equal-width buckets spanning the range b1 to b2;   
returns 0 or count+1 for an input outside the range	  
width_bucket(5.35, 0.024, 10.06, 5)	  
3  

2、求geometry对象的x,y,z值

http://postgis.net/docs/manual-2.4/reference.html

ST_X — Return the X coordinate of the point, or NULL if not available. Input must be a point.  
ST_XMax — Returns X maxima of a bounding box 2d or 3d or a geometry.  
ST_XMin — Returns X minima of a bounding box 2d or 3d or a geometry.  
ST_Y — Return the Y coordinate of the point, or NULL if not available. Input must be a point.  
ST_YMax — Returns Y maxima of a bounding box 2d or 3d or a geometry.  
ST_YMin — Returns Y minima of a bounding box 2d or 3d or a geometry.  
ST_Z — Return the Z coordinate of the point, or NULL if not available. Input must be a point.  
ST_ZMax — Returns Z minima of a bounding box 2d or 3d or a geometry.  
ST_Zmflag — Returns ZM (dimension semantic) flag of the geometries as a small int. Values are: 0=2d, 1=3dm, 2=3dz, 3=4d.  
ST_ZMin — Returns Z minima of a bounding box 2d or 3d or a geometry.  

3、求point对象的x,y值