Author: digoal
imgsmlr是PostgreSQL的一款支持以图搜图的插件， 支持
1、几种图像特征值数据类型，
2、图像特征值相似算子，
3、图像特征值相似排序索引支持，
4、图像相似排序的索引（通过扩展GiST索引接口实现）支持，
5、png,gif等图像格式特征值提取函数。
图像特征值为64*64的16个区域经过小波转换后的16个浮点数。
在数据量（图片数）非常庞大时，输入一个图片特征值，搜索相似度排行前N的图片，性能如何呢？如何优化呢？
接下来的3篇文档将分别介绍如下三种场景的图像特征值搜索性能以及优化思路：
https://github.com/digoal/blog/blob/master/201809/20180904_03.md
https://github.com/digoal/blog/blob/master/201809/20180904_04.md
1、在需要使用图像搜索的数据库中创建插件
create extension imgsmlr;
1、创建生成随机图像特征值signature的UDF。
create or replace function gen_rand_img_sig(int) returns signature as $$
select ('('||rtrim(ltrim(array(select (random()*$1)::float4 from generate_series(1,16))::text,'{'),'}')||')')::signature;
$$ language sql strict;
postgres=# select * from gen_rand_img_sig(10);
gen_rand_img_sig
------------------------------------------------------------------------------------------------------------------------------------------------------------------
(6.744310, 5.105020, 0.087113, 3.808010, 8.129480, 2.834540, 2.495250, 0.940481, 0.033208, 6.583490, 2.840330, 1.422440, 6.683830, 0.080847, 8.327730, 2.471430)
(1 row)
postgres=# select * from gen_rand_img_sig(10);
gen_rand_img_sig
------------------------------------------------------------------------------------------------------------------------------------------------------------------
(3.013650, 6.170690, 0.601905, 2.692030, 1.268540, 7.803740, 9.757770, 5.537750, 0.391753, 4.440790, 1.201580, 5.501380, 6.166980, 0.240686, 9.768680, 2.911290)
(1 row)
2、建表，建图像特征值索引
create table t_img_sig (id int primary key, sig signature);
create index idx_t_img_sig_1 on t_img_sig using gist(sig);
3、写入约4亿随机图像特征值
vi testsig.sql
\set id random(1,2000000000)
insert into t_img_sig values (:id, gen_rand_img_sig(10)) on conflict(id) do nothing;
pgbench -M prepared -n -r -P 1 -f ./testsig.sql -c 32 -j 32 -t 20000000
postgres=# select * from t_img limit 10;
id | sig
-----------+------------------------------------------------------------------------------------------------------------------------------------------------------------------
47902935 | (5.861920, 1.062770, 8.318020, 2.205840, 0.202951, 6.956610, 1.413190, 2.898480, 8.961630, 6.377800, 1.110450, 6.684520, 2.286290, 7.850760, 1.832650, 0.074348)
174656795 | (2.165030, 0.183753, 9.913950, 9.208260, 5.165660, 6.603510, 2.008380, 8.117910, 2.358590, 5.466330, 9.139280, 8.893700, 4.664190, 9.361670, 9.016990, 2.271000)
96186891 | (9.605980, 4.395920, 4.336720, 3.174360, 8.706960, 0.155107, 9.408940, 4.531100, 2.783530, 5.681780, 9.792380, 6.428320, 2.983760, 9.733290, 7.635160, 7.035780)
55061667 | (7.567960, 5.874530, 5.222040, 5.638520, 3.488960, 8.770750, 7.054610, 7.239630, 9.202280, 9.465020, 4.079080, 5.729770, 0.475227, 8.434800, 6.873730, 5.140080)
64659434 | (4.860650, 3.984440, 3.009900, 5.116680, 6.489150, 4.224800, 0.609752, 8.731120, 6.577390, 8.542540, 9.096120, 8.976700, 8.936000, 2.836270, 7.186250, 6.264300)
87143098 | (4.801570, 7.870150, 0.939599, 3.666670, 1.102340, 5.819580, 6.511330, 6.430760, 0.584531, 3.024190, 6.255460, 8.823820, 5.076960, 0.181344, 8.137380, 1.230360)
109245945 | (7.541850, 7.201460, 6.858400, 2.605210, 1.283090, 7.525200, 4.213240, 8.413760, 9.707390, 1.916970, 1.719320, 1.255280, 9.006780, 4.851420, 2.168250, 5.997360)
4979218 | (8.463000, 4.051410, 9.057320, 1.367980, 3.344340, 7.032640, 8.583770, 1.873090, 5.524810, 0.187254, 5.783270, 6.141040, 2.479410, 6.406450, 9.371700, 0.050690)
72846137 | (7.018560, 4.039150, 9.114800, 2.911170, 5.531180, 8.557330, 6.739050, 0.103649, 3.691390, 7.584640, 8.184180, 0.599390, 9.037130, 4.090610, 4.369770, 6.480000)
36813995 | (4.643480, 8.704640, 1.073880, 2.665530, 3.298300, 9.244280, 5.768050, 0.887555, 5.990350, 2.991390, 6.186550, 6.464940, 6.187140, 0.150242, 2.123070, 2.932270)
(10 rows)
Time: 58.101 ms
写入约4.39亿图像特征值。
postgres=# select count(*) from t_img_sig;
count
-----------
438924137
(1 row)
4、输入一个图像特征值，搜索与之最相似的图像。
explain (analyze,verbose,timing,costs,buffers) select * from t_img_sig order by sig <-> '(5.07998,6.80827,5.42024,2.53619,4.10843,0.532198,4.33886,9.60262,6.68369,8.01305,9.60298,8.087,1.25819,6.54424,6.04902,5.3444)' limit 1;
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from t_img_sig where signature_distance(sig,'(5.07998,6.80827,5.42024,2.53619,4.10843,0.532198,4.33886,9.60262,6.68369,8.01305,9.60298,8.087,1.25819,6.54424,6.04902,5.3444)') > 0.9 order by sig <-> '(5.07998,6.80827,5.42024,2.53619,4.10843,0.532198,4.33886,9.60262,6.68369,8.01305,9.60298,8.087,1.25819,6.54424,6.04902,5.3444)' limit 1;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.48..0.51 rows=1 width=72) (actual time=4094.810..4094.812 rows=1 loops=1)
Output: id, sig, ((sig <-> '(5.079980, 6.808270, 5.420240, 2.536190, 4.108430, 0.532198, 4.338860, 9.602620, 6.683690, 8.013050, 9.602980, 8.087000, 1.258190, 6.544240, 6.049020, 5.344400)'::signature))
Buffers: shared hit=205999
-> Index Scan using idx_t_img_sig_1 on public.t_img_sig (cost=0.48..5361351.06 rows=146395778 width=72) (actual time=4094.808..4094.808 rows=1 loops=1)
Output: id, sig, (sig <-> '(5.079980, 6.808270, 5.420240, 2.536190, 4.108430, 0.532198, 4.338860, 9.602620, 6.683690, 8.013050, 9.602980, 8.087000, 1.258190, 6.544240, 6.049020, 5.344400)'::signature)
Order By: (t_img_sig.sig <-> '(5.079980, 6.808270, 5.420240, 2.536190, 4.108430, 0.532198, 4.338860, 9.602620, 6.683690, 8.013050, 9.602980, 8.087000, 1.258190, 6.544240, 6.049020, 5.344400)'::signature)
Filter: (signature_distance(t_img_sig.sig, '(5.079980, 6.808270, 5.420240, 2.536190, 4.108430, 0.532198, 4.338860, 9.602620, 6.683690, 8.013050, 9.602980, 8.087000, 1.258190, 6.544240, 6.049020, 5.344400)'::signature) > '0.9'::double precision)
Buffers: shared hit=205999
Planning Time: 0.073 ms
Execution Time: 4194.485 ms
(10 rows)
性能：4.39亿图像特征值，以图搜图约4.2秒。
瓶颈:
1、扫描了大量的索引页（205999）。
1、压缩精度，比如使用3位小数。据用户说有10倍性能提升。
精度优化如下，使用新的生成图像特征值的函数，使用3位小数。
create or replace function gen_rand_img_sig3(int) returns signature as $$
select ('('||rtrim(ltrim(array(select trunc((random()*$1)::numeric,3) from generate_series(1,16))::text,'{'),'}')||')')::signature;
$$ language sql strict;
例子如下
postgres=# select gen_rand_img_sig3(10);
gen_rand_img_sig3
------------------------------------------------------------------------------------------------------------------------------------------------------------------
(2.984000, 3.323000, 4.083000, 6.292000, 5.008000, 9.029000, 6.208000, 1.141000, 1.796000, 9.257000, 1.397000, 1.235000, 7.157000, 3.745000, 0.112000, 7.723000)
(1 row)
2、使用分区表+dblink异步接口并行调用。（内核层面直接支持imgsmlr gist index scan并行更好）
下一篇介绍
3、使用citus sharding。多机，提高整体计算能力。 （因为扫描大量索引页，即使CPU没有瓶颈，将来内存带宽也会成为瓶颈。多机可以解决这个问题。）
下一篇介绍
4、内核层面，支持维度更低的signature，现在是16片，比如支持降低到4片，性能也可以提升。
1、当有记录可以完全匹配时，扫描少量INDEX PAGE。
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from t_img_sig order by sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179000)'::signature limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.48..0.49 rows=1 width=72) (actual time=1.596..1.598 rows=1 loops=1)
Output: id, sig, ((sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179000)'::signature))
Buffers: shared hit=125
-> Index Scan using t_img_sig1_sig_idx on public.t_img_sig (cost=0.48..7318159.22 rows=785457848 width=72) (actual time=1.594..1.595 rows=1 loops=1)
Output: id, sig, (sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179000)'::signature)
Order By: (t_img_sig.sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179000)'::signature)
Buffers: shared hit=125
Planning Time: 0.072 ms
Execution Time: 1.621 ms
(9 rows)
2、当修改少量内容，少量值完全匹配，其他值不完全匹配时，扫描的INDEX PAGE增加。
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from t_img_sig order by sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179001)'::signature limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.48..0.49 rows=1 width=72) (actual time=7.051..7.052 rows=1 loops=1)
Output: id, sig, ((sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179001)'::signature))
Buffers: shared hit=454
-> Index Scan using t_img_sig1_sig_idx on public.t_img_sig (cost=0.48..7324626.56 rows=786152016 width=72) (actual time=7.049..7.049 rows=1 loops=1)
Output: id, sig, (sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179001)'::signature)
Order By: (t_img_sig.sig <-> '(3.727000, 2.594000, 0.185000, 3.996000, 6.450000, 7.126000, 5.499000, 1.540000, 8.239000, 6.262000, 2.053000, 2.566000, 4.522000, 6.929000, 1.582000, 2.179001)'::signature)
Buffers: shared hit=454
Planning Time: 0.074 ms
Execution Time: 7.076 ms
(9 rows)
3、当大量修改值，不能完全匹配时，需要扫描大量INDEX PAGE。
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from t_img_sig order by sig <-> '(7.727000, 3.594000, 1.185000, 4.996000, 6.950000, 7.129000, 5.429000, 1.520000, 8.219000, 6.222000, 2.013000, 2.536000, 4.532000, 6.939000, 1.538000, 2.178000)'::signature limit 1;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.47..0.48 rows=1 width=72) (actual time=2528.890..2528.891 rows=1 loops=1)
Output: id, sig, ((sig <-> '(7.727000, 3.594000, 1.185000, 4.996000, 6.950000, 7.129000, 5.429000, 1.520000, 8.219000, 6.222000, 2.013000, 2.536000, 4.532000, 6.939000, 1.538000, 2.178000)'::signature))
Buffers: shared hit=121510
-> Index Scan using t_img_sig1_sig_idx on public.t_img_sig (cost=0.47..1361409.21 rows=146121007 width=72) (actual time=2528.887..2528.888 rows=1 loops=1)
Output: id, sig, (sig <-> '(7.727000, 3.594000, 1.185000, 4.996000, 6.950000, 7.129000, 5.429000, 1.520000, 8.219000, 6.222000, 2.013000, 2.536000, 4.532000, 6.939000, 1.538000, 2.178000)'::signature)
Order By: (t_img_sig.sig <-> '(7.727000, 3.594000, 1.185000, 4.996000, 6.950000, 7.129000, 5.429000, 1.520000, 8.219000, 6.222000, 2.013000, 2.536000, 4.532000, 6.939000, 1.538000, 2.178000)'::signature)
Buffers: shared hit=121510
Planning Time: 0.092 ms
Execution Time: 2582.558 ms
(9 rows)
具体原因可以参考
https://github.com/postgrespro/imgsmlr/blob/master/imgsmlr_idx.c
https://www.postgresql.org/docs/devel/static/xindex.html
https://github.com/postgrespro/imgsmlr
《PostgreSQL 相似搜索插件介绍大汇总 (rum,pg_trgm,smlar,imgsmlr,pg_similarity) (rum,gin,gist)》