问题描述
CREATE TABLE tsvector_business_objects (
id int4 NULL,"type" varchar NULL,value varchar NULL,label varchar NULL,synonyms _text NULL,label_tsvector tsvector NULL
);
CREATE INDEX label_tsvector_idx ON tsvector_business_objects USING gin (label_tsvector);
CREATE INDEX lower_case_synonym_tsvector_business_objects ON tsvector_business_objects USING gin (synonyms);
CREATE INDEX txt_label_tsvector_business_objects ON tsvector_business_objects USING gin (to_tsvector('english'::regconfig,lower((label)::text)));
CREATE INDEX txt_value_tsvector_business_objects ON tsvector_business_objects USING gin (to_tsvector('english'::regconfig,lower((value)::text)));
CREATE INDEX type_tsvector_business_objects ON tsvector_business_objects USING btree (type);
CREATE INDEX type_value_label_lower_case_tsvector_business_objects ON tsvector_business_objects USING btree (lower((type)::text),lower((value)::text),lower((label)::text));
我在数据中有一些像 OID-0127820 这样的文本，当我使用 ts_vector 搜索令牌 OID 时，Postgres 总是使用序列扫描而不是使用 GIN 索引。
有 500k 条带有 OID 文本的记录。但是当我禁用序列扫描时，索引正在被使用
SET enable_seqscan = ON;
explain (analyze,verbose,buffers,timing,costs)
SELECT TYPE,value,label,synonyms,((to_tsvector('english',lower(value)) @@ plainto_tsquery('english','OID') AND numnode(plainto_tsquery('english','OID')) > 0) OR (lower(value) LIKE '% OID %' AND numnode(plainto_tsquery('english','OID')) = 0) OR ((to_tsvector('english',lower(label))) @@ plainto_tsquery('english','OID')) > 0) OR (lower(label) LIKE '% OID %' AND numnode(plainto_tsquery('english','OID')) = 0)) as partial_value_label_match,(lower(value) ='OID' OR lower(label) ='OID') as exact_value_label_match,(synonyms @> '{OID}') is true as synonym_match FROM tsvector_business_objects AS business_objects_alias
WHERE synonyms @> '{OID}'
OR (to_tsvector('english','OID')) > 0)
OR (lower(value) LIKE '% OID %' AND numnode(plainto_tsquery('english','OID')) = 0)
OR ((to_tsvector('english','OID')) > 0)
OR (lower(label) LIKE '% OID %' AND numnode(plainto_tsquery('english','OID')) = 0) OR (lower(value) ='OID' OR lower(label) ='OID')
limit 30
QUERY PLAN |
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
Limit (cost=0.00..41.91 rows=30 width=82) (actual time=3804.998..3805.163 rows=30 loops=1) |
Output: type,(((to_tsvector('english'::regconfig,lower((value)::text)) @@ '''oid'''::tsquery) OR (to_tsvector('english'::regconfig,lower((label)::text)) @@ '''oid'''::tsquery))),(((lower((value)::text) = 'OID'::text) OR (lowe|
Buffers: shared hit=21217 |
-> Seq Scan on mdlz_performancebenchmarking.tsvector_business_objects business_objects_alias (cost=0.00..717278.98 rows=513425 width=82) (actual time=3804.997..3805.159 rows=30 loops=1) |
Output: type,((to_tsvector('english'::regconfig,lower((label)::text)) @@ '''oid'''::tsquery)),((lower((value)::text) = 'OID'::text) OR (l|
Filter: ((business_objects_alias.synonyms @> '{OID}'::text[]) OR (to_tsvector('english'::regconfig,lower((business_objects_alias.value)::text)) @@ '''oid'''::tsquery) OR (to_tsvector('english'::regconfig,lower((business_objects_alias.label)::tex|
Rows Removed by Filter: 575042 |
Buffers: shared hit=21217 |
Planning time: 0.226 ms |
Execution time: 3805.210 ms
SET enable_seqscan = OFF;
explain (analyze,'OID')) = 0) OR (lower(value) ='OID' OR lower(label) ='OID')
limit 30
QUERY PLAN |
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
Limit (cost=66996.67..67036.72 rows=30 width=82) (actual time=312.131..312.251 rows=30 loops=1) |
Output: type,(((lower((value)::text) = 'OID'::text) OR (lowe|
Buffers: shared hit=45103 |
-> Bitmap Heap Scan on mdlz_performancebenchmarking.tsvector_business_objects business_objects_alias (cost=66996.67..752444.39 rows=513425 width=82) (actual time=312.130..312.247 rows=30 loops=1) |
Output: type,((lower((value)::text) = 'OID'::text) OR (l|
Recheck Cond: ((business_objects_alias.synonyms @> '{OID}'::text[]) OR (to_tsvector('english'::regconfig,lower((business_objects_alias.label|
Heap Blocks: exact=1 |
Buffers: shared hit=45103 |
-> BitmapOr (cost=66996.67..66996.67 rows=666282 width=0) (actual time=310.732..310.732 rows=0 loops=1) |
Buffers: shared hit=45102 |
-> Bitmap Index Scan on lower_case_synonym_tsvector_business_objects (cost=0.00..189.20 rows=1 width=0) (actual time=0.008..0.008 rows=0 loops=1) |
Index Cond: (business_objects_alias.synonyms @> '{OID}'::text[]) |
Buffers: shared hit=2 |
-> Bitmap Index Scan on txt_value_tsvector_business_objects (cost=0.00..2638.25 rows=333140 width=0) (actual time=49.131..49.131 rows=575000 loops=1) |
Index Cond: (to_tsvector('english'::regconfig,lower((business_objects_alias.value)::text)) @@ '''oid'''::tsquery) |
Buffers: shared hit=86 |
-> Bitmap Index Scan on txt_label_tsvector_business_objects (cost=0.00..2638.25 rows=333140 width=0) (actual time=49.956..49.956 rows=575000 loops=1) |
Index Cond: (to_tsvector('english'::regconfig,lower((business_objects_alias.label)::text)) @@ '''oid'''::tsquery) |
Buffers: shared hit=86 |
-> Bitmap Index Scan on type_value_label_lower_case_tsvector_business_objects (cost=0.00..30444.59 rows=1 width=0) (actual time=93.168..93.169 rows=0 loops=1) |
Index Cond: (lower((business_objects_alias.value)::text) = 'OID'::text) |
Buffers: shared hit=22464 |
-> Bitmap Index Scan on type_value_label_lower_case_tsvector_business_objects (cost=0.00..30444.59 rows=1 width=0) (actual time=118.465..118.465 rows=0 loops=1) |
Index Cond: (lower((business_objects_alias.label)::text) = 'OID'::text) |
Buffers: shared hit=22464 |
Planning time: 0.249 ms |
Execution time: 312.579 ms
输出数据
type |value |label |synonyms|label_tsvector |partial_value_label_match|exact_value_label_match|synonym_match|
---------------------|-----------|-----------|--------|--------------------|-------------------------|-----------------------|-------------|
orderid_1621409737948|OID-0127820|OID-0127820|NULL |'-0127820':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0355880|OID-0355880|NULL |'-0355880':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0041048|OID-0041048|NULL |'-0041048':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0499716|OID-0499716|NULL |'-0499716':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0219268|OID-0219268|NULL |'-0219268':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0000560|OID-0000560|NULL |'-0000560':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0355656|OID-0355656|NULL |'-0355656':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0355628|OID-0355628|NULL |'-0355628':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0219380|OID-0219380|NULL |'-0219380':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0063896|OID-0063896|NULL |'-0063896':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0054740|OID-0054740|NULL |'-0054740':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0447020|OID-0447020|NULL |'-0447020':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0040964|OID-0040964|NULL |'-0040964':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0499744|OID-0499744|NULL |'-0499744':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0219800|OID-0219800|NULL |'-0219800':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0355180|OID-0355180|NULL |'-0355180':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0530348|OID-0530348|NULL |'-0530348':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0219996|OID-0219996|NULL |'-0219996':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0220024|OID-0220024|NULL |'-0220024':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0446936|OID-0446936|NULL |'-0446936':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0220108|OID-0220108|NULL |'-0220108':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0020692|OID-0020692|NULL |'-0020692':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0354872|OID-0354872|NULL |'-0354872':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0097496|OID-0097496|NULL |'-0097496':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0354648|OID-0354648|NULL |'-0354648':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0128268|OID-0128268|NULL |'-0128268':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0354536|OID-0354536|NULL |'-0354536':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0530432|OID-0530432|NULL |'-0530432':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0128324|OID-0128324|NULL |'-0128324':2 'oid':1|true |false |false |
orderid_1621409737948|OID-0354256|OID-0354256|NULL |'-0354256':2 'oid':1|true |false |false |
random_page_cost 为 1.1。
我认为 Postgres 查询规划器认为使用序列扫描会更快，但事实并非如此。
解决方法
问题在于许多条件高度相关，因为它们测试几乎相同的东西。由于优化器不知道这一点并将它们视为统计独立的，因此它得出了错误的估计。
您可以尝试简化 WHERE 条件，或者您可以升级到 PostgreSQL v14，其中 commit a4d75c86bf1 引入了表达式的扩展统计信息。
一个基本问题是规划器认为它会在 seq 扫描中很早就找到 30 行，因此可以很早就停止。这是错误的，显然是因为与您的条件匹配的行很少出现在表的早期部分。仅通过改进选择性估计很难解决这个问题，因为即使行估计是正确的，但行没有均匀地散布在表中，它也会犯这个错误。