PostgreSQL:未使用索引导致查询性能不好

Question

我们有一个查询来获取在特定时间段内发生更改的所有作业。根据所选择的时段，性能从<每天100毫秒到一周约7秒。

我发现，如果时间段足够小，就会使用索引，并且查询速度很快。如果句点过大，则不会使用索引，查询速度也会变慢。

服务器运行版本为9.2。

为什么会出现这种情况，以及如何解决此问题？

创建脚本：

CREATE TABLE IF NOT EXISTS "Job" 
(
"id" serial PRIMARY KEY,
"serial" TEXT NOT NULL
);
CREATE UNIQUE INDEX "index_Job_serial" ON "Job" ("serial" ASC);
CREATE TABLE IF NOT EXISTS "Property" 
(
"id" serial PRIMARY KEY,
"name" TEXT NOT NULL
);
CREATE TABLE IF NOT EXISTS "Timestamp"
(
"id" serial PRIMARY KEY,
"usSince1970" BIGINT NOT NULL ,
"localTime" TEXT
);
CREATE INDEX "index_Timestamp_usSince1970" ON "Timestamp" USING btree ("usSince1970");
CREATE  TABLE IF NOT EXISTS "Changes" 
(
"idJob" INTEGER  NOT NULL ,
"idProperty" INTEGER  NOT NULL ,
"idTimestamp" INTEGER  NOT NULL ,
"value1" decimal(25,5),
"value2" INTEGER ,
"value3" TEXT ,
PRIMARY KEY ("idJob", "idProperty", "idTimestamp") ,
FOREIGN KEY ("idJob" ) REFERENCES "Job" ("id" ) ,
FOREIGN KEY ("idProperty" ) REFERENCES "Property" ("id" ) ,
FOREIGN KEY ("idTimestamp" ) REFERENCES "Timestamp" ("id" )
);
CREATE INDEX "index_Changes_idJob" ON "Changes" ("idJob" ASC);
CREATE INDEX "index_Changes_idProperty" ON "Changes" ("idProperty" ASC);
CREATE INDEX "index_Changes_idTimestamp" ON "Changes" ("idTimestamp" DESC);

快速查询：

-- fast query (1 day)
SELECT DISTINCT "idJob"
FROM "Changes" 
INNER JOIN "Timestamp" ON "Timestamp"."id" = "Changes"."idTimestamp" 
WHERE "Timestamp"."usSince1970" between 1584831600000000 and 1584745200000000 
-- explain
HashAggregate  (cost=26383.48..26444.33 rows=6085 width=4) (actual time=8.039..8.078 rows=179 loops=1)
->  Nested Loop  (cost=0.00..26368.26 rows=6085 width=4) (actual time=0.031..7.059 rows=6498 loops=1)
->  Index Scan using "index_Timestamp_usSince1970" on "Timestamp"  (cost=0.00..96.25 rows=2510 width=4) (actual time=0.022..0.514 rows=2671 loops=1)
Index Cond: (("usSince1970" >= 1584745200000000::bigint) AND ("usSince1970" <= 1584831600000000::bigint))
->  Index Scan using "index_Changes_idTimestamp" on "Changes"  (cost=0.00..10.27 rows=20 width=8) (actual time=0.002..0.002 rows=2 loops=2671)
Index Cond: ("idTimestamp" = "Timestamp".id)
Total runtime: 8.204 ms

慢速查询：

-- slow query (7 days)
SELECT distinct "idJob"
FROM "Changes" 
INNER JOIN "Timestamp" ON "Timestamp"."id" = "Changes"."idTimestamp" 
WHERE "Timestamp"."usSince1970" between 1583708400000000 and 1584313200000000
-- explain
Unique  (cost=570694.82..571824.16 rows=92521 width=4) (actual time=8869.569..8930.545 rows=3695 loops=1)
->  Sort  (cost=570694.82..571259.49 rows=225867 width=4) (actual time=8869.568..8915.372 rows=260705 loops=1)
Sort Key: "Changes"."idJob"
Sort Method: external merge  Disk: 3552kB
->  Hash Join  (cost=4926.44..547518.97 rows=225867 width=4) (actual time=6325.494..8734.353 rows=260705 loops=1)
Hash Cond: ("Changes"."idTimestamp" = "Timestamp".id)
->  Seq Scan on "Changes"  (cost=0.00..250722.43 rows=16238343 width=8) (actual time=0.004..2505.794 rows=16238343 loops=1)
->  Hash  (cost=3397.79..3397.79 rows=93172 width=4) (actual time=42.392..42.392 rows=107093 loops=1)
Buckets: 4096  Batches: 4  Memory Usage: 948kB
->  Index Scan using "index_Timestamp_usSince1970" on "Timestamp"  (cost=0.00..3397.79 rows=93172 width=4) (actual time=0.006..20.831 rows=107093 loops=1)
Index Cond: (("usSince1970" >= 1583708400000000::bigint) AND ("usSince1970" <= 1584313200000000::bigint))
Total runtime: 8932.374 ms

提前谢谢。

Answer 1

慢速查询处理的数据要多得多(100000行，而"Timestamp"中有2500行(，因此速度较慢也就不足为奇了。

您也可以强制PostgreSQL使用嵌套循环联接和慢速查询：

BEGIN;
SET LOCAL enable_hashjoin = off;
SET LOCAL enable_mergejoin = off;
SELECT ...;
COMMIT;

试试看PostgreSQL是否是对的，散列连接是否真的很慢。

我怀疑PostgreSQL在这里做的是正确的，提高性能的最好方法是增加work_mem。

如果您愿意经常添加另一个索引和VACUUM"Changes"，则使用仅索引扫描可以获得更好的性能：

CREATE INDEX ON "Changes" ("idTimestamp") INCLUDE ("idJob");

在旧版本的PostgreSQL上，这将是

CREATE INDEX ON "Changes" ("idTimestamp", "idJob");

然后最好去掉现在不必要的索引"index_Changes_idTimestamp"。

顺便说一句，你使用骆驼格和引用的标识符让你的生活变得不必要地艰难。

Answer 2

BTW：您的查询相当于：

---

SELECT DISTINCT "idJob"
FROM "Changes" ch
WHERE EXISTS ( 
SELECT * FROM "Timestamp"  ts
WHERE ts.id = ch."idTimestamp" 
AND ts."usSince1970" between 1584831600000000 and 1584745200000000 
);