mil*_*oxe 15 sql postgresql exists bigdata sql-delete
有两个表由id链接:
item_tbl (id)
link_tbl (item_id)
有些记录中item_tbl没有匹配的行link_tbl.一个可以计算其金额的选择将是:
SELECT COUNT(*)
FROM link_tbl lnk LEFT JOIN item_tbl itm ON lnk.item_id=itm.id
WHERE itm.id IS NULL
我想删除那些孤儿记录(那些在其他表中没有匹配的记录),link_tbl但我能想到的唯一方法是:
DELETE FROM link_tbl lnk
WHERE lnk.item_id NOT IN (SELECT itm.id FROM item_tbl itm)
在2003年的16,844,347个孤儿记录中有
3,033,811个记录的link_tbl
262,086,253条记录.
服务器有4GB RAM和8核CPU.item_tbl
link_tbl
EXPLAIN DELETE FROM link_tbl lnk
WHERE lnk.item_id NOT IN (SELECT itm.id FROM item_tbl itm)
返回:
Delete on link lnk (cost=0.00..11395249378057.98 rows=131045918 width=6)
-> Seq Scan on link lnk (cost=0.00..11395249378057.98 rows=131045918 width=6)
Filter: (NOT (SubPlan 1))
SubPlan 1
-> Materialize (cost=0.00..79298.10 rows=3063207 width=4)
-> Seq Scan on item itm (cost=0.00..52016.07 rows=3063207 width=4)
问题是:
link_tbl？上面的解释有多准确,或删除这些记录需要多长时间？
解析度:
谢谢大家的建议,非常有帮助.我终于使用了Erwin Brandstetter建议的删除/sf/answers/1117192751/,但我稍微调整了一下:
DELETE FROM link_tbl lnk
WHERE lnk.item_id BETWEEN 0 AND 10000
AND lnk.item_id NOT IN (SELECT itm.id FROM item itm
WHERE itm.id BETWEEN 0 AND 10000)
我比较了NOT IN和NOT EXISTS的结果,输出在下面,虽然我使用COUNT而不是DELETE,我认为应该是相同的(我的意思是为了相对比较):
EXPLAIN ANALYZE SELECT COUNT(*)
FROM link_tbl lnk
WHERE lnk.item_id BETWEEN 0 AND 20000
AND lnk.item_id NOT IN (SELECT itm.id
FROM item_tbl itm
WHERE itm.id BETWEEN 0 AND 20000);
QUERY PLAN
Aggregate (cost=6002667.56..6002667.57 rows=1 width=0) (actual time=226817.086..226817.088 rows=1 loops=1)
-> Seq Scan on link_tbl lnk (cost=1592.50..5747898.65 rows=101907564 width=0) (actual time=206.029..225289.570 rows=566625 loops=1)
Filter: ((item_id >= 0) AND (item_id <= 20000) AND (NOT (hashed SubPlan 1)))
SubPlan 1
-> Index Scan using item_tbl_pkey on item_tbl itm (cost=0.00..1501.95 rows=36221 width=4) (actual time=0.056..99.266 rows=17560 loops=1)
Index Cond: ((id >= 0) AND (id <= 20000))
Total runtime: 226817.211 ms
EXPLAIN ANALYZE SELECT COUNT(*)
FROM link_tbl lnk WHERE lnk.item_id>0 AND lnk.item_id<20000
AND NOT EXISTS (SELECT 1 FROM item_tbl itm WHERE itm.id=lnk.item_id);
QUERY PLAN
Aggregate (cost=8835772.00..8835772.01 rows=1 width=0)
(actual time=1209235.133..1209235.135 rows=1 loops=1)
-> Hash Anti Join (cost=102272.16..8835771.99 rows=1 width=0)
(actual time=19315.170..1207900.612 rows=566534 loops=1)
Hash Cond: (lnk.item_id = itm.id)
-> Seq Scan on link_tbl lnk (cost=0.00..5091076.55 rows=203815128 width=4) (actual time=0.016..599147.604 rows=200301872 loops=1)
Filter: ((item_id > 0) AND (item_id < 20000))
-> Hash (cost=52016.07..52016.07 rows=3063207 width=4) (actual time=19313.976..19313.976 rows=3033811 loops=1)
Buckets: 131072 Batches: 4 Memory Usage: 26672kB
-> Seq Scan on item_tbl itm (cost=0.00..52016.07 rows=3063207 width=4) (actual time=0.013..9274.158 rows=3033811 loops=1)
Total runtime: 1209260.228 ms
NOT EXISTS慢了5倍.
只要我担心,实际删除的数据就没有了,我能够分5批删除它(10000-20000,20000-100000,100000-200000,200000-1000000和1000000-1755441).起初我发现了max item_id,我只需要经过一半的表.
当我尝试NOT IN或EXISTS没有范围(选择计数)它甚至没有完成,我让它在夜间运行,它仍然在早上运行.
我想我正在寻找使用来自wildplasser的回答/sf/answers/1119162341/的 DELETE 但它来得太晚了.
DELETE FROM one o
USING (
SELECT o2.id
FROM one o2
LEFT JOIN two t ON t.one_id = o2.id
WHERE t.one_id IS NULL
) sq
WHERE sq.id = o.id
;
wil*_*ser 17
我使用{work_mem,effective_cache_size,random_page_cost}的不同设置对四个典型查询进行基准测试,这些设置对所选计划的影响最大.我首先使用我的默认设置"运行"来加热缓存.注意:测试集足够小,可以在缓存中显示所有需要的页面.
测试集
SET search_path=tmp;
/************************/
DROP SCHEMA tmp CASCADE;
CREATE SCHEMA tmp ;
SET search_path=tmp;
CREATE TABLE one
( id SERIAL NOT NULL PRIMARY KEY
, payload varchar
);
CREATE TABLE two
( id SERIAL NOT NULL PRIMARY KEY
, one_id INTEGER REFERENCES one
, payload varchar
);
INSERT INTO one (payload) SELECT 'Text_' || gs::text FROM generate_series(1,30000) gs;
INSERT INTO two (payload) SELECT 'Text_' || gs::text FROM generate_series(1,30000) gs;
UPDATE two t
SET one_id = o.id
FROM one o
WHERE o.id = t.id
AND random() < 0.1;
INSERT INTO two (one_id,payload) SELECT one_id,payload FROM two;
INSERT INTO two (one_id,payload) SELECT one_id,payload FROM two;
INSERT INTO two (one_id,payload) SELECT one_id,payload FROM two;
VACUUM ANALYZE one;
VACUUM ANALYZE two;
/***************/
查询:
\echo NOT EXISTS()
EXPLAIN ANALYZE
DELETE FROM one o
WHERE NOT EXISTS ( SELECT * FROM two t
WHERE t.one_id = o.id
);
\echo NOT IN()
EXPLAIN ANALYZE
DELETE FROM one o
WHERE o.id NOT IN ( SELECT one_id FROM two t)
;
\echo USING (subquery self LEFT JOIN two where NULL)
EXPLAIN ANALYZE
DELETE FROM one o
USING (
SELECT o2.id
FROM one o2
LEFT JOIN two t ON t.one_id = o2.id
WHERE t.one_id IS NULL
) sq
WHERE sq.id = o.id
;
\echo USING (subquery self WHERE NOT EXISTS(two)))
EXPLAIN ANALYZE
DELETE FROM one o
USING (
SELECT o2.id
FROM one o2
WHERE NOT EXISTS ( SELECT *
FROM two t WHERE t.one_id = o2.id
)
) sq
WHERE sq.id = o.id
;
结果(总结)
NOT EXISTS() NOT IN() USING(LEFT JOIN NULL) USING(NOT EXISTS)
1) rpc=4.0.csz=1M wmm=64 80.358 14389.026 77.620 72.917
2) rpc=4.0.csz=1M wmm=64000 60.527 69.104 51.851 51.004
3) rpc=1.5.csz=1M wmm=64 69.804 10758.480 80.402 77.356
4) rpc=1.5.csz=1M wmm=64000 50.872 69.366 50.763 53.339
5) rpc=4.0.csz=1G wmm=64 84.117 7625.792 69.790 69.627
6) rpc=4.0.csz=1G wmm=64000 49.964 67.018 49.968 49.380
7) rpc=1.5.csz=1G wmm=64 68.567 3650.008 70.283 69.933
8) rpc=1.5.csz=1G wmm=64000 49.800 67.298 50.116 50.345
legend:
rpc := "random_page_cost"
csz := "effective_cache_size"
wmm := "work_mem"
如您所见,该NOT IN()变体对缺乏非常敏感work_mem.同意,设置64(KB)非常低,但这个"或多或少"对应于大数据集,它们也不适合哈希表.
EXTRA:在暖入阶段,NOT EXISTS()查询遭受极端FK触发器争用.这是与真空守护体发生冲突的结果,真空守护在表格设置后仍然有效:
PostgreSQL 9.1.2 on x86_64-unknown-linux-gnu, compiled by gcc (Ubuntu/Linaro 4.6.1-9ubuntu3) 4.6.1, 64-bit
NOT EXISTS()
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------
Delete on one o (cost=6736.00..7623.94 rows=27962 width=12) (actual time=80.596..80.596 rows=0 loops=1)
-> Hash Anti Join (cost=6736.00..7623.94 rows=27962 width=12) (actual time=49.174..61.327 rows=27050 loops=1)
Hash Cond: (o.id = t.one_id)
-> Seq Scan on one o (cost=0.00..463.00 rows=30000 width=10) (actual time=0.003..5.156 rows=30000 loops=1)
-> Hash (cost=3736.00..3736.00 rows=240000 width=10) (actual time=49.121..49.121 rows=23600 loops=1)
Buckets: 32768 Batches: 1 Memory Usage: 1015kB
-> Seq Scan on two t (cost=0.00..3736.00 rows=240000 width=10) (actual time=0.006..33.790 rows=240000 loops=1)
Trigger for constraint two_one_id_fkey: time=467720.117 calls=27050
Total runtime: 467824.652 ms
(9 rows)
首先:你的文字说:
我想从中删除那些孤儿记录
item_tbl.
但是你的代码说:
DELETE FROM link_tbl lnk ...更新:
在重新读取QI时,发现它更有可能要删除孤立的行link_tbl.行数指向该方向.@Lucas)查询在这种情况下是正确的.但是我很害怕,NOT EXISTS是实际速度较慢比NOT IN在这种情况下.
为了验证我运行了一个测试用例,这就像你的设置一样远程.无法使它大得多,或者SQLfiddle会遇到超时.
NOT EXISTS反转的情况会更快.(我也测试过这个.)EXISTS更适合测试"很多".通常情况下,有更多的收获而EXISTS不是NOT EXISTS- 无论如何,这种形式必须检查整个表格.这是更难证明的东西不存在,而不是证明的东西存在.这个普遍的事实也适用于数据库.
此操作适合拆分.特别是如果您有并发事务(但即使没有),我会考虑将其拆分DELETE为多个片段,以便事务可以COMMIT在相当长的时间之后完成.
就像是:
DELETE FROM link_tbl l
WHERE l.item_id < 1000000
AND l.item_id NOT IN (SELECT i.id FROM item_tbl i)
然后l.item_id BETWEEN 100001 AND 200000,等等
您无法使用函数自动执行此操作.这会把所有东西都包装成一个交易并且无视目的.所以你必须从任何客户端编写脚本.
或者你可以使用..
这个附加模块允许您在任何数据库中运行单独的事务,包括它运行的事务.这可以通过持久连接完成,这应该可以消除大部分连接开销.有关如何安装它的说明:
如何在PostgreSQL中使用(安装)dblink？
DO会做的工作(PostgreSQL 9.0或更高版本).一次运行DELETE50000个50000命令item_id:
DO
$$
DECLARE
_sql text;
BEGIN
PERFORM dblink_connect('port=5432 dbname=mydb'); -- your connection parameters
FOR i IN 0 .. 100
LOOP
_sql := format('
DELETE FROM link_tbl l
WHERE l.item_id BETWEEN %s AND %s
AND l.item_id NOT IN (SELECT i.id FROM item_tbl i)'
, (50000 * i)::text
, (50000 * (i+1))::text);
PERFORM dblink_exec(_sql);
END LOOP;
PERFORM dblink_disconnect();
END
$$
如果脚本应该被中断:dblink_connect向DB日志写入它执行的内容,这样你就可以看到已经完成的操作了.
也许这：
DELETE FROM link_tbl lnk
WHERE NOT EXISTS
( SELECT 1 FROM item_tbl item WHERE item.id = lnk.item_id );
当处理大量记录时，创建临时表，执行INSERT INTO SELECT * FROM ...然后删除原始表，重命名临时表，然后再添加索引可能会更加高效。