(警告道歉和黑客入侵......)
背景:
我有一个遗留应用程序,我想避免重写它的大量 SQL 代码。我正在尝试加速它所做的特定类型的非常昂贵的查询(即:低悬的果实)。
它有一个由transactions
表表示的金融交易分类帐。当插入新行时,触发函数(此处未显示)为给定实体结转新余额。
某些类型的交易模型外部性(如进行中的支付)通过使用“相关”交易标记新交易,以便应用程序可以将相关交易组合在一起。
\d transactions
Table "public.transactions"
Column | Type | Modifiers
---------------------+-----------+-----------
entityid | bigint | not null
transactionid | bigint | not null default nextval('tid_seq')
type | smallint | not null
status | smallint | not null
related | bigint |
amount | bigint | not null
abs_amount | bigint | not null
is_credit | boolean | not null
inserted | timestamp | not null default now()
description | text | not null
balance | bigint | not null
Indexes:
"transactions_pkey" PRIMARY KEY, btree (transactionid)
"transactions by entityid" btree (entityid)
"transactions by initial trans" btree ((COALESCE(related, transactionid)))
Foreign-key constraints:
"invalid related transaction!" FOREIGN KEY (related)
REFERENCES transactions(transactionid)
在我的测试数据集中,我有:
- 总共大约 550 万行
- 大约 370 万行没有“相关”事务
- 大约 180 万行“相关”交易
- 大约 55k 个不同的 entityids(客户)。
因此,大约 1/3 的事务行是与一些早期事务“相关”的更新。生产数据大约大 25 倍transactionid
,不同大约大 8 倍entityid
,1/3 的比率用于事务更新。
该代码查询一个特别低效的 VIEW 定义为:
CREATE VIEW collapsed_transactions AS
SELECT t.entityid,
g.initial,
g.latest,
i.inserted AS created,
t.inserted AS updated,
t.type,
t.status,
t.amount,
t.abs_amount,
t.is_credit,
t.balance,
t.description
FROM ( SELECT
COALESCE(x.related, x.transactionid) AS initial,
max(x.transactionid) AS latest
FROM transactions x
GROUP BY COALESCE(x.related, x.transactionid)
) g
INNER JOIN transactions t ON t.transactionid = g.latest
INNER JOIN transactions i ON i.transactionid = g.initial;
典型的查询采用以下形式:
SELECT * FROM collapsed_transactions WHERE entityid = 204425;
如您所见,该where entityid = 204425
子句不会用于约束GROUP BY
子查询,因此所有实体的事务将被分组,从而产生 55,000 个更大的子查询结果集和更长的查询时间......所有这些都达到平均 40 行(本例中为 71)在撰写本文时。
在不重写代码库的数百个 SQL 查询的情况下,我无法transactions
进一步规范化表(比如将表连接到),其中许多以不同的方式使用自连接语义。initial_transactions
updated_transactions
related
洞察力:
我最初尝试使用 WINDOW 函数重写查询,但遇到了各种各样的问题(另一个 SO 问题再次出现),当我看到www_fdw将其 WHERE 子句作为 GET/POST 参数传递给 HTTP 时,我很感兴趣无需太多重组即可优化非常幼稚的查询的可能性。
F.31.4。远程查询优化
postgres_fdw 尝试优化远程查询以减少从外部服务器传输的数据量。这是通过将查询 WHERE 子句发送到远程服务器执行,并且不检索当前查询不需要的表列来完成的。为了降低查询错误执行的风险,WHERE 子句不会发送到远程服务器,除非它们仅使用内置数据类型、运算符和函数。子句中的运算符和函数也必须是 IMMUTABLE。
可以使用 EXPLAIN VERBOSE 检查实际发送到远程服务器执行的查询。
试图:
所以我认为也许我可以将 GROUP-BY 放入一个视图中,将该视图视为一个外部表,并且优化器将通过 WHERE 子句传递到该外部表,从而产生更有效的查询......
CREATE VIEW foreign_transactions_grouped_by_initial_transaction AS
SELECT
entityid,
COALESCE(t.related, t.transactionid) AS initial,
MAX(t.transactionid) AS latest
FROM transactions t
GROUP BY
t.entityid,
COALESCE(t.related, t.transactionid);
CREATE FOREIGN TABLE transactions_grouped_by_initial_transaction
(entityid bigint, initial bigint, latest bigint)
SERVER local_pg_server
OPTIONS (table_name 'foreign_transactions_grouped_by_initial_transaction');
EXPLAIN ANALYSE VERBOSE
SELECT
t.entityid,
g.initial,
g.latest,
i.inserted AS created,
t.inserted AS updated,
t.type,
t.status,
t.amount,
t.abs_amount,
t.is_credit,
t.balance,
t.description
FROM transactions_grouped_by_initial_transaction g
INNER JOIN transactions t on t.transactionid = g.latest
INNER JOIN transactions i on i.transactionid = g.initial
WHERE g.entityid = 204425;
效果很好!
Nested Loop (cost=100.87..305.05 rows=10 width=116)
(actual time=4.113..16.646 rows=71 loops=1)
Output: t.entityid, g.initial, g.latest, i.inserted,
t.inserted, t.type, t.status, t.amount, t.abs_amount,
t.balance, t.description
-> Nested Loop (cost=100.43..220.42 rows=10 width=108)
(actual time=4.017..10.725 rows=71 loops=1)
Output: g.initial, g.latest, t.entityid, t.inserted,
t.type, t.status, t.amount, t.abs_amount, t.is_credit,
t.balance, t.description
-> Foreign Scan on public.transactions_grouped_by_initial_transaction g
(cost=100.00..135.80 rows=10 width=16)
(actual time=3.914..4.694 rows=71 loops=1)
Output: g.entityid, g.initial, g.latest
Remote SQL:
SELECT initial, latest
FROM public.foreign_transactions_grouped_by_initial_transaction
WHERE ((entityid = 204425))
-> Index Scan using transactions_pkey on public.transactions t
(cost=0.43..8.45 rows=1 width=100)
(actual time=0.023..0.035 rows=1 loops=71)
Output: t.entityid, t.transactionid, t.type, t.status,
t.related, t.amount, t.abs_amount, t.is_credit,
t.inserted, t.description, t.balance
Index Cond: (t.transactionid = g.latest)
-> Index Scan using transactions_pkey on public.transactions i
(cost=0.43..8.45 rows=1 width=16)
(actual time=0.021..0.033 rows=1 loops=71)
Output: i.entityid, i.transactionid, i.type, i.status,
i.related, i.amount, i.abs_amount, i.is_credit,
i.inserted, i.description, i.balance
Index Cond: (i.transactionid = g.initial)
Total runtime: 20.363 ms
问题:
但是,当我尝试将其烘焙到 VIEW 中(有或没有另一层postgres_fdw
)时,查询优化器似乎没有通过 WHERE 子句:-(
CREATE view collapsed_transactions_fast AS
SELECT
t.entityid,
g.initial,
g.latest,
i.inserted AS created,
t.inserted AS updated,
t.type,
t.status,
t.amount,
t.abs_amount,
t.is_credit,
t.balance,
t.description
FROM transactions_grouped_by_initial_transaction g
INNER JOIN transactions t on t.transactionid = g.latest
INNER JOIN transactions i on i.transactionid = g.initial;
EXPLAIN ANALYSE VERBOSE
SELECT * FROM collapsed_transactions_fast WHERE entityid = 204425;
结果是:
Nested Loop (cost=534.97..621.88 rows=1 width=117)
(actual time=104720.383..139307.940 rows=71 loops=1)
Output: t.entityid, g.initial, g.latest, i.inserted, t.inserted, t.type,
t.status, t.amount, t.abs_amount, t.is_credit, t.balance,
t.description
-> Hash Join (cost=534.53..613.66 rows=1 width=109)
(actual time=104720.308..139305.522 rows=71 loops=1)
Output: g.initial, g.latest, t.entityid, t.inserted, t.type,
t.status, t.amount, t.abs_amount, t.is_credit, t.balance,
t.description
Hash Cond: (g.latest = t.transactionid)
-> Foreign Scan on public.transactions_grouped_by_initial_transaction g
(cost=100.00..171.44 rows=2048 width=16)
(actual time=23288.569..108916.051 rows=3705600 loops=1)
Output: g.entityid, g.initial, g.latest
Remote SQL:
SELECT initial, latest
FROM public.foreign_transactions_grouped_by_initial_transaction
-> Hash (cost=432.76..432.76 rows=142 width=101)
(actual time=2.103..2.103 rows=106 loops=1)
Output:
t.entityid, t.inserted, t.type, t.status, t.amount,
t.abs_amount, t.is_credit, t.balance, t.description,
t.transactionid
Buckets: 1024 Batches: 1 Memory Usage: 14kB
-> Index Scan using "transactions by entityid"
on public.transactions t
(cost=0.43..432.76 rows=142 width=101)
(actual time=0.049..1.241 rows=106 loops=1)
Output: t.entityid, t.inserted, t.type, t.status,
t.amount, t.abs_amount, t.is_credit,
t.balance, t.description, t.transactionid
Index Cond: (t.entityid = 204425)
-> Index Scan using transactions_pkey on public.transactions i
(cost=0.43..8.20 rows=1 width=16)
(actual time=0.013..0.018 rows=1 loops=71)
Output: i.entityid, i.transactionid, i.type, i.status, i.related,
i.amount, i.abs_amount, i.is_credit, i.inserted, i.description,
i.balance
Index Cond: (i.transactionid = g.initial)
Total runtime: 139575.140 ms
如果我可以将该行为烘焙到 VIEW 或 FDW 中,那么我可以在极少数查询中替换 VIEW的名称,以提高效率。我不在乎对于其他用例(更复杂的 WHERE 子句)是否超级慢,我将命名 VIEW 以反映其预期用途。
有其use_remote_estimate
默认值,FALSE
但无论哪种方式都没有区别。
问题:
我可以使用一些技巧来使这个公认的黑客工作吗?