我有一个中等大小的数据库,分布在几个表上,粗略的架构是:
- 输入数据(数据 ID、会话 ID 和一些具有统计重要性的字段)
- 输入文件(数据 ID 和 blob)
- 第 1 阶段输出文件(数据 ID 和 blob)
- 第 2 阶段输出文件(数据 ID 和 blob)
- 第 1 类结果(数据 ID 和一些布尔值)
- 2 类结果(数据 ID 和一些整数)
- 第 3 类结果(数据 ID 和一些整数)
每个表有约 200,000 行。
我还有一个视图,它基本上将所有这些粘合在一起,以便我可以SELECT
使用一堆 ID(通常根据会话 ID 选择它们)并在一个页面上查看所有相关数据。
视图工作正常,查询计划的索引利用率看起来很正常,但结果并不快:
> EXPLAIN ANALYZE SELECT(*) FROM overlay WHERE test_session=12345;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Merge Right Join (cost=7.19..74179.49 rows=10 width=305) (actual time=10680.129..10680.494 rows=4 loops=1)
Merge Cond: (p.data_id = d.id)
-> Merge Join (cost=7.19..75077.04 rows=183718 width=234) (actual time=0.192..10434.995 rows=173986 loops=1)
Merge Cond: (p.data_id = input_file.data_id)
-> Merge Join (cost=7.19..69917.74 rows=183718 width=222) (actual time=0.173..9255.653 rows=173986 loops=1)
Merge Cond: (p.data_id = stage1_output_file.data_id)
-> Merge Join (cost=5.50..62948.54 rows=183718 width=186) (actual time=0.153..8081.949 rows=173986 loops=1)
Merge Cond: (p.data_id = stage2_output_file.data_id)
-> Merge Join (cost=3.90..55217.36 rows=183723 width=150) (actual time=0.132..6918.814 rows=173986 loops=1)
Merge Cond: (p.data_id = stage3_output_file.data_id)
-> Nested Loop (cost=2.72..47004.01 rows=183723 width=114) (actual time=0.111..5753.105 rows=173986 loops=1)
Join Filter: (p.impression = istr.id)
-> Merge Join (cost=1.68..30467.90 rows=183723 width=102) (actual time=0.070..2675.733 rows=173986 loops=1)
Merge Cond: (p.data_id = s.data_id)
-> Merge Join (cost=1.68..19031.56 rows=183723 width=58) (actual time=0.049..1501.546 rows=173986 loops=1)
Merge Cond: (p.data_id = t.data_id)
-> Index Scan using Category1_Results_pkey on Category1_Results p (cost=0.00..7652.17 rows=183723 width=18) (actual time=0.025..315.531 rows=173986 loops=1)
-> Index Scan using Category3_Results_pkey on Category3_Results t (cost=0.00..8624.43 rows=183787 width=40) (actual time=0.016..321.460 rows=173986 loops=1)
-> Index Scan using Category2_Results_pkey on Category2_Results s (cost=0.00..8681.47 rows=183787 width=44) (actual time=0.014..320.794 rows=173986 loops=1)
-> Materialize (cost=1.04..1.08 rows=4 width=20) (actual time=0.001..0.007 rows=4 loops=173986)
-> Seq Scan on Category1_impression_str istr (cost=0.00..1.04 rows=4 width=20) (actual time=0.005..0.012 rows=4 loops=1)
-> Index Scan using Stage3_Output_file_pkey on Stage3_Output_file stage3 (cost=0.00..8178.35 rows=183871 width=36) (actual time=0.015..317.698 rows=173986 loops=1)
-> Index Scan using analysis_file_pkey on analysis_file Stage2_Output (cost=0.00..8168.99 rows=183718 width=36) (actual time=0.014..317.776 rows=173986 loops=1)
-> Index Scan using Stage1_output_file_pkey on Stage1_output_file stg1 (cost=0.00..8199.07 rows=183856 width=36) (actual time=0.014..321.648 rows=173986 loops=1)
-> Index Scan using input_file_pkey on input_file input (cost=0.00..8618.05 rows=183788 width=36) (actual time=0.014..328.968 rows=173986 loops=1)
-> Materialize (cost=0.00..39.59 rows=10 width=75) (actual time=0.046..0.150 rows=4 loops=1)
-> Nested Loop Left Join (cost=0.00..39.49 rows=10 width=75) (actual time=0.039..0.128 rows=4 loops=1)
Join Filter: (t.id = d.input_quality)
-> Index Scan using input_data_exists_index on input_data d (cost=0.00..28.59 rows=10 width=45) (actual time=0.013..0.025 rows=4 loops=1)
Index Cond: (test_session = 1040)
-> Seq Scan on quality_codes t (cost=0.00..1.04 rows=4 width=38) (actual time=0.002..0.009 rows=4 loops=4)
Total runtime: 10680.902 ms
其基础视图是我们的“完整结果”视图,定义为:
SELECT p.data_id, p.x2, istr.str AS impression, input.h, p.x3, p.x3, p.x4, s.x5,
s.x6, s.x7, s.x8, s.x9, s.x10, s.x11, s.x12, s.x13, s.x14, t.x15,
t.x16, t.x17, t.x18, t.x19, t.x20, t.x21, t.x22, t.x23,
input.data AS input, stage1_output_file.data AS stage1,
stage2_output_file.data AS stage2, stage3_output_file.data AS stage3
FROM category1_results p, category1_impression_str istr, input_file input,
stage1_output_file, stage2_output_file, stage3_output_file,
category2_results s, category3_results t
WHERE p.impression = istr.id AND p.data_id = input.data_id AND p.data_id = stage1_output_file.data_id
AND p.data_id = stage2_output_file.data_id AND p.data_id = stage3_output_file.data_id AND p.data_id = s.data_id AND p.data_id = t.data_id;
以及生成上述查询计划的覆盖视图,定义为:
SELECT d.data_id, d.test_session, d.a, d.b, t.c, d.d, d.e, d.f, r.*
FROM input_data d LEFT JOIN quality_codes t ON t.id = d.input_quality
LEFT JOIN full_results r ON r.data_id = d.data_id
WHERE NOT d.deleted;
我们似乎在整个链条中的大部分时间都在加入我们的整个数据集,我非常确信这是我们的性能问题——有人对优化这只猪的方法有什么建议吗?
我在这里推测,但我猜你对视图的事实
LEFT JOIN
使计划者在加入查询的第一部分之前从整个视图计算结果。尝试从视图中内联查询并将其设为 a
JOIN
而不是 LEFT JOIN,以查看规划器现在是否找到更快的方法:我清理了您的语法以使其更易于管理。为第二
data_id
列添加了别名,因此它可以执行。如果这会导致执行时间大大加快,您可以尝试添加缺失的行,原因
INNER JOIN
如下:盯着这个看了几天,我很确定一种可能的解决方案是对表进行非规范化并在所有表上粘贴会话 ID。这应该让查询计划器
JOIN
更快地将 s 减少到更小的行子集。这里最大的缺点是非规范化数据库 - 可能不会破坏交易,但如果可能的话我会避免......