高级查询
除了基本的连接、聚合、分组操作外,YMatrix 还提供了许多高级分析函数,如窗口函数、CTE(Comman Table Expression)、 有序集聚合函数(Ordered-Set Aggregate Functions)、常用时序函数等。本节将向你介绍 YMatrix 中常用的高级查询方法。
下面以统计磁盘使用量为例,演示如何在 YMatrix 中进行高级查询。在示例指标表 disk 中,我们设计了磁盘的读、写速度字段,使用 MARS2 存储引擎。
MARS2 表依赖 matrixts 扩展,在建表前,首先需要你在使用该存储引擎的数据库中创建扩展。此扩展为数据库级别,无需重复创建。
CREATE EXTENSION matrixts;
创建指标表 disk。
CREATE TABLE disk (
time timestamp with time zone,
tag_id int,
read float,
write float
)
USING mars2
DISTRIBUTED BY (tag_id);
创建MARS2表成功后,你必须额外创建一个 mars2_btree 类型的索引,这样才能进行正常的数据读写。
CREATE INDEX idx_mars2 ON disk
USING mars2_btree(tag_id);
更多 MARS2 相关信息请见 存储引擎
1 窗口函数
窗口函数主要用于处理相对复杂的报表统计分析场景。“窗口”限定了一个数据集合,在与当前行相关的行之间执行聚合查询。窗口函数与其他 SQL 函数的区别在于 OVER 子句的存在。 如果一个函数有一个 OVER 子句,那么它就是一个窗口函数。你可以通过以下表格理解窗口函数与普通聚合函数的区别。
普通聚合函数 | 窗口函数 | |
---|---|---|
输出 | 一条记录 | 多条记录 |
函数 | max()、min()、count()、sum()等 | avg()、sum()、rank()等 |
用法 | 通常与 GROUP BY 子句组合使用 | 与 OVER 子句组合使用。一般情况下,OVER 子句直接写在窗口函数的名称和参数之后。OVER 子句中通常可以使用 PARTITION BY、ORDER BY、ROWS BETWEEN 三种子句。只使用 PARTITION BY 子句会形成静态窗口,窗口大小、位置不会发生变化;除了 PARTITION BY 语句外,还使用了 ORDER BY、ROWS BETWEEN 语句中的一种或两种,会形成滑动窗口,即窗口大小、位置会不断变化 |
你可以根据以下案例实践常用的窗口函数。
1.1 累积和
通过嵌套使用 SUM 方法,可以计算累积和。
如下 SQL 计算了 2021-04-10 21:00:00 到 2021-04-10 21:00:10 间 tag_id 为 1 的磁盘读写累积和:
ymatrix=# SELECT time,
sum(sum(read)) OVER (ORDER BY time) AS read,
sum(sum(write)) OVER (ORDER BY time) AS write
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 21:00:10'::timestamp
AND tag_id = 1
GROUP BY time
ORDER BY time;
time | read | write
------------------------+--------+--------------------
2021-04-10 21:00:00+08 | 81.07 | 73.3
2021-04-10 21:00:01+08 | 110.63 | 121.77
2021-04-10 21:00:02+08 | 202.12 | 201.36
2021-04-10 21:00:03+08 | 263.74 | 257.88
2021-04-10 21:00:04+08 | 361.6 | 299.3
2021-04-10 21:00:05+08 | 394.49 | 327.33000000000004
2021-04-10 21:00:06+08 | 438.3 | 334.98
2021-04-10 21:00:07+08 | 523.35 | 431.39
2021-04-10 21:00:08+08 | 583.15 | 461.84
2021-04-10 21:00:09+08 | 609.01 | 533.03
2021-04-10 21:00:10+08 | 669.52 | 535.9
(11 rows)
1.2 移动平均值
移动平均值用来计算该条记录与前 n 条的平均值。
如下 SQL 计算了 tag_id 为 1 的磁盘,在 2021-4-10 21:00:00 到 21:01:00 每 10 秒的平均读写(对于前 9 条数据,只是计算了满足条件行数的平均值):
ymatrix=# SELECT time,
round(AVG(read) OVER(ORDER BY time ROWS BETWEEN 9 PRECEDING AND CURRENT ROW)) AS read,
round(AVG(write) OVER(ORDER BY time ROWS BETWEEN 9 PRECEDING AND CURRENT ROW)) AS write
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 21:01:00'::timestamp
AND tag_id = 1
ORDER BY time DESC;
time | read | write
------------------------+------+-------
2021-04-10 21:01:00+08 | 57 | 57
2021-04-10 21:00:59+08 | 49 | 60
2021-04-10 21:00:58+08 | 52 | 56
2021-04-10 21:00:57+08 | 51 | 57
2021-04-10 21:00:56+08 | 53 | 65
2021-04-10 21:00:55+08 | 48 | 64
2021-04-10 21:00:54+08 | 49 | 64
2021-04-10 21:00:53+08 | 47 | 54
2021-04-10 21:00:52+08 | 44 | 54
2021-04-10 21:00:51+08 | 41 | 56
......
1.3 增量
增量通常用来计算对于一个单调序列增幅或降幅,也可以简单的用来计算与前一条数据的变化。
如下语句计算了 tag_id 为 1 的磁盘,在 2021-4-10 21:00:00 到 21:01:00 期间磁盘读的变化值,正数为相比上一秒增长,负数为相比上一秒下降:
ymatrix=# SELECT
time,
(
CASE WHEN lag(read) OVER (ORDER BY time) IS NULL THEN NULL
ELSE round(read - lag(read) OVER (ORDER BY time))
END
) AS read
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 21:01:00'::timestamp
AND tag_id = 1
ORDER BY time;
time | read
------------------------+------
2021-04-10 21:00:00+08 |
2021-04-10 21:00:01+08 | -52
2021-04-10 21:00:02+08 | 62
2021-04-10 21:00:03+08 | -30
2021-04-10 21:00:04+08 | 36
2021-04-10 21:00:05+08 | -65
2021-04-10 21:00:06+08 | 11
2021-04-10 21:00:07+08 | 41
2021-04-10 21:00:08+08 | -25
2021-04-10 21:00:09+08 | -34
......
1.4 增速
在增量的基础上,再除以时间间隔,就可以得到增速(因为样例数据是 1 秒采样一次,所以看到的和增量的结果相同):
ymatrix=# SELECT
time,
(
CASE WHEN lag(read) OVER (ORDER BY time) IS NULL THEN NULL
ELSE round(read - lag(read) OVER (ORDER BY time))
END
) / extract(epoch from time - lag(time) OVER (ORDER BY time)) AS read_rate,
extract(epoch from time - lag(time) OVER (ORDER BY time)) AS "time lag"
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 21:01:00'::timestamp
AND tag_id = 1
ORDER BY time;
time | read_rate | time lag
------------------------+-----------+----------
2021-04-10 21:00:00+08 | |
2021-04-10 21:00:01+08 | -52 | 1
2021-04-10 21:00:02+08 | 62 | 1
2021-04-10 21:00:03+08 | -30 | 1
2021-04-10 21:00:04+08 | 36 | 1
2021-04-10 21:00:05+08 | -65 | 1
2021-04-10 21:00:06+08 | 11 | 1
2021-04-10 21:00:07+08 | 41 | 1
2021-04-10 21:00:08+08 | -25 | 1
2021-04-10 21:00:09+08 | -34 | 1
......
1.5 变化点
变化点列出相比于前一条有变化的记录,该类型查询适合在比较平稳的数据集中找发生变化的点:
ymatrix=# SELECT time, read FROM (
SELECT time,
read,
read - lag(read) OVER (ORDER BY TIME) AS diff
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 21:01:00'::timestamp
AND tag_id = 1 ) ht
WHERE diff IS NULL OR diff != 0
ORDER BY time;
time | read
------------------------+-------
2021-04-10 21:00:00+08 | 81.07
2021-04-10 21:00:01+08 | 29.56
2021-04-10 21:00:02+08 | 91.49
2021-04-10 21:00:03+08 | 61.62
2021-04-10 21:00:04+08 | 97.86
2021-04-10 21:00:05+08 | 32.89
2021-04-10 21:00:06+08 | 43.81
2021-04-10 21:00:07+08 | 85.05
2021-04-10 21:00:08+08 | 59.8
2021-04-10 21:00:09+08 | 25.86
(10 rows)
2 有序集聚合函数(Ordered-Set Aggregate Functions)
2.1 计算连续百分率
你可以使用以下有序集聚合函数来计算连续百分率。
percentile_cont(fractions) WITHIN GROUP (ORDER BY sort_expression)
示例及返回结果如下。
ymatrix=# SELECT tag_id,
percentile_cont(0.2) WITHIN GROUP
(ORDER BY read) AS read,
percentile_cont(0.3) WITHIN GROUP (ORDER BY write) AS write
FROM disk
GROUP BY tag_id
ORDER BY tag_id;
tag_id | read | write
--------+-------+-------
1 | 19.87 | 29.86
2 | 19.95 | 29.88
3 | 20.06 | 29.93
(3 rows)
参数表示百分比,如果传 0.5 相当于计算中位数。
3. 通用表表达式(Comman Table Expression)
通用表表达式即 CTE,你可以使用它定义构建一个临时的视图,从而使大型查询语句简化。它通过 WITH 关键字实现。在使用之前,可通过以下表格理解 CTE 与 CREATE VIEW 语句的区别。
CTE | CREATE VIEW | |
---|---|---|
表述 | 并非独立SQL语句,而是语句的一部分,即表达式 | 独立SQL语句 |
应用范围 | 建立的临时视图只用于所属查询语句 | 建立的视图可用于所有查询语句 |
如下 CTE 计算了所有设备中读速度平均值的最大值和最小值:
WITH avg_read (tag_id, avg_read) AS (
SELECT tag_id, AVG(read) AS read FROM disk GROUP BY tag_id
) SELECT MAX(avg_read), MIN(avg_read) FROM avg_read;
4. 时序函数
YMatrix 提供的时序组件 matrixts
中还提供了时序场景中经常使用的时序函数,首先要创建该组件:
stats=# CREATE EXTENSION matrixts;
4.1 time_bucket
time_bucket 可以计算出给定时间段的平均值。
参数名 | 说明 | 数据类型 | 可缺省 |
---|---|---|---|
period | 时间窗口大小 | int16;int32;int64;interval | 否 |
timestamp | 需要被转换的列 | int16;int32;int64;date;timestamp;timestamptz | 否 |
如下 SQL 计算了 tag_id 为 1 的磁盘,在 2021-4-10 21:00:00 到 22:00:00 之间每 5 分钟的平均读写速度:
ymatrix=# SELECT time_bucket('5 minutes', time) AS five_min,
AVG(read) as read,
AVG(write) as write
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 22:00:00'::timestamp
AND tag_id = 1
GROUP BY five_min
ORDER BY five_min;
five_min | read | write
------------------------+--------------------+--------------------
2021-04-10 21:00:00+08 | 48.614599999999996 | 49.48656666666666
2021-04-10 21:05:00+08 | 50.73533333333335 | 49.992566666666654
2021-04-10 21:10:00+08 | 51.6102333333333 | 49.99359999999999
2021-04-10 21:15:00+08 | 49.29116666666669 | 53.89146666666666
2021-04-10 21:20:00+08 | 49.67863333333332 | 50.47406666666665
2021-04-10 21:25:00+08 | 51.09013333333332 | 47.766733333333335
2021-04-10 21:30:00+08 | 49.55949999999999 | 50.440766666666654
2021-04-10 21:35:00+08 | 48.86253333333333 | 50.57290000000001
2021-04-10 21:40:00+08 | 51.061299999999974 | 47.028766666666684
2021-04-10 21:45:00+08 | 52.10353333333333 | 49.861466666666665
2021-04-10 21:50:00+08 | 51.780566666666694 | 51.4159
2021-04-10 21:55:00+08 | 51.83549999999998 | 49.124366666666674
2021-04-10 22:00:00+08 | 93.96 | 91.07
(13 rows)
4.2 time_bucket_gapfill
当时间段中有数据缺失,需要进行数据清洗,可以使用 time_bucket_gapfill 函数为缺失数据做填充,使数据按指定时间间隔在时序上平滑分布,从而便于分析。有两种填充策略:locf (last observation carried forward) 和 interpolate。
- locf:用聚合组中之前出现的值填充
- interpolate:对缺失的值做线性插值填充
假设示例 disk 表中数据情况如下(只适用于 4.2 节):
ymatrix=# SELECT * FROM disk ORDER BY tag_id;
time | tag_id | read | write
------------------------+--------+------+-------
2021-04-10 21:00:00+08 | 1 | 3.4 | 4.6
2021-04-10 21:50:00+08 | 1 | 4 | 2.7
2021-04-10 21:40:00+08 | 1 | 8.4 | 12
2021-04-10 21:20:00+08 | 1 | 2.9 | 6
2021-04-10 21:30:00+08 | 1 | 9 | 10.2
2021-04-10 21:10:00+08 | 1 | 5.2 | 6.6
2021-04-10 22:00:00+08 | 1 | 10 | 7
(7 rows)
使用 time_bucket_gapfill 函数对 tag_id 为 1 的磁盘, 2021-4-10 21:00:00 到 22:00:00 之间每隔 5 分钟进行数据填充,得到结果。
SELECT time_bucket_gapfill('5 minutes', time) AS five_min,
locf(AVG(read)) as locf_read,
interpolate(AVG(read)) as interpolate_read
FROM disk
WHERE time BETWEEN '2021-04-10 21:00:00'::timestamp AND '2021-04-10 22:00:00'::timestamp
AND tag_id = 1
GROUP BY five_min
ORDER BY five_min;
five_min | locf_read | interpolate_read
------------------------+-----------+------------------
2021-04-10 21:00:00+08 | 3.4 | 3.4
2021-04-10 21:05:00+08 | 3.4 | 4.3
2021-04-10 21:10:00+08 | 5.2 | 5.2
2021-04-10 21:15:00+08 | 5.2 | 4.05
2021-04-10 21:20:00+08 | 2.9 | 2.9
2021-04-10 21:25:00+08 | 2.9 | 5.95
2021-04-10 21:30:00+08 | 9 | 9
2021-04-10 21:35:00+08 | 9 | 8.7
2021-04-10 21:40:00+08 | 8.4 | 8.4
2021-04-10 21:45:00+08 | 8.4 | 6.2
2021-04-10 21:50:00+08 | 4 | 4
2021-04-10 21:55:00+08 | 4 | 7
2021-04-10 22:00:00+08 | 10 | 10
(13 rows)
4.3 first/last
first 返回时间最早的值:
ymatrix=# SELECT tag_id,
first(read, time) AS read,
first(write, time) AS write
FROM disk
GROUP BY tag_id
ORDER BY tag_id;
tag_id | read | write
--------+-------+-------
1 | 11.51 | 86.61
2 | 50.07 | 25.9
3 | 83.72 | 10.5
(3 rows)
last 返回时间最晚的值:
ymatrix=# SELECT tag_id,
last(read, time) AS read,
last(write, time) AS write
FROM disk
GROUP BY tag_id
ORDER BY tag_id;
tag_id | read | write
--------+-------+-------
1 | 5.32 | 4.96
2 | 5.73 | 34.73
3 | 49.03 | 86.02
(3 rows)
4.4 last_not_null_value
last_not_null_value 相当于在 last 基础上增加了 not null 的过滤,返回最后一个非空的值:
ymatrix=# SELECT last_not_null_value(read, time)
FROM disk WHERE tag_id = 1;
last_not_null_value
---------------------
3.1
(1 row)
4.5 last_not_null
last_not_null 和 last_not_null_value 相比,不仅返回值,时间也会返回。返回的类型是字符串,格式为 '["value", "time"]':
ymatrix=# SELECT last_not_null(read, time)
FROM disk WHERE tag_id = 1;
last_not_null
-----------------------------------------
["3.1","2021-11-05 17:32:51.754457+08"]
(1 row)