Postgres的：获得最大值和最小值，和时间戳他们发生

Question

我跑的Postgres 9.2，并有温度和时间标记，每分钟一个时间戳毫秒时代时间的表：

weather=# \d weather_data 
     Table "public.weather_data" 
    Column |  Type  | Modifiers 
-------------+--------------+----------- 
timestamp | bigint  | not null 
sensor_id | integer  | not null 
temperature | numeric(4,1) | 
humidity | integer  | 
date  | date   | not null 
Indexes: 
    "weather_data_pkey" PRIMARY KEY, btree ("timestamp", sensor_id) 
    "weather_data_date_idx" btree (date) 
    "weather_data_humidity_idx" btree (humidity) 
    "weather_data_sensor_id_idx" btree (sensor_id) 
    "weather_data_temperature_idx" btree (temperature) 
    "weather_data_time_idx" btree ("timestamp") 
Foreign-key constraints: 
    "weather_data_sensor_id_fkey" FOREIGN KEY (sensor_id) REFERENCES weather_sensors(sensor_id) 

weather=# select * from weather_data order by timestamp desc; 
    timestamp | sensor_id | temperature | humidity | date  
---------------+-----------+-------------+----------+------------ 
1483272420000 |   2 |  22.3 |  57 | 2017-01-01 
1483272420000 |   1 |  24.9 |  53 | 2017-01-01 
1483272360000 |   2 |  22.3 |  57 | 2017-01-01 
1483272360000 |   1 |  24.9 |  58 | 2017-01-01 
1483272300000 |   2 |  22.4 |  57 | 2017-01-01 
1483272300000 |   1 |  24.9 |  57 | 2017-01-01 
[...] 

我有这个现有的查询得到的高点和每一天的低点，但不是具体时间是高还是低发生：

WITH t AS (
    SELECT date, highest, lowest 
    FROM (
     SELECT date, max(temperature) AS highest 
     FROM weather_data 
     WHERE sensor_id = (SELECT sensor_id FROM weather_sensors WHERE sensor_name = 'outdoor') 
     GROUP BY date 
     ORDER BY date ASC 
    ) h 
    INNER JOIN (
     SELECT date, min(temperature) AS lowest 
     FROM weather_data 
     WHERE sensor_id = (SELECT sensor_id FROM weather_sensors WHERE sensor_name = 'outdoor') 
     GROUP BY date 
     ORDER BY date ASC 
    ) l 
    USING (date) 
    ORDER BY date DESC 
) 
SELECT * from t ORDER BY date ASC; 

有一点超过两个百万行的数据库，它需要〜1.2秒运行，这不是 太糟糕了。我想现在得到的具体时间，高或低的是，我想出了这个利用窗口函数，这确实工作，但需要〜5.6秒时：

SELECT h.date, high_time, high_temp, low_time, low_temp FROM (
    SELECT date, high_temp, high_time FROM (
     SELECT date, temperature AS high_temp, timestamp AS high_time, row_number() 
     OVER (PARTITION BY date ORDER BY temperature DESC, timestamp DESC) 
     FROM weather_data 
     WHERE sensor_id = (SELECT sensor_id FROM weather_sensors WHERE sensor_name = 'outdoor') 
    ) highs 
    WHERE row_number = 1 
) h 
INNER JOIN (
    SELECT * FROM (
     SELECT date, temperature AS low_temp, timestamp AS low_time, row_number() 
     OVER (PARTITION BY date ORDER BY temperature ASC, timestamp DESC) 
     FROM weather_data 
     WHERE sensor_id = (SELECT sensor_id FROM weather_sensors WHERE sensor_name = 'outdoor') 
    ) lows 
    WHERE row_number = 1 
) l 
ON h.date = l.date 
ORDER BY h.date ASC; 

有一些相对简单的除我可以做的第一个查询不会增加大量的执行时间？我假设有，但我认为我处于这个问题太久的地步了！

Answer 1

SELECT 
     DISTINCT ON (zdate) zdate 
     , first_value(ztimestamp) OVER www AS stamp_at_min 
     , first_value(temperature) OVER www AS tmin 
     , last_value(ztimestamp) OVER www AS stamp_at_max 
     , last_value(temperature) OVER www AS tmax 
FROM weather_data 
WHERE sensor_id = 2 
WINDOW www AS (PARTITION BY zdate ORDER BY temperature, ztimestamp 
       ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING 
       ) 
     ; 

---

• 前缀ž日期和ž时间戳
• 我加ztimestamp到排序作为决胜

Answer 2

这确实与您的第二个查询，但只需要在weather_data表中的单个扫描：

select date, 
     max(case when high_rn = 1 then timestamp end) as high_time, 
     max(case when high_rn = 1 then temperature end) as high_temp, 
     max(case when low_rn = 1 then timestamp end) as low_time, 
     max(case when low_rn = 1 then temperature end) as low_temp 
from (
    select timestamp, temperature, date, 
     row_number() OVER (PARTITION BY date ORDER BY temperature DESC, timestamp DESC) as high_rn, 
     row_number() OVER (PARTITION BY date ORDER BY temperature ASC, timestamp DESC) as low_rn 
    from weather_data 
    where sensor_id = ... 
) t 
where (high_rn = 1 or low_rn = 1) 
group by date; 

它使用条件聚集做一个交叉表（又名“转动”）的结果查询只包含最低和最高温度。

---

^{无关，而是：date和timestamp是列名可怕。首先是因为它们是关键字，但更重要的是因为它们没有记录列的实际含义。它是“到期日期”吗？ “阅读日期”？ “处理日期”？}