按日期分组Spark数据帧 | 按日期分组Spark数据帧

本文介绍了按日期分组Spark数据帧的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！

问题描述

我已经从SQLServer表中加载了DataFrame.看起来像这样:

I've loaded a DataFrame from a SQLServer table. It looks like this:

>>> df.show()
+--------------------+----------+
|           timestamp|    Value |
+--------------------+----------+
|2015-12-02 00:10:...|     652.8|
|2015-12-02 00:20:...|     518.4|
|2015-12-02 00:30:...|     524.6|
|2015-12-02 00:40:...|     382.9|
|2015-12-02 00:50:...|     461.6|
|2015-12-02 01:00:...|     476.6|
|2015-12-02 01:10:...|     472.6|
|2015-12-02 01:20:...|     353.0|
|2015-12-02 01:30:...|     407.9|
|2015-12-02 01:40:...|     475.9|
|2015-12-02 01:50:...|     513.2|
|2015-12-02 02:00:...|     569.0|
|2015-12-02 02:10:...|     711.4|
|2015-12-02 02:20:...|     457.6|
|2015-12-02 02:30:...|     392.0|
|2015-12-02 02:40:...|     459.5|
|2015-12-02 02:50:...|     560.2|
|2015-12-02 03:00:...|     252.9|
|2015-12-02 03:10:...|     228.7|
|2015-12-02 03:20:...|     312.2|
+--------------------+----------+

现在，我想按小时(或日，月或月...)对值进行分组(和求和)，但是我真的不知道如何执行此操作.

Now I'd like to group (and sum) values by hour (or day, or month or...), but I don't really have a clue about how can I do that.

这就是我加载DataFrame的方式.我感觉这不是正确的方法，但是:

That's how I load the DataFrame. I've got the feeling that this isn't the right way to do it, though:

query = """
SELECT column1 AS timestamp, column2 AS value
FROM table
WHERE  blahblah
"""

sc = SparkContext("local", 'test')
sqlctx = SQLContext(sc)

df = sqlctx.load(source="jdbc",
                 url="jdbc:sqlserver://<CONNECTION_DATA>",
                 dbtable="(%s) AS alias" % query)

可以吗?

推荐答案

自1.5.0起，Spark提供了许多功能，例如dayofmonth，hour，month或year，它们可以在日期和时间戳上运行. .因此，如果timestamp是TimestampType，则只需要一个正确的表达式即可.例如:

Since 1.5.0 Spark provides a number of functions like dayofmonth, hour, month or year which can operate on dates and timestamps. So if timestamp is a TimestampType all you need is a correct expression. For example:

from pyspark.sql.functions import hour, mean

(df
    .groupBy(hour("timestamp").alias("hour"))
    .agg(mean("value").alias("mean"))
    .show())

## +----+------------------+
## |hour|              mean|
## +----+------------------+
## |   0|508.05999999999995|
## |   1| 449.8666666666666|
## |   2| 524.9499999999999|
## |   3|264.59999999999997|
## +----+------------------+

1.5.0之前的版本，最好的选择是将HiveContext和Hive UDF与selectExpr一起使用:

Pre-1.5.0 your best option is to use HiveContext and Hive UDFs either with selectExpr:

df.selectExpr("year(timestamp) AS year", "value").groupBy("year").sum()

## +----+---------+----------+
## |year|SUM(year)|SUM(value)|
## +----+---------+----------+
## |2015|    40300|    9183.0|
## +----+---------+----------+

或原始SQL:

df.registerTempTable("df")

sqlContext.sql("""
    SELECT MONTH(timestamp) AS month, SUM(value) AS values_sum
    FROM df
    GROUP BY MONTH(timestamp)""")

请记住，聚合是由Spark执行的，而不是下推到外部源.通常这是一种期望的行为，但是在某些情况下，您可能更愿意将聚合作为子查询来限制数据传输.

Just remember that aggregation is performed by Spark not pushed-down to the external source. Usually it is a desired behavior but there are situations when you may prefer to perform aggregation as a subquery to limit data transfer.

这篇关于按日期分组Spark数据帧的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！