PySpark入门级学习教程，框架思维（中）
SAMshare
共 27164字，需浏览 55分钟
·
2021-04-21 05:30
“这周工作好忙，晚上陆陆续续写了好几波，周末来一次集合输出，不过这个PySpark原定是分上下两篇的，但是越学感觉越多，所以就分成了3 Parts，今天这一part主要就是讲一下Spark SQL，这个实在好用！建议收藏学习哈哈。上一节的可点击回顾下哈。《PySpark入门级学习教程，框架思维（上）》
🎳 Spark SQL使用
在讲Spark SQL前，先解释下这个模块。这个模块是Spark中用来处理结构化数据的，提供一个叫SparkDataFrame的东西并且自动解析为分布式SQL查询数据。我们之前用过Python的Pandas库，也大致了解了DataFrame，这个其实和它没有太大的区别，只是调用的API可能有些不同罢了。
我们通过使用Spark SQL来处理数据，会让我们更加地熟悉，比如可以用SQL语句、用SparkDataFrame的API或者Datasets API，我们可以按照需求随心转换，通过SparkDataFrame API 和 SQL 写的逻辑，会被Spark优化器Catalyst自动优化成RDD，即便写得不好也可能运行得很快（如果是直接写RDD可能就挂了哈哈）。
创建SparkDataFrame
开始讲SparkDataFrame，我们先学习下几种创建的方法，分别是使用RDD来创建、使用python的DataFrame来创建、使用List来创建、读取数据文件来创建、通过读取数据库来创建。
1. 使用RDD来创建
主要使用RDD的toDF方法。
rdd = sc.parallelize([("Sam", 28, 88), ("Flora", 28, 90), ("Run", 1, 60)])
df = rdd.toDF(["name", "age", "score"])
df.show()
df.printSchema()
# +-----+---+-----+
# | name|age|score|
# +-----+---+-----+
# | Sam| 28| 88|
# |Flora| 28| 90|
# | Run| 1| 60|
# +-----+---+-----+
# root
# |-- name: string (nullable = true)
# |-- age: long (nullable = true)
# |-- score: long (nullable = true)
2. 使用python的DataFrame来创建
df = pd.DataFrame([['Sam', 28, 88], ['Flora', 28, 90], ['Run', 1, 60]],
columns=['name', 'age', 'score'])
print(">> 打印DataFrame:")
print(df)
print("\n")
Spark_df = spark.createDataFrame(df)
print(">> 打印SparkDataFrame:")
Spark_df.show()
# >> 打印DataFrame:
# name age score
# 0 Sam 28 88
# 1 Flora 28 90
# 2 Run 1 60
# >> 打印SparkDataFrame:
# +-----+---+-----+
# | name|age|score|
# +-----+---+-----+
# | Sam| 28| 88|
# |Flora| 28| 90|
# | Run| 1| 60|
# +-----+---+-----+
3. 使用List来创建
list_values = [['Sam', 28, 88], ['Flora', 28, 90], ['Run', 1, 60]]
Spark_df = spark.createDataFrame(list_values, ['name', 'age', 'score'])
Spark_df.show()
# +-----+---+-----+
# | name|age|score|
# +-----+---+-----+
# | Sam| 28| 88|
# |Flora| 28| 90|
# | Run| 1| 60|
# +-----+---+-----+
4. 读取数据文件来创建
# 4.1 CSV文件
df = spark.read.option("header", "true")\
.option("inferSchema", "true")\
.option("delimiter", ",")\
.csv("./test/data/titanic/train.csv")
df.show(5)
df.printSchema()
# 4.2 json文件
df = spark.read.json("./test/data/hello_samshare.json")
df.show(5)
df.printSchema()
5. 通过读取数据库来创建
# 5.1 读取hive数据
spark.sql("CREATE TABLE IF NOT EXISTS src (key INT, value STRING) USING hive")
spark.sql("LOAD DATA LOCAL INPATH 'data/kv1.txt' INTO TABLE src")
df = spark.sql("SELECT key, value FROM src WHERE key < 10 ORDER BY key")
df.show(5)
# 5.2 读取mysql数据
url = "jdbc:mysql://localhost:3306/test"
df = spark.read.format("jdbc") \
.option("url", url) \
.option("dbtable", "runoob_tbl") \
.option("user", "root") \
.option("password", "8888") \
.load()\
df.show()
常用的SparkDataFrame API
这里我大概是分成了几部分来看这些APIs，分别是查看DataFrame的APIs、简单处理DataFrame的APIs、DataFrame的列操作APIs、DataFrame的一些思路变换操作APIs、DataFrame的一些统计操作APIs，这样子也有助于我们了解这些API的功能，以后遇见实际问题的时候可以解决。
首先我们这小节全局用到的数据集如下：
from pyspark.sql import functions as F
from pyspark.sql import SparkSession
# SparkSQL的许多功能封装在SparkSession的方法接口中, SparkContext则不行的。
spark = SparkSession.builder \
.appName("sam_SamShare") \
.config("master", "local[4]") \
.enableHiveSupport() \
.getOrCreate()
sc = spark.sparkContext
# 创建一个SparkDataFrame
rdd = sc.parallelize([("Sam", 28, 88, "M"),
("Flora", 28, 90, "F"),
("Run", 1, 60, None),
("Peter", 55, 100, "M"),
("Mei", 54, 95, "F")])
df = rdd.toDF(["name", "age", "score", "sex"])
df.show()
df.printSchema()
# +-----+---+-----+----+
# | name|age|score| sex|
# +-----+---+-----+----+
# | Sam| 28| 88| M|
# |Flora| 28| 90| F|
# | Run| 1| 60|null|
# |Peter| 55| 100| M|
# | Mei| 54| 95| F|
# +-----+---+-----+----+
# root
# |-- name: string (nullable = true)
# |-- age: long (nullable = true)
# |-- score: long (nullable = true)
# |-- sex: string (nullable = true)
1、查看DataFrame的APIs
# DataFrame.collect
# 以列表形式返回行
df.collect()
# [Row(name='Sam', age=28, score=88, sex='M'),
# Row(name='Flora', age=28, score=90, sex='F'),
# Row(name='Run', age=1, score=60, sex=None),
# Row(name='Peter', age=55, score=100, sex='M'),
# Row(name='Mei', age=54, score=95, sex='F')]
# DataFrame.count
df.count()
# 5
# DataFrame.columns
df.columns
# ['name', 'age', 'score', 'sex']
# DataFrame.dtypes
df.dtypes
# [('name', 'string'), ('age', 'bigint'), ('score', 'bigint'), ('sex', 'string')]
# DataFrame.describe
# 返回列的基础统计信息
df.describe(['age']).show()
# +-------+------------------+
# |summary| age|
# +-------+------------------+
# | count| 5|
# | mean| 33.2|
# | stddev|22.353970564532826|
# | min| 1|
# | max| 55|
# +-------+------------------+
df.describe().show()
# +-------+-----+------------------+------------------+----+
# |summary| name| age| score| sex|
# +-------+-----+------------------+------------------+----+
# | count| 5| 5| 5| 4|
# | mean| null| 33.2| 86.6|null|
# | stddev| null|22.353970564532826|15.582040944625966|null|
# | min|Flora| 1| 60| F|
# | max| Sam| 55| 100| M|
# +-------+-----+------------------+------------------+----+
# DataFrame.select
# 选定指定列并按照一定顺序呈现
df.select("sex", "score").show()
# DataFrame.first
# DataFrame.head
# 查看第1条数据
df.first()
# Row(name='Sam', age=28, score=88, sex='M')
df.head(1)
# [Row(name='Sam', age=28, score=88, sex='M')]
# DataFrame.freqItems
# 查看指定列的枚举值
df.freqItems(["age","sex"]).show()
# +---------------+-------------+
# | age_freqItems|sex_freqItems|
# +---------------+-------------+
# |[55, 1, 28, 54]| [M, F,]|
# +---------------+-------------+
# DataFrame.summary
df.summary().show()
# +-------+-----+------------------+------------------+----+
# |summary| name| age| score| sex|
# +-------+-----+------------------+------------------+----+
# | count| 5| 5| 5| 4|
# | mean| null| 33.2| 86.6|null|
# | stddev| null|22.353970564532826|15.582040944625966|null|
# | min|Flora| 1| 60| F|
# | 25%| null| 28| 88|null|
# | 50%| null| 28| 90|null|
# | 75%| null| 54| 95|null|
# | max| Sam| 55| 100| M|
# +-------+-----+------------------+------------------+----+
# DataFrame.sample
# 按照一定规则从df随机抽样数据
df.sample(0.5).show()
# +-----+---+-----+----+
# | name|age|score| sex|
# +-----+---+-----+----+
# | Sam| 28| 88| M|
# | Run| 1| 60|null|
# |Peter| 55| 100| M|
# +-----+---+-----+----+
2、简单处理DataFrame的APIs
# DataFrame.distinct
# 对数据集进行去重
df.distinct().show()
# DataFrame.dropDuplicates
# 对指定列去重
df.dropDuplicates(["sex"]).show()
# +-----+---+-----+----+
# | name|age|score| sex|
# +-----+---+-----+----+
# |Flora| 28| 90| F|
# | Run| 1| 60|null|
# | Sam| 28| 88| M|
# +-----+---+-----+----+
# DataFrame.exceptAll
# DataFrame.subtract
# 根据指定的df对df进行去重
df1 = spark.createDataFrame(
[("a", 1), ("a", 1), ("b", 3), ("c", 4)], ["C1", "C2"])
df2 = spark.createDataFrame([("a", 1), ("b", 3)], ["C1", "C2"])
df3 = df1.exceptAll(df2) # 没有去重的功效
df4 = df1.subtract(df2) # 有去重的奇效
df1.show()
df2.show()
df3.show()
df4.show()
# +---+---+
# | C1| C2|
# +---+---+
# | a| 1|
# | a| 1|
# | b| 3|
# | c| 4|
# +---+---+
# +---+---+
# | C1| C2|
# +---+---+
# | a| 1|
# | b| 3|
# +---+---+
# +---+---+
# | C1| C2|
# +---+---+
# | a| 1|
# | c| 4|
# +---+---+
# +---+---+
# | C1| C2|
# +---+---+
# | c| 4|
# +---+---+
# DataFrame.intersectAll
# 返回两个DataFrame的交集
df1 = spark.createDataFrame(
[("a", 1), ("a", 1), ("b", 3), ("c", 4)], ["C1", "C2"])
df2 = spark.createDataFrame([("a", 1), ("b", 4)], ["C1", "C2"])
df1.intersectAll(df2).show()
# +---+---+
# | C1| C2|
# +---+---+
# | a| 1|
# +---+---+
# DataFrame.drop
# 丢弃指定列
df.drop('age').show()
# DataFrame.withColumn
# 新增列
df1 = df.withColumn("birth_year", 2021 - df.age)
df1.show()
# +-----+---+-----+----+----------+
# | name|age|score| sex|birth_year|
# +-----+---+-----+----+----------+
# | Sam| 28| 88| M| 1993|
# |Flora| 28| 90| F| 1993|
# | Run| 1| 60|null| 2020|
# |Peter| 55| 100| M| 1966|
# | Mei| 54| 95| F| 1967|
# +-----+---+-----+----+----------+
# DataFrame.withColumnRenamed
# 重命名列名
df1 = df.withColumnRenamed("sex", "gender")
df1.show()
# +-----+---+-----+------+
# | name|age|score|gender|
# +-----+---+-----+------+
# | Sam| 28| 88| M|
# |Flora| 28| 90| F|
# | Run| 1| 60| null|
# |Peter| 55| 100| M|
# | Mei| 54| 95| F|
# +-----+---+-----+------+
# DataFrame.dropna
# 丢弃空值，DataFrame.dropna(how='any', thresh=None, subset=None)
df.dropna(how='all', subset=['sex']).show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# | Sam| 28| 88| M|
# |Flora| 28| 90| F|
# |Peter| 55| 100| M|
# | Mei| 54| 95| F|
# +-----+---+-----+---+
# DataFrame.fillna
# 空值填充操作
df1 = spark.createDataFrame(
[("a", None), ("a", 1), (None, 3), ("c", 4)], ["C1", "C2"])
# df2 = df1.na.fill({"C1": "d", "C2": 99})
df2 = df1.fillna({"C1": "d", "C2": 99})
df1.show()
df2.show()
# DataFrame.filter
# 根据条件过滤
df.filter(df.age>50).show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# |Peter| 55| 100| M|
# | Mei| 54| 95| F|
# +-----+---+-----+---+
df.where(df.age==28).show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# | Sam| 28| 88| M|
# |Flora| 28| 90| F|
# +-----+---+-----+---+
df.filter("age<18").show()
# +----+---+-----+----+
# |name|age|score| sex|
# +----+---+-----+----+
# | Run| 1| 60|null|
# +----+---+-----+----+
# DataFrame.join
# 这个不用多解释了，直接上案例来看看具体的语法即可，DataFrame.join(other, on=None, how=None)
df1 = spark.createDataFrame(
[("a", 1), ("d", 1), ("b", 3), ("c", 4)], ["id", "num1"])
df2 = spark.createDataFrame([("a", 1), ("b", 3)], ["id", "num2"])
df1.join(df2, df1.id == df2.id, 'left').select(df1.id.alias("df1_id"),
df1.num1.alias("df1_num"),
df2.num2.alias("df2_num")
).sort(["df1_id"], ascending=False)\
.show()
# DataFrame.agg(*exprs)
# 聚合数据，可以写多个聚合方法，如果不写groupBy的话就是对整个DF进行聚合
# DataFrame.alias
# 设置列或者DataFrame别名
# DataFrame.groupBy
# 根据某几列进行聚合，如有多列用列表写在一起，如 df.groupBy(["sex", "age"])
df.groupBy("sex").agg(F.min(df.age).alias("最小年龄"),
F.expr("avg(age)").alias("平均年龄"),
F.expr("collect_list(name)").alias("姓名集合")
).show()
# +----+--------+--------+------------+
# | sex|最小年龄|平均年龄| 姓名集合|
# +----+--------+--------+------------+
# | F| 28| 41.0|[Flora, Mei]|
# |null| 1| 1.0| [Run]|
# | M| 28| 41.5|[Sam, Peter]|
# +----+--------+--------+------------+
# DataFrame.foreach
# 对每一行进行函数方法的应用
def f(person):
print(person.name)
df.foreach(f)
# Peter
# Run
# Sam
# Flora
# Mei
# DataFrame.replace
# 修改df里的某些值
df1 = df.na.replace({"M": "Male", "F": "Female"})
df1.show()
# DataFrame.union
# 相当于SQL里的union all操作
df1 = spark.createDataFrame(
[("a", 1), ("d", 1), ("b", 3), ("c", 4)], ["id", "num"])
df2 = spark.createDataFrame([("a", 1), ("b", 3)], ["id", "num"])
df1.union(df2).show()
df1.unionAll(df2).show()
# 这里union没有去重，不知道为啥，有知道的朋友麻烦解释下，谢谢了。
# +---+---+
# | id|num|
# +---+---+
# | a| 1|
# | d| 1|
# | b| 3|
# | c| 4|
# | a| 1|
# | b| 3|
# +---+---+
# DataFrame.unionByName
# 根据列名来进行合并数据集
df1 = spark.createDataFrame([[1, 2, 3]], ["col0", "col1", "col2"])
df2 = spark.createDataFrame([[4, 5, 6]], ["col1", "col2", "col0"])
df1.unionByName(df2).show()
# +----+----+----+
# |col0|col1|col2|
# +----+----+----+
# | 1| 2| 3|
# | 6| 4| 5|
# +----+----+----+
3、DataFrame的列操作APIs
这里主要针对的是列进行操作，比如说重命名、排序、空值判断、类型判断等，这里就不展开写demo了，看看语法应该大家都懂了。
Column.alias(*alias, **kwargs) # 重命名列名
Column.asc() # 按照列进行升序排序
Column.desc() # 按照列进行降序排序
Column.astype(dataType) # 类型转换
Column.cast(dataType) # 强制转换类型
Column.between(lowerBound, upperBound) # 返回布尔值，是否在指定区间范围内
Column.contains(other) # 是否包含某个关键词
Column.endswith(other) # 以什么结束的值，如 df.filter(df.name.endswith('ice')).collect()
Column.isNotNull() # 筛选非空的行
Column.isNull()
Column.isin(*cols) # 返回包含某些值的行 df[df.name.isin("Bob", "Mike")].collect()
Column.like(other) # 返回含有关键词的行
Column.when(condition, value) # 给True的赋值
Column.otherwise(value) # 与when搭配使用，df.select(df.name, F.when(df.age > 3, 1).otherwise(0)).show()
Column.rlike(other) # 可以使用正则的匹配 df.filter(df.name.rlike('ice$')).collect()
Column.startswith(other) # df.filter(df.name.startswith('Al')).collect()
Column.substr(startPos, length) # df.select(df.name.substr(1, 3).alias("col")).collect()
4、DataFrame的一些思路变换操作APIs
# DataFrame.createOrReplaceGlobalTempView
# DataFrame.dropGlobalTempView
# 创建全局的试图，注册后可以使用sql语句来进行操作，生命周期取决于Spark application本身
df.createOrReplaceGlobalTempView("people")
spark.sql("select * from global_temp.people where sex = 'M' ").show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# | Sam| 28| 88| M|
# |Peter| 55| 100| M|
# +-----+---+-----+---+
# DataFrame.createOrReplaceTempView
# DataFrame.dropTempView
# 创建本地临时试图，生命周期取决于用来创建此数据集的SparkSession
df.createOrReplaceTempView("tmp_people")
spark.sql("select * from tmp_people where sex = 'F' ").show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# |Flora| 28| 90| F|
# | Mei| 54| 95| F|
# +-----+---+-----+---+
# DataFrame.cache\DataFrame.persist
# 可以把一些数据放入缓存中，default storage level (MEMORY_AND_DISK).
df.cache()
df.persist()
df.unpersist()
# DataFrame.crossJoin
# 返回两个DataFrame的笛卡尔积关联的DataFrame
df1 = df.select("name", "sex")
df2 = df.select("name", "sex")
df3 = df1.crossJoin(df2)
print("表1的记录数", df1.count())
print("表2的记录数", df2.count())
print("笛卡尔积后的记录数", df3.count())
# 表1的记录数 5
# 表2的记录数 5
# 笛卡尔积后的记录数 25
# DataFrame.toPandas
# 把SparkDataFrame转为 Pandas的DataFrame
df.toPandas()
# DataFrame.rdd
# 把SparkDataFrame转为rdd，这样子可以用rdd的语法来操作数据
df.rdd
5、DataFrame的一些统计操作APIs
# DataFrame.cov
# 计算指定两列的样本协方差
df.cov("age", "score")
# 324.59999999999997
# DataFrame.corr
# 计算指定两列的相关系数，DataFrame.corr(col1, col2, method=None)，目前method只支持Pearson相关系数
df.corr("age", "score", method="pearson")
# 0.9319004030498815
# DataFrame.cube
# 创建多维度聚合的结果，通常用于分析数据，比如我们指定两个列进行聚合，比如name和age，那么这个函数返回的聚合结果会
# groupby("name", "age")
# groupby("name")
# groupby("age")
# groupby(all)
# 四个聚合结果的union all 的结果
df1 = df.filter(df.name != "Run")
print(df1.show())
df1.cube("name", "sex").count().show()
# +-----+---+-----+---+
# | name|age|score|sex|
# +-----+---+-----+---+
# | Sam| 28| 88| M|
# |Flora| 28| 90| F|
# |Peter| 55| 100| M|
# | Mei| 54| 95| F|
# +-----+---+-----+---+
# cube 聚合之后的结果
# +-----+----+-----+
# | name| sex|count|
# +-----+----+-----+
# | null| F| 2|
# | null|null| 4|
# |Flora|null| 1|
# |Peter|null| 1|
# | null| M| 2|
# |Peter| M| 1|
# | Sam| M| 1|
# | Sam|null| 1|
# | Mei| F| 1|
# | Mei|null| 1|
# |Flora| F| 1|
# +-----+----+-----+
嘻嘻，恭喜你读完啦，奖励你一首歌，一起加油。
评论