近期致力于总结科研或者工作中用到的主要技术栈，从技术原理到常用语法，这次查缺补漏当作我的小百科。主要技术包括：

• ✅数据库常用：MySQL, Hive SQL, Spark SQL
• ✅大数据处理常用：Pyspark, Pandas
• ⚪ 图像处理常用：OpenCV, matplotlib
• ⚪ 机器学习常用：SciPy, Sklearn
• ⚪ 深度学习常用：Pytorch, numpy
• ⚪ 常用数据结构语法糖：itertools, collections
• ⚪ 常用命令: Shell, Git, Vim

以下整理错误或者缺少的部分欢迎指正！！！

大数据处理常用：Pyspark, Pandas

性能对比

常用语法对比

# 头文件
from pyspark.sql import SparkSession
import pyspark.sql.functions as F
from pyspark.sql.types import StructType, StructField, LongType, StringType, ArrayType  # 或者直接导入*
import pandas as pd


# 创建SparkSession对象
spark = SparkSession.builder \
    .appName("username") \
    .getOrCreate()


# 创建空表
schema = StructType([
                StructField('id', LongType()),
                StructField('type', StringType()),
            ])  # spark需要指定列名和类型
spark_df = spark.createDataFrame(spark.sparkContext.emptyRDD(), schema=schema)
pandas_df = pd.DataFrame(columns=['id', 'type'], index=[0, 1, 2])


# 根据现有数据创建
data = [(1, "Alice", 2000), (2, "Bob", 2001), (3, "Charlie", 2002)]
schema = StructType([
    StructField("id", IntegerType(), True),
    StructField("name", StringType(), True),
    StructField("birth_year", IntegerType(), True)
])
spark_df = spark.createDataFrame(data, ["id", "name", "birth_year"])
spark_df = spark.createDataFrame(data, schema)
pandas_df = pd.DataFrame(data=data, columns=["id", "name", "birth_year"])


# 读取csv文件
spark_df = spark.read.csv("data.csv", header=True, inferSchema=True)
pandas_df = pd.read_csv("data.csv", sep="\t")  # read_excel
# 保存数据到csv
spark_df.write.csv('data.csv', header=True)
pandas_df.to_csv("data.csv", index=False)

# 读取hive表数据
spark_df = spark.sql('select * from tab')
# 保存数据到hive表
spark_df.write.mode('overwrite').saveAsTable('db_name.tab_name')


# 相互转换
spark_df = SQLContext.createDataFrame(pandas_df)
pandas_df = spark_df.toPandas()


# 转换数据类型
spark_df = spark_df.withColumn("A", col("age").cast(StringType))
pandas_df["A"] = pandas_df['A'].astype("int")


# 重置索引
spark_df = spark_df.withColumn("id", monotonically_increasing_id())  # 生成一个增长的id列
pandas_df.reset_index()


# 切片
pandas_df['a':'c']  # a-c三行
pandas_df.iloc[1:3, 0:2]  # 1-2行，0-1列。左闭右开
pandas_df.iloc[[0, 2], [1, 2]] # 第0，2行第0，2列
pandas_df.loc['a':'c', ['A', 'B']] # 第a-c行A，B列


# 选择列
spark_df.select('A', 'B')
pandas_df[['A', 'B']]

# 删除列
spark_df.drop('A', 'B')
pandas_df.drop(['A', 'B'], axis=1, inplace=True)  # inplace表示是否创建新对象

# 新增列，设置列值
spark_df = spark_df.withColumn('name', F.lit(0))
pandas_df['name'] = 0

# 修改列值
spark_df.withColumn('name', 1)
pandas_df['name'] = 1
# 使用函数修改列值
spark_df = spark_df.withColumn('code', F.when(F.isnull(spark_df.code), 0).otherwise(spark_df.code))

# 修改列名
spark_df.withColumnRenamed('old_name', 'new_name')
pandas_df.rename(columns={'old_name1': 'new_name1', 'old_name1': 'new_name2'}, inplace=True)


# 显示数据
spark_df.limit(10) # 前10行
spark_df.show/take(10)  # collect()返回全部数据
spark_df/pandas_df.first/head/tail(10)


# 表格遍历
saprk_df.collect()[:10]
spark_df.foreach(lambda row: print(row['c1'], row['c2']))
for i, row in pandas_df.iterrows():
    print(row["c1"], row["c2"])


# 排序
spark/pandas_df.sort()  # 按列值排序
pandas_df.sort_index()  # 按轴排序
pandas_df.sort_values(by=["A", "B"], axis=0, ascending=[True, False], inplace=True)  # 指定列升序/降序排序


# 过滤
spark_df.filter(df['col_name'] > 1)     # spark_df.where(df['col_name'] > 1)
pandas_df[pandas_df['col_name'] > 1]
pandas_df_new = pandas_df[pandas_df["code"].apply(lambda x: len(x) == 11)]


# 去重
spark_df.select('col_name').distinct()
spark_df_filter = spark_df.drop_duplicates(["col_name"])
pandas_df.drop_duplicates(["col_name"], keep='first', inplace=True)

# 缺失数据处理
spark_df.na.fill()
spark_df.na.drop(subset=['A', "B"])  # 同dropna
pandas_df.fillna()
pandas_df.dropna(subset=['A', "B"], how="any", inplace=True)

# 空值过滤 filter=choose
spark_df.filter(~(F.isnull(spark_df.d)))
spark_df.filter(~(spark_df['A'].isNull() | spark_df['B'].isNull()))   # 选出列值不为空的行  isnan()=isNull()<->isNOtnan()
pandas_df[pandas_df['A'].isna()]  # 选出列值为空的行
pandas_df[pandas_df['A'].notna()] # 选出列值不为空的行


# 统计
spark/pandas_df.count()  # spark返回总行数，pandas返回列非空总数
spark/pandas_df.describe() # 描述列的count, mean, min, max...

# 计算某一列均值
average_value = spark_df.select("col_name").agg({"col_name": "avg"}).collect()[0][0]
average_value = pandas_df["col_name"].mean()


# 表合并
# 按行合并，相当于追加
spark_df = spark_df.unionAll(spark_df1)
pandas_df = pd.concat([df_up, df_down], axis=0)
# 按列合并
spark_df = spark_df.join(df1, df1.id==spark_df.id, 'inner').drop(df1.id)  # df1.id==spark_df.id也可写成['id]（当且仅当列名相同）
pd.merge(df_left, df_right, left_on="a", right_on="b", how="left|right|inner|outer")  


# 聚合函数
spark_df_collect = spark_df.groupBy('number').agg(
    F.collect_set('province').alias('set_province'),
    F.first('city').alias('set_city'),
    F.collect_list('district').alias('set_district'),
    F.max('report_user').alias('set_report_user'),
    F.min('first_type').alias('set_first_type'))
# 分组聚合
spark_df.groupBy('A').agg(F.avg('B'), F.min('B'))
spark/pandas_df.groupby('A').avg('B')

# 根据函数分组聚合
def func(x):
    return pd.DataFrame({
        "A": x["A"].tolist()[0],
        "B": sum(x["B"])}, index=[0])
pandas_df_result = pandas_df.groupby(["A"]).apply(func)


# spark udf函数和pandas apply函数
def func1(a, b):
    return a + b
spark_df.withColumn("col_name", F.udf(func1, IntegerType())(spark_df.a, spark_df.b))  # spark_df['a']或F.col("a")))
def func2(x,y):
    return 1 if x > np.mean(y) else 0
pandas_df['A'].apply(func2, args=(pandas_df['B'],))
pandas_df['C'] = pandas_df.apply(lambda x: 1 if x['A'] > (x['B']*0.5) else 0, axis=1)


# spark创建临时表
spark_df.createOrReplaceTempView('tmp_table')  # 用sql API
res1 = spark.sql('select * from tmp_table')
spark_df.registerTempTable('tmp_table') # 用dataframe API
res2 = spark.table('tmp_table') 

其他常用设置

class SparkUtils:
    def __init__(self):
        self.spark = None

    def get_spark(self):
        if self.spark is None:
            self.spark = SparkSession.builder.appName("username") \
                .enableHiveSupport().config("spark.sql.shuffle.partitions", "500") \
                .config("spark.sql.broadcastTimeout", "3600") \
                .config("spark.driver.memory", "200g") \
                .config("spark.executor.memory", "40g") \
                .config("spark.yarn.appMasterEnv.yarn.nodemanager.container-executor.class", "DockerLinuxContainer") \
                .config("spark.executorEnv.yarn.nodemanager.container-executor.class", "DockerLinuxContainer") \
                .config("spark.yarn.appMasterEnv.yarn.nodemanager.docker-container-executor.image-name",
                        "bdp-docker.jd.com:5000/wise_mart_bag:latest") \
                .config("spark.executorEnv.yarn.nodemanager.docker-container-executor.image-name",
                        "bdp-docker.jd.com:5000/wise_mart_bag:latest") \
                .getOrCreate()
        self.spark.sql('SET hive.exec.dynamic.partition=true')
        self.spark.sql('SET hive.exec.dynamic.partition.mode=nonstrict')
        return self.spark

spark = SparkUtils()

# 生成dataframe
spark_data = spark.sql("""
    select 
      id, 
      username
    from 
      tab1
    where 
      status in (1, 2, 3)
      and dt = '{}'
  """.format(date))

# pandas常用显示设置
pd.set_option('display.max_rows', 100)
pd.set_option('display.max_columns', None)
pd.set_option('display.width',1000)
pd.set_option('display.max_colwidth',1000)