pyspark sparksession_pyspark入门 | spark-submit 提交pyspark任务

https://github.com/hxjcarrie/pyspark_study

以LogisticRegression为例

• 输入数据样例（第一列为label，后面为feature）

• lrDemo.py（基于RDD的mllib）

#!coding=utf8
'''
author: huangxiaojuan
'''
import sys
reload(sys)
sys.setdefaultencoding('utf8')
from pyspark.sql import SparkSession,Row
from pyspark.sql.types import *
from time import *
import numpy
import os

from pyspark.mllib.linalg import Vectors
from pyspark.mllib.regression import LabeledPoint
from pyspark.mllib.classification import LogisticRegressionWithLBFGS, LogisticRegressionModel
#os.environ['PYSPARK_PYTHON'] = './python_env/py27/bin/python2' 

def parseFloat(x):
    try:
        rx = float(x)
    except:
        rx = 0.0
    return rx

def parse(line, ifUid=False):
    l = line.split('t')
    uid = l[0]
    label = parseFloat(l[1])
    features = map(lambda x: parseFloat(x), l[2:])
    if ifUid:
        return (uid, LabeledPoint(label, features))
    else:
        return LabeledPoint(label, features)

def main():
    #spark = SparkSession.builder.master("yarn").appName("spark_demo").getOrCreate()
    spark = SparkSession.builder.getOrCreate()
    print "Session created!"
    sc = spark.sparkContext
    #打印追踪任务url
    print "The url to track the job: http://namenode-01:8088/proxy/" + sc.applicationId

    sampleHDFS_train = sys.argv[1]
    sampleHDFS_test = sys.argv[2]
    outputHDFS = sys.argv[3]

    sampleRDD = sc.textFile(sampleHDFS_train).map(parse)
    predictRDD = sc.textFile(sampleHDFS_test).map(lambda x: parse(x, True))

    # 训练
    model = LogisticRegressionWithLBFGS.train(sampleRDD)
    model.clearThreshold() #删除默认阈值（否则后面直接输出0、1）

    # 预测，保存结果
    labelsAndPreds = predictRDD.map(lambda p: (p[0], p[1].label, model.predict(p[1].features)))
    labelsAndPreds.map(lambda p: 't'.join(map(str, p))).saveAsTextFile(outputHDFS + "/target/output")

    # 评估不同阈值下的准确率、召回率
    labelsAndPreds_label_1 = labelsAndPreds.filter(lambda lp: int(lp[1]) == 1)
    labelsAndPreds_label_0 = labelsAndPreds.filter(lambda lp: int(lp[1]) == 0)
    t_cnt = labelsAndPreds_label_1.count()
    f_cnt = labelsAndPreds_label_0.count()
    print "threttpttntfptfntaccuracytrecall"
    for thre in [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95]:
        tp = labelsAndPreds_label_1.filter(lambda lp: lp[2] > thre).count()
        tn = t_cnt - tp
        fp = labelsAndPreds_label_0.filter(lambda lp: lp[2] > thre).count()
        fn = f_cnt - fp
        print("%.1ft%dt%dt%dt%dt%.4ft%.4f"%(thre, tp, tn, fp, fn, float(tp)/(tp+fp), float(tp)/(t_cnt)))

    # 保存模型、加载模型
    model.save(sc, outputHDFS + "/target/tmp/pythonLogisticRegressionWithLBFGSModel")
    sameModel = LogisticRegressionModel.load(sc, outputHDFS + "/target/tmp/pythonLogisticRegressionWithLBFGSModel")

    print "output:", outputHDFS


if __name__ == '__main__':
    main()

• lrDemo_df.py（基于DataFrame的ml）

#!coding=utf8
'''
author: huangxiaojuan
'''
import sys
reload(sys)
sys.setdefaultencoding('utf8')
from pyspark.sql import SparkSession,Row
from pyspark.sql.types import *
from time import *
import numpy
import os

from pyspark.ml.linalg import Vectors
from pyspark.ml.classification import LogisticRegression
from pyspark.ml.feature import StringIndexer, VectorAssembler
from pyspark.ml import Pipeline
from pyspark.sql.functions import udf, col

def getFeatureName():
    featureLst = ['feature1', 'feature2', 'feature3', 'feature4', 'feature5', 'feature6', 'feature7', 'feature8', 'feature9']
    colLst = ['uid', 'label'] + featureLst
    return featureLst, colLst

def parseFloat(x):
    try:
        rx = float(x)
    except:
        rx = 0.0
    return rx


def getDict(dictDataLst, colLst):
    dictData = {}
    for i in range(len(colLst)):
        dictData[colLst[i]] = parseFloat(dictDataLst[i])
        if colLst[i] == "label":
            dictData["weight"] = 1.0 if dictDataLst[i] == '1' else 1.0
    return dictData

def to_array(col):
    def to_array_(v):
        return v.toArray().tolist()
    return udf(to_array_, ArrayType(DoubleType())).asNondeterministic()(col)


def main():
    #spark = SparkSession.builder.master("yarn").appName("spark_demo").getOrCreate()
    spark = SparkSession.builder.getOrCreate()
    print "Session created!"
    sc = spark.sparkContext
    print "The url to track the job: http://bx-namenode-02:8088/proxy/" + sc.applicationId

    sampleHDFS_train = sys.argv[1]
    sampleHDFS_test = sys.argv[2]
    outputHDFS = sys.argv[3]

    featureLst, colLst = getFeatureName()

    #读取hdfs上数据，将RDD转为DataFrame
    #训练数据
    rdd_train = sc.textFile(sampleHDFS_train)
    rowRDD_train = rdd_train.map(lambda x: getDict(x.split('t'), colLst))
    trainDF = spark.createDataFrame(rowRDD_train)
    #测试数据
    rdd_test = sc.textFile(sampleHDFS_test)
    rowRDD_test = rdd_test.map(lambda x: getDict(x.split('t'), colLst))
    testDF = spark.createDataFrame(rowRDD_test)

    #用于训练的特征featureLst
    vectorAssembler = VectorAssembler().setInputCols(featureLst).setOutputCol("features")

    #### 训练 ####
    print "step 1"
    lr = LogisticRegression(regParam=0.01, maxIter=100, weightCol="weight")  # regParam 正则项参数

    pipeline = Pipeline(stages=[vectorAssembler, lr])
    model = pipeline.fit(trainDF)
    #打印参数
    print "n-------------------------------------------------------------------------"
    print "LogisticRegression parameters:n" + lr.explainParams() + "n"
    print "-------------------------------------------------------------------------n"

    #### 预测, 保存结果 ####
    print "step 2"
    labelsAndPreds = model.transform(testDF).withColumn("probability_xj", to_array(col("probability"))[1])
                                            .select("uid", "label", "prediction", "probability_xj")
    labelsAndPreds.show()
    labelsAndPreds.write.mode("overwrite").options(header="true").csv(outputHDFS + "/target/output")


    #### 评估不同阈值下的准确率、召回率
    print "step 3"
    labelsAndPreds_label_1 = labelsAndPreds.where(labelsAndPreds.label == 1)
    labelsAndPreds_label_0 = labelsAndPreds.where(labelsAndPreds.label == 0)
    labelsAndPreds_label_1.show(3)
    labelsAndPreds_label_0.show(3)
    t_cnt = labelsAndPreds_label_1.count()
    f_cnt = labelsAndPreds_label_0.count()
    print "threttpttntfptfntaccuracytrecall"
    for thre in [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95]:
        tp = labelsAndPreds_label_1.where(labelsAndPreds_label_1.probability_xj > thre).count()
        tn = t_cnt - tp
        fp = labelsAndPreds_label_0.where(labelsAndPreds_label_0.probability_xj > thre).count()
        fn = f_cnt - fp
        print("%.1ft%dt%dt%dt%dt%.4ft%.4f"%(thre, tp, tn, fp, fn, float(tp)/(tp+fp), float(tp)/(t_cnt)))

    # 保存模型
    model.write().overwrite().save(outputHDFS + "/target/model/lrModel")
    #加载模型
    #model.load(outputHDFS + "/target/model/lrModel")

    print "output:", outputHDFS


if __name__ == '__main__':
    main()

• 日志打印模型效果：

spark-submit_lr.sh 提交任务到集群

ModelType=lrDemo
ModelType=lrDemo_df
#ModelType=xgbDemo

CUR_PATH=$(cd "$(dirname "$0")";pwd)
echo $CUR_PATH

SPARK_PATH=/user/spark/spark
YARN_QUEUE=

DEPLOY_MODE=cluster
DEPLOY_MODE=client

input_path_train=hdfs://
input_path_test=hdfs://
output_path=hdfs://user/huangxiaojuan/program/sparkDemo/${ModelType}

hadoop fs -rmr $output_path


${SPARK_PATH}/bin/spark-submit 
 --master yarn 
 --name "spark_demo_lr" 
 --queue ${YARN_QUEUE} 
 --deploy-mode ${DEPLOY_MODE} 
 --driver-memory 6g 
 --driver-cores 4 
 --executor-memory 12g 
 --executor-cores 15 
 --num-executors 10 
 --archives ./source/py27.zip#python_env 
 --conf spark.default.parallelism=150 
 --conf spark.executor.memoryOverhead=4g 
 --conf spark.driver.memoryOverhead=2g 
 --conf spark.yarn.maxAppAttempts=3 
 --conf spark.yarn.submit.waitAppCompletion=true 
 --conf spark.pyspark.driver.python=./source/py27/bin/python2 
 --conf spark.yarn.appMasterEnv.PYSPARK_PYTHON=./python_env/py27/bin/python2 
 --conf spark.pyspark.python=./python_env/py27/bin/python2 
 ./${ModelType}.py $input_path_train $input_path_test $output_path

• nohup sh -x spark-submit_lr.sh > spark-submit_lr.log 2>&1 &
• kill任务: yarn application -kill application_xxxxxxxxx_xxxxx

上传python包

• 需要保证driver和executor上的python版本一致
• 若executor上的python不满足要求，可通过如下参数上传打包好的python到executor上

#上传python包到executor
--archives ./source/py27.zip  
#指定executor上python路径
--conf spark.yarn.appMasterEnv.PYSPARK_PYTHON=./python_env/py27/bin/python2 
--conf spark.pyspark.python=./python_env/py27/bin/python2 
#指定driver上python路径
--conf spark.pyspark.driver.python=./source/py27/bin/python2 

#或者先上传至hdfs
--conf spark.yarn.dist.archives=hdfs://user/huangxiaojuan/py27.zip#python_env

参考资料

初学者可直接参考：

Submitting Applications

https://spark.apache.org/docs/latest/quick-start.html

spark.mllib参考：

MLlib: Main Guide​spark.apache.org

spark.ml参考：

https://spark.apache.org/docs/latest/api/python/pyspark.ml.html​spark.apache.org

两者区别：

• spark.sql参考：

Getting Started - Spark 2.4.4 Documentation​spark.apache.org

• 使用dataFrame进行表的关联，例子：

https://github.com/hxjcarrie/pyspark_study/blob/master/df.py

• 使用spark.sql进行表的关联，例子：

https://github.com/hxjcarrie/pyspark_study/blob/master/df_sql.py

• 若要写入hive表，可参考: https://blog.csdn.net/lulynn/article/details/51543567

• xgboost 参考：

叶一一：pyspark 数据分析-xgboost​zhuanlan.zhihu.com

• scala参考：

叶一一：scala入门 | spark-submit提交scala任务​zhuanlan.zhihu.com

-------------------------------------------------------------------------------------------

若需要RDD嵌套RDD，或要使用的算法只有python自己的sklearn里有，可以考虑对样本分组做分布式的（但模型训练是单机的，所以这种方法的前提是：分完组的数据量在单机训练占用的内存不多）

Say you find yourself in the peculiar situation where you need to train a whole bunch ofscikit-learnmodels over different groups from a large amount of data. And say you want to leverage Spark to distribute the process to do it all in a scalable fashion

Using PySpark and Pandas UDFs to train many scikit-learn models distributedly​patrickroos.org