最近要在 Spark job 中通过 Spark SQL 的方式读取 Elasticsearch 数据,踩了一些坑,总结于此。
环境说明-
Spark job 的编写语言为 Scala,scala-library 的版本为 2.11.8。
-
Spark 相关依赖包的版本为 2.3.2,如 spark-core、spark-sql。
-
Elasticsearch 数据
schema
{ "settings": { "number_of_replicas": 1 }, "mappings": { "label": { "properties": { "docId": { "type": "keyword" }, "labels": { "type": "nested", "properties": { "id": { "type": "long" }, "label": { "type": "keyword" } } }, "itemId": { "type": "long" } } } } }
sample data
{ "took" : 141, "timed_out" : false, "_shards" : { "total" : 5, "successful" : 5, "skipped" : 0, "failed" : 0 }, "hits" : { "total" : 17370929, "max_score" : 1.0, "hits" : [ { "_index" : "aen-label-v1", "_type" : "label", "_id" : "123_ITEM", "_score" : 1.0, "_source" : { "docId" : "123_ITEM", "labels" : [ { "id" : 7378, "label" : "1kg" } ], "itemId" : 123 } }, { "_index" : "aen-label-v1", "_type" : "label", "_id" : "456_ITEM", "_score" : 1.0, "_source" : { "docId" : "456_ITEM", "labels" : [ { "id" : 7378, "label" : "2kg" } ], "itemId" : 456 } } ] } }
既然要用 Spark SQL,当然少不了其对应的依赖,
dependencies {
implementation 'org.apache.spark:spark-core_2.11:2.3.2'
implementation 'org.apache.spark:spark-sql_2.11:2.3.2'
}
对于 ES 的相关库,如同 官网 所说,要在 Spark 中访问 ES,需要将 elasticsearch-hadoop
依赖包加入到 Spark job 运行的类路径中,具体而言就是添加到 Spark job 工程的依赖中,公司的 nexus 中当前最新的版本为 7.15.0,且目前我们是使用 gradle 管理依赖,故添加依赖的代码如下,
dependencies {
implementation 'org.elasticsearch:elasticsearch-hadoop:7.15.0'
}
本地测试
对于 Spark,基于资源管理器的不同,可以在两种模式下运行:本地模式和集群模式,可通过 --master
参数来指定资源管理器的方式。本地模式时,不依赖额外的 Spark 集群,Spark 将在同一台机器上运行所有内容,非常方便用于本地测试,对于 Spark SQL,只需要在创建 SparkSession 时采用 local 的模式即可,
class MyUtils extends Serializable {
def esHost() = s"es.sherlockyb.club"
// local mode
def getLocalSparkSession: SparkSession = SparkSession.builder()
.master("local")
.getOrCreate()
// cluster mode
def getSparkSession: SparkSession = SparkSession.builder()
.enableHiveSupport()
.config("spark.sql.broadcastTimeout", "3600")
.getOrCreate()
}
测试代码
object LocalTest extends LazyLogging {
def main(args: Array[String]): Unit = {
new LocalTest().run()
}
}
class LocalTest {
def run(): Unit = {
val myUtils = new MyUtils
val spark = myUtils.getLocalSparkSession
import spark.implicits._
var start = System.currentTimeMillis()
val attributeId = 7378L
val labelNames = Array("aen-label-retail", "aen-label-seller")
spark.read
.format("es")
.option("es.nodes", myUtils.esHost())
.option("es.port", "9200")
.option("es.nodes.wan.only", value = true)
.option("es.resource", Joiner.on(",").join(java.util.Arrays.asList(labelNames:_*)) + "/label")
.option("es.scroll.size", 2000)
.load()
.createOrReplaceTempView("temp_labels")
val sqlDf = spark.sql("select itemId, labels from temp_labels where itemId in (123, 456)")
val newDf = sqlDf
.map(row => {
val labels = row.getAs[Seq[Row]]("labels")
val labelValue = labels.find(p => p.getAs[Long]("id") == attributeId).map(p => p.getAs[String]("label"))
(row.getAs[Long]("itemId"), attributeId, labelValue.orNull)
})
.withColumn("final_result", lit("PASS"))
.toDF("itemId", "attributeId", "label", "final_result")
val finalDf = newDf.toDF("itemId", "attributeId", "label", "result")
finalDf.printSchema()
finalDf.show()
var emptyDf = newDf
.filter(col("label").isNotNull)
.toDF("itemId", "attributeId", "label", "result")
emptyDf = emptyDf.union(finalDf)
emptyDf.printSchema()
emptyDf.show()
emptyDf.filter(col("itemId") === 6238081929L and col("label").notEqual(col("result")))
.show()
val attributeTypeIds = Array.fill(3)(100)
val attributeTypeIdsStr = Joiner.on(",").join(java.util.Arrays.asList(attributeTypeIds:_*))
println(attributeTypeIdsStr)
import scala.collection.JavaConverters._
emptyDf = emptyDf.filter(!col("itemId").isin(trainItemIds.asScala.map(Long2long).toList:_*))
emptyDf.show(false)
}
}
知识点
Spark SQL Data Sources
Spark SQL 通过 DataFrameReader
类支持读取各种类型的数据源,比如 Parquet、ORC、JSON、CSV 等格式的文件,Hive table,以及其他 database。而 Elasticsearch 只不过是众多数据源中的一种,DataFrameReader
通过 format(...)
指定数据源格式,通过 option(...)
定制对应数据源下的配置,最后通过 load()
加载生成 DataFrame
,也就是 Dataset[Row]
的类型别名。有了 DataFrame
,就可以创建一个临时表,然后就能以 SQL 的方式读取数据。
在 Spark 1.5 以前,Elasticsearch 在 format(...)
中对应的 source 名需要是全包名 org.elasticsearch.spark.sql
,而在 Spark 1.5 以及之后的版本,source 名称简化为 es
。
- df.printSchema(),打印 schema
- df.show(),查看数据列表,默认是 truncate 前 20 条,传 false 时列出全部数据。
- df.createOrReplaceTempView("view_name"),构建临时表视图,方便后续 SQL 操作。
- df.withColumn(),添加新列或替换现有列。
- df.withColumn("final_result", lit("PASS")) ,通过
lit
添加常量列。
- df.withColumn("final_result", lit("PASS")) ,通过
- df.filter(col("label").isNotNull),用指定的条件过滤行。
- df.dropDuplicates("itemId","attributeId"),按指定列对行去重,返回新的数据集。
- df.union(otherDf),将两个 DataFrame 的记录合并且不去重,相当于 union all。
- df.toDF("itemId", "attributeId", "label", "final_result"),为 df 各列指定一个有意义的名称。
- scala.Long -> long
- Array[T] -> T[]
import scala.collection.JavaConverters._
newDf = df.filter(!col("itemId").isin(trainItemIds.asScala.map(Long2long).toList:_*))
Scala 中的 : _*
:_*
是 type ascription 的一个特例,它会告诉编译器将序列类型的单个参数视为变参数序列,即 varargs。应用例子,
val indices = Array("aen-label", "aen-label-seller")
Joiner.on(",").join(java.util.Arrays.asList(indices:_*))
踩的坑
es.nodes.wan.only
该配置项表示连接器是否用于 WAN 上的云或受限环境如 AWS 中的 Elasticsearch 实例,默认为 false,而公司的 Elasticsearch 集群是在 AWS 上的,endpoint 只能在内网访问,因而刚开始测试时,遇到如下报错,
Exception in thread "main" org.elasticsearch.hadoop.EsHadoopIllegalArgumentException: No data nodes with HTTP-enabled available
at org.elasticsearch.hadoop.rest.InitializationUtils.filterNonDataNodesIfNeeded(InitializationUtils.java:159)
at org.elasticsearch.hadoop.rest.RestService.findPartitions(RestService.java:223)
at org.elasticsearch.spark.rdd.AbstractEsRDD.esPartitions$lzycompute(AbstractEsRDD.scala:73)
at org.elasticsearch.spark.rdd.AbstractEsRDD.esPartitions(AbstractEsRDD.scala:72)
at org.elasticsearch.spark.rdd.AbstractEsRDD.getPartitions(AbstractEsRDD.scala:44)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:253)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:251)
at scala.Option.getOrElse(Option.scala:121)
at org.apache.spark.rdd.RDD.partitions(RDD.scala:251)
at org.apache.spark.rdd.MapPartitionsRDD.getPartitions(MapPartitionsRDD.scala:46)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:253)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:251)
at scala.Option.getOrElse(Option.scala:121)
at org.apache.spark.rdd.RDD.partitions(RDD.scala:251)
at org.apache.spark.rdd.MapPartitionsRDD.getPartitions(MapPartitionsRDD.scala:46)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:253)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:251)
at scala.Option.getOrElse(Option.scala:121)
at org.apache.spark.rdd.RDD.partitions(RDD.scala:251)
at org.apache.spark.rdd.MapPartitionsRDD.getPartitions(MapPartitionsRDD.scala:46)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:253)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:251)
at scala.Option.getOrElse(Option.scala:121)
at org.apache.spark.rdd.RDD.partitions(RDD.scala:251)
at org.apache.spark.rdd.MapPartitionsRDD.getPartitions(MapPartitionsRDD.scala:46)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:253)
at org.apache.spark.rdd.RDD$$anonfun$partitions$2.apply(RDD.scala:251)
at scala.Option.getOrElse(Option.scala:121)
at org.apache.spark.rdd.RDD.partitions(RDD.scala:251)
at org.apache.spark.sql.execution.SparkPlan.executeTake(SparkPlan.scala:340)
at org.apache.spark.sql.execution.CollectLimitExec.executeCollect(limit.scala:38)
at org.apache.spark.sql.Dataset.org$apache$spark$sql$Dataset$$collectFromPlan(Dataset.scala:3278)
at org.apache.spark.sql.Dataset$$anonfun$head$1.apply(Dataset.scala:2489)
at org.apache.spark.sql.Dataset$$anonfun$head$1.apply(Dataset.scala:2489)
at org.apache.spark.sql.Dataset$$anonfun$52.apply(Dataset.scala:3259)
at org.apache.spark.sql.execution.SQLExecution$.withNewExecutionId(SQLExecution.scala:77)
at org.apache.spark.sql.Dataset.withAction(Dataset.scala:3258)
at org.apache.spark.sql.Dataset.head(Dataset.scala:2489)
at org.apache.spark.sql.Dataset.take(Dataset.scala:2703)
at org.apache.spark.sql.Dataset.showString(Dataset.scala:254)
at org.apache.spark.sql.Dataset.show(Dataset.scala:723)
通过 option("es.nodes.wan.only", value = true)
将配置项设置为 true 后恢复正常。
在遍历 DataFrame 时遇到如下编译错误,
Unable to find encoder for type stored in a Dataset. Primitive types (Int, String, etc) and Product types (case classes) are supported by importing spark.implicits._
在处理 DataFrame 之前需要加上 importing spark.implicits._
,用于将常见的 Scala 对象转换为 DataFrame,通常在获取 SparkSession 后立马 import。
WrappedArray
而不是 Array
当我们通过 createOrReplaceTempView("temp_labels")
构建一个临时表视图后,就可以通过 SQL 像操作 hive 表那样读取数据。例如读取指定的列,
val sqlDf = spark.sql("select itemId, labels from temp_labels where itemId in (123, 456)")
通过 sqlDf.printSchema()
可以看到 sqlDf 的 schema 长这样,
root
|-- itemId: long (nullable = true)
|-- labels: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- id: long (nullable = true)
| | |-- label: string (nullable = true)
labels
是包含 struct 的数组,于是从 row 中将 labels
列读出时想尝试转换为 Array,
val newDf = sqlDf.map(
row => {
val labels = row.getAs[Array[Row]]("labels")
val labelValue = labels.find(p => p.getAs[Long]("id") == attributeId).map(p => p.getAs[String]("label"))
(row.getAs[Long]("itemId"), attributeId, labelValue.orNull)
}
)
结果报错如下,
java.lang.ClassCastException: scala.collection.mutable.WrappedArray$ofRef cannot be cast to [Lorg.apache.spark.sql.Row;
可以看到 Spark SQL 在读取表中数组列时,是用的 scala.collection.mutable.WrappedArray
来存储结果的,看其类定义可知,它是间接实现 Seq 接口的,所以也可用 row.getAs[Seq[Row]]("labels")
来读取。这里需要注意的是,Array[T] 虽然在 Scala 源码定义中是 class,但其对标的 Java 类型是原生数组 T[]。
is null
或 is not null
,而不是 ===
或 !==
对于错误的用法,filter 并不会生效,就像下面这样
newDf.filter(col("label") !== null)
这一点和 hive 表以及 MySQL 表判断字段是否为 null,是保持一致的,应该像下面这样,
newDf.filter(col("label").isNotNull)
最终代码
import com.google.common.base.Joiner
import com.typesafe.scalalogging.LazyLogging
import org.apache.spark.sql.{DataFrame, Row, SaveMode, SparkSession}
object TestMain extends LazyLogging {
def main(args: Array[String]): Unit = {
val myUtils = new MyUtils
new TestApp(myUtils).run()
}
}
class TestApp(myUtils: MyUtils) extends Serializable with LazyLogging {
def esDf(spark: SparkSession, indices: Array[String]): DataFrame = {
spark.read
.format("es")
.option("es.nodes", myUtils.esHost())
.option("es.port", "9200")
.option("es.nodes.wan.only", value = true)
.option("es.resource", Joiner.on(",").join(java.util.Arrays.asList(indices:_*)) + "/label")
.option("es.scroll.size", 2000)
.load()
}
def run(): Unit = {
val spark = myUtils.getSparkSession
import spark.implicits._
val esTempView = "es_label"
val labelNames = Array("aen-label-retail", "aen-label-seller")
esDf(spark, labelNames).createOrReplaceTempView(esTempView)
val labelDf = getLabelDf(spark, itemIdsStr, attributeTypeIds, esTempView)
println("debug log")
labelDf.printSchema()
labelDf.show()
labelDf.createOrReplaceTempView("final_labels")
val data = spark.sql(
s"""
|select cc.*, pp.final_result, pp.label, null as remark
|from temp_request cc
|left join final_labels pp
|on cc.itemid = pp.itemId
|and cc.attributetypeid = pp.attributeId
|where cc.profile = '$jobId'
|""".stripMargin)
data.distinct().write.mode(SaveMode.Overwrite)
.option("compression", "gzip")
.json(s"s3://sherlockyb-test/check-precision/job_id=$jobId")
}
def getLabelDf(spark: SparkSession, itemIdsStr: String, attributeTypeIds: Array[String], esTempView: String): DataFrame = {
import spark.implicits._
val sqlDf = spark.sql(s"select itemId, labels from $esTempView where itemId in ($itemIdsStr)")
val emptyDf = spark.emptyDataFrame
var labelDf = emptyDf
attributeTypeIds.foreach(attributeTypeId => {
val attributeDf = sqlDf
.map(row => {
val labels = row.getAs[Seq[Row]]("labels")
val labelValue = labels.find(p => p.getAs[Long]("id") == attributeTypeId.toLong).map(p => p.getAs[String]("label"))
(row.getAs[Long]("itemId"), attributeTypeId.toLong, labelValue.orNull)
})
.withColumn("final_result", lit("PASS"))
.toDF("itemId", "attributeId", "label", "final_result")
.filter(col("label").isNotNull)
if (labelDf == emptyDf) {
labelDf = attributeDf
} else {
labelDf = labelDf.union(attributeDf)
}
})
labelDf.dropDuplicates("itemId","attributeId")
}
}
补充:提交 spark job
将 job 工程打包为 Jar,上传到 AWS 的 s3,比如 s3://sherlockyb-test/1.0.0/artifacts/spark/
目录下,然后通过 Genie 提交 spark job 到 Spark 集群运行。Genie 是 Netflix 研发的联合作业执行引擎,提供 REST-full API 来运行各种大数据作业,如 Hadoop、Pig、Hive、Spark、Presto、Sqoop 等。
def run_spark(job_name, spark_jar_name, spark_class_name, arg_str, spark_param=''):
import pygenie
pygenie.conf.DEFAULT_GENIE_URL = "genie.sherlockyb.club"
job = pygenie.jobs.GenieJob() \
.genie_username('sherlockyb') \
.job_name(job_name) \
.job_version('0.0.1') \
.metadata(teamId='team_account') \
.metadata(teamCredential='team_password')
job.cluster_tags(['type:yarn-kerberos', 'sched:default'])
job.command_tags(['type:spark-submit-kerberos', 'ver:2.3.2'])
job.command_arguments(
f"--class {spark_class_name} {spark_param} "
f"s3a://sherlockyb-test/1.0.0/artifacts/spark/{spark_jar_name} "
f"{arg_str}"
)
# Submit the job to Genie
running_job = job.execute()
running_job.wait()
return running_job.status
首发链接: https://www.yangbing.club/2022/06/03/Spark-reading-elasticsearch-guide/
许可协议: 除特殊声明外,本博文均采用 CC BY-NC-SA 3.0 CN 许可协议,转载请注明出处!