使用Spark in Cluster模式将文件写入本地系统

Question

I know this is a weird way of using Spark but I'm trying to save a dataframe to the local file system (not hdfs) using Spark even though I'm in cluster mode . I know I can use client mode but I do want to run in cluster mode and don't care which node (out of 3) the application is going to run on as driver. The code below is the pseudo code of what I'm trying to do.

// create dataframe
val df = Seq(Foo("John", "Doe"), Foo("Jane", "Doe")).toDF()
// save it to the local file system using 'file://' because it defaults to hdfs://
df.coalesce(1).rdd.saveAsTextFile(s"file://path/to/file")

And this is how I'm submitting the spark application.

spark-submit --class sample.HBaseSparkRSample --master yarn-cluster hbase-spark-r-sample-assembly-1.0.jar

This works fine if I'm in local mode but doesn't in yarn-cluster mode .

For example, java.io.IOException: Mkdirs failed to create file occurs with the above code.

I've changed the df.coalesce(1) part to df.collect and attempted to save a file using plain Scala but it ended up with a Permission denied .

I've also tried:

• spark-submit with root user
• chown ed yarn:yarn , yarn:hadoop , spark:spark
• gave chmod 777 to related directories

but no luck.

I'm assuming this has to do something with clusters , drivers and executors , and the user who's trying to write to the local file system but am pretty much stuck in solving this problem by myself.

I'm using:

• Spark: 1.6.0-cdh5.8.2
• Scala: 2.10.5
• Hadoop: 2.6.0-cdh5.8.2

Any support is welcome and thanks in advance.

Some articles I've tried:

• "Spark saveAsTextFile() results in Mkdirs failed to create for half of the directory" -> Tried changing users but nothing changed
• "Failed to save RDD as text file to local file system" -> chmod didn't help me

Edited (2016/11/25)

This is the Exception I get.

java.io.IOException: Mkdirs failed to create file:/home/foo/work/rhbase/r/input/input.csv/_temporary/0/_temporary/attempt_201611242024_0000_m_000000_0 (exists=false, cwd=file:/yarn/nm/usercache/foo/appcache/application_1478068613528_0143/container_e87_1478068613528_0143_01_000001)
    at org.apache.hadoop.fs.ChecksumFileSystem.create(ChecksumFileSystem.java:449)
    at org.apache.hadoop.fs.ChecksumFileSystem.create(ChecksumFileSystem.java:435)
    at org.apache.hadoop.fs.FileSystem.create(FileSystem.java:920)
    at org.apache.hadoop.fs.FileSystem.create(FileSystem.java:813)
    at org.apache.hadoop.mapred.TextOutputFormat.getRecordWriter(TextOutputFormat.java:135)
    at org.apache.spark.SparkHadoopWriter.open(SparkHadoopWriter.scala:91)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$saveAsHadoopDataset$1$$anonfun$13.apply(PairRDDFunctions.scala:1193)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$saveAsHadoopDataset$1$$anonfun$13.apply(PairRDDFunctions.scala:1185)
    at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:66)
    at org.apache.spark.scheduler.Task.run(Task.scala:89)
    at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:214)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
    at java.lang.Thread.run(Thread.java:745)
16/11/24 20:24:12 WARN scheduler.TaskSetManager: Lost task 0.0 in stage 0.0 (TID 0, localhost): java.io.IOException: Mkdirs failed to create file:/home/foo/work/rhbase/r/input/input.csv/_temporary/0/_temporary/attempt_201611242024_0000_m_000000_0 (exists=false, cwd=file:/yarn/nm/usercache/foo/appcache/application_1478068613528_0143/container_e87_1478068613528_0143_01_000001)
    at org.apache.hadoop.fs.ChecksumFileSystem.create(ChecksumFileSystem.java:449)
    at org.apache.hadoop.fs.ChecksumFileSystem.create(ChecksumFileSystem.java:435)
    at org.apache.hadoop.fs.FileSystem.create(FileSystem.java:920)
    at org.apache.hadoop.fs.FileSystem.create(FileSystem.java:813)
    at org.apache.hadoop.mapred.TextOutputFormat.getRecordWriter(TextOutputFormat.java:135)
    at org.apache.spark.SparkHadoopWriter.open(SparkHadoopWriter.scala:91)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$saveAsHadoopDataset$1$$anonfun$13.apply(PairRDDFunctions.scala:1193)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$saveAsHadoopDataset$1$$anonfun$13.apply(PairRDDFunctions.scala:1185)
    at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:66)
    at org.apache.spark.scheduler.Task.run(Task.scala:89)
    at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:214)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
    at java.lang.Thread.run(Thread.java:745)

Answer 1

I'm going to answer my own question because eventually, none of the answers didn't seem to solve my problem. None the less, thanks for all the answers and pointing me to alternatives I can check.

I think @Ricardo was the closest in mentioning the user of the Spark application. I checked whoami with Process("whoami") and the user was yarn . The problem was probably that I tried to output to /home/foo/work/rhbase/r/input/input.csv and although /home/foo/work/rhbase was owned by yarn:yarn , /home/foo was owned by foo:foo . I haven't checked in detail but this may have been the cause of this permission problem.

When I hit pwd in my Spark application with Process("pwd") , it output /yarn/path/to/somewhere . So I decided to output my file to /yarn/input.csv and it was successful despite in cluster mode .

I probably can conclude that this was just a simple permission issue. Any further solution would be welcome but for now, this was the way how I solved this question.

Answer 2

Use forEachPartition method, and then for each partition get file system object and write one by one record to it, below is the sample code here i am writing to hdfs, instead you can use local file system as well

Dataset<String> ds=....

ds.toJavaRdd.foreachPartition(new VoidFunction<Iterator<String>>() {
    @Override
    public void call(Iterator<String> iterator) throws Exception {

    final FileSystem hdfsFileSystem = FileSystem.get(URI.create(finalOutPathLocation), hadoopConf);

    final FSDataOutputStream fsDataOutPutStream = hdfsFileSystem.exists(finalOutPath)
            ? hdfsFileSystem.append(finalOutPath) : hdfsFileSystem.create(finalOutPath);


    long processedRecCtr = 0;
    long failedRecsCtr = 0;


    while (iterator.hasNext()) {

        try {
            fsDataOutPutStream.writeUTF(iterator.next);
        } catch (Exception e) {
            failedRecsCtr++;
        }
        if (processedRecCtr % 3000 == 0) {
            LOGGER.info("Flushing Records");
            fsDataOutPutStream.flush();
        }
    }

    fsDataOutPutStream.close();
        }
});

Answer 3

If you run the job as yarn-cluster mode , the driver will be running in any of the machine which is managed by YARN, so if saveAsTextFile has local file path, then it will store the output in any of the machine where driver is running.

Try running the job as yarn-client mode so the driver runs in the client machine

Answer 4

Please refer to the spark documentation to understand the use of --master option in spark-submit .

• --master local is supposed to be used when running locally.

• --master yarn --deploy-mode cluster is supposed to be used when actually running on a yarn cluster.

Refer this and this .

Answer 5

Check if you are trying to run/write the file with a user other than the Spark service. On that situation you can solve the permission issue by presetting the directory ACLs. Example:

setfacl -d -m group:spark:rwx /path/to/

(modify "spark" to your user group trying to write the file)