So, I have a 16 node cluster where every node has Spark and Cassandra installed with a replication factor of 3 and spark.sql.shuffle.partitions of 96. I am using the Spark-Cassandra Connector 3.0.0 and I am trying to join a dataset with a cassandra table on the partition key, while also using .repartitionByCassandraReplica.
However repartitionByCassandraReplica is implemented only on RDDs so I am converting my dataset to JavaRDD, do the repartitionByCassandraReplica, then converting it back to dataset and do a Direct Join with the cassandra table. It seems though, that in the process of that the number of partitions is "changing" or is not as expected.
I am doing a PCA on 4 partition keys which have some thousands of rows and for which I know the nodes where they are stored according to nodetool getendpoints . It looks like not only the number of partitions is changing but also the nodes where data are pulled are not the ones that actually have the data. Below is the code.
//FYI experimentlist is a List<String> which is converted to Dataset,then JavaRDD, then partitioned
//according to repartitionByCassandraReplica and then back to Dataset. The table with which I want to
//join it, is called experiment.
List<ExperimentForm> tempexplist = experimentlist.stream()
.map(s -> { ExperimentForm p = new ExperimentForm(); p.setExperimentid(s); return p; })
.collect(Collectors.toList());
Encoder<ExperimentForm> ExpEncoder = Encoders.bean(ExperimentForm.class);
Dataset<ExperimentForm> dfexplistoriginal = sp.createDataset(tempexplist, Encoders.bean(ExperimentForm.class));
//Below prints DATASET: PartNum 4
System.out.println("DATASET: PartNum "+dfexplistoriginal.rdd().getNumPartitions());
JavaRDD<ExperimentForm> predf = CassandraJavaUtil.javaFunctions(dfexplistoriginal.javaRDD()).repartitionByCassandraReplica("mdb","experiment",experimentlist.size(),CassandraJavaUtil.someColumns("experimentid"),CassandraJavaUtil.mapToRow(ExperimentForm.class));
//Below prints RDD: PartNum 64
System.out.println("RDD: PartNum "+predf.getNumPartitions());
Dataset<ExperimentForm> newdfexplist = sp.createDataset(predf.rdd(), ExpEncoder);
Dataset<Row> readydfexplist = newdfexplist.as(Encoders.STRING()).toDF("experimentid");
//Below prints DATASET: PartNum 64
System.out.println("DATASET: PartNum "+readydfexplist.rdd().getNumPartitions());
//and finally the DirectJoin which for some reason is not mentioned as DirectJoin in DAGs like other times
Dataset<Row> metlistinitial = sp.read().format("org.apache.spark.sql.cassandra")
.options(new HashMap<String, String>() {
{
put("keyspace", "mdb");
put("table", "experiment");
}
})
.load().select(col("experimentid"), col("description"), col("intensity")).join(readydfexplist,"experimentid");
Is the code wrong? Below are also some images from SparkUI the Stages Tab with DAGs. At first I have 4 tasks/partitions and after repartitionByCassandraReplica I get 64 or more. Why?
All the Stages:
Stage 0 DAG
Stage 0 Metrics
Stage 1 DAG
Stage 1 Some Metrics
Looks like the code I wrote above is not entirely correct! I managed to get repartitionByCassandraReplica working by just converting dataset to RDD, performing the repartitionByCassandraReplica doing the join with JoinWithCassandraTable and THEN converting back to dataset! Now it is indeed repartitioned on the nodes that actually have the data! Partitions are maintained between these conversions!
I am trying to read csv file and then adding some columns . After that trying to save in orc format.
I could not understand how spark decided number of tasks for different stages.
Why number of task for CSV stage is 1 and for ORC stage it is 39?
val c1c8 = spark.read.option("header",true).csv("/user/DEEPAK_TEST/C1C6_NEW/")
val c1c8new = { c1c8.withColumnRenamed("c1c6_F","c1c8").withColumnRenamed("Network_Out","c1c8_network").withColumnRenamed("Access NE Out","c1c8_access_ne")
.withColumn("c1c8_signalling",when (col("signalling_Out") === "SIP Cl4" , "SIP CL4").when (col("signalling_Out") === "SIP cl4" , "SIP CL4").when (col("signalling_Out") === "Other" , "other").otherwise(col("signalling_Out")))
.withColumnRenamed("access type Out","c1c8_access_type").withColumnRenamed("Type_of_traffic_C","c1c8_typeoftraffic")
.withColumnRenamed("BOS traffic type Out","c1c8_bos_trafc_typ").withColumnRenamed("Scope_Out","c1c8_scope")
.withColumnRenamed("Join with UP-DWN SIP cl5 T1T7 Out","c1c8_join_indicator")
.select("c1c8","c1c8_network", "c1c8_access_ne", "c1c8_signalling", "c1c8_access_type", "c1c8_typeoftraffic",
"c1c8_bos_trafc_typ", "c1c8_scope","c1c8_join_indicator")
}
c1c8new.write.orc("/user/DEEPAK_TEST/C1C8_MAPPING_NEWT/")
Below is my understanding from looking at Spark 2.x source code.
Stage 0 is a file scan that creates FileScanRDD which is an RDD that scans a list of file partitions. This stage can have more than one task when you are reading from multiple partitioned directories, such as a partitioned Hive table.
The number of tasks in Stage 1 will be equals to the number of RDD partitions. In your case c1c8new.rdd.getNumPartitions will be 39. This number is calculated using:
config value spark.files.maxPartitionBytes (128MB by default)
sparkContext.defaultParallelism returned by task scheduler (equal to number of cores when running in local mode)
totalBytes
DataSourceScanExec.scala#L423
val defaultMaxSplitBytes =
fsRelation.sparkSession.sessionState.conf.filesMaxPartitionBytes
val openCostInBytes = fsRelation.sparkSession.sessionState.conf.filesOpenCostInBytes
val defaultParallelism = fsRelation.sparkSession.sparkContext.defaultParallelism
val totalBytes = selectedPartitions.flatMap(_.files.map(_.getLen + openCostInBytes)).sum
val bytesPerCore = totalBytes / defaultParallelism
val maxSplitBytes = Math.min(defaultMaxSplitBytes, Math.max(openCostInBytes, bytesPerCore))
logInfo(s"Planning scan with bin packing, max size: $maxSplitBytes bytes, " +
s"open cost is considered as scanning $openCostInBytes bytes.")
You can see actual calculated values in the above log message if you set the log level to INFO - spark.sparkContext.setLogLevel("INFO")
In your case, I think the split size is 128 and so, number of tasks/partitions is roughly 4.6G/128MB
As a side note, you can change the number of partitions (and hence the number of tasks in the subsequent stage) by using repartition() or coalesce() on the dataframe. More importantly, the number of partitions after a shuffle is determined by spark.sql.shuffle.partitions (200 by default). If you have a shuffle, it is better to use this configuration to control the number of tasks because inserting repartition() or coalesce() between stages adds extra overhead.
For large spark SQL workloads, setting optimum values for spark.sql.shuffle.partitions in each stage was always a pain point. Spark 3.x has better support for this if Adaptive Query Execution is enabled, but I haven't tried it for any production workloads.
When using tensorflow java for inference the amount of memory to make the job run on YARN is abnormally large. The job run perfectly with spark on my computer (2 cores 16Gb of RAM) and take 35 minutes to complete. But when I try to run it on YARN with 10 executors 16Gb memory and 16 Gb memoryOverhead the executors are killed for using too much memory.
Prediction Run on an Hortonworks cluster with YARN 2.7.3 and Spark 2.2.1. Previously we used DL4J to do inference and everything run under 3 min.
Tensor are correctly closed after usage and we use a mapPartition to do prediction. Each task contain approximately 20.000 records (1Mb) so this will make input tensor of 2.000.000x14 and output tensor of 2.000.000 (5Mb).
option passed to spark when running on YARN
--master yarn --deploy-mode cluster --driver-memory 16G --num-executors 10 --executor-memory 16G --executor-cores 2 --conf spark.driver.memoryOverhead=16G --conf spark.yarn.executor.memoryOverhead=16G --conf spark.sql.shuffle.partitions=200 --conf spark.tasks.cpu=2
This configuration may work if we set spark.sql.shuffle.partitions=2000 but it take 3 hours
UPDATE:
The difference between local and cluster was in fact due to a missing filter. we actually run the prediction on more data than we though.
To reduce memory footprint of each partition you must create batch inside each partition (use grouped(batchSize)). Thus you are faster than running predict for each row and you allocate tensor of predermined size (batchSize). If you investigate the code of tensorflowOnSpark scala inference this is what they did. Below you will find a reworked example of an implementation this code may not compile but you get the idea of how to do it.
lazy val sess = SavedModelBundle.load(modelPath, "serve").session
lazy val numberOfFeatures = 1
lazy val laggedFeatures = Seq("cost_day1", "cost_day2", "cost_day3")
lazy val numberOfOutputs = 1
val predictionsRDD = preprocessedData.rdd.mapPartitions { partition =>
partition.grouped(batchSize).flatMap { batchPreprocessed =>
val numberOfLines = batchPreprocessed.size
val featuresShape: Array[Long] = Array(numberOfLines, laggedFeatures.size / numberOfFeatures, numberOfFeatures)
val featuresBuffer: FloatBuffer = FloatBuffer.allocate(numberOfLines)
for (
featuresWithKey <- batchPreprocessed;
feature <- featuresWithKey.features
) {
featuresBuffer.put(feature)
}
featuresBuffer.flip()
val featuresTensor = Tensor.create(featuresShape, featuresBuffer)
val results: Tensor[_] = sess.runner
.feed("cost", featuresTensor)
.fetch("prediction")
.run.get(0)
val output = Array.ofDim[Float](results.numElements(), numberOfOutputs)
val outputArray: Array[Array[Float]] = results.copyTo(output)
results.close()
featuresTensor.close()
outputArray
}
}
spark.createDataFrame(predictionsRDD)
We use FloatBuffer and Shape to create Tensor as recommended in this issue
I have spark job code as below. Which works fine with below configuration on cluster.
String path = "/tmp/one.txt";
JavaRDD<SomeClass> jRDD = spark.read()
.textFile(path)
.javaRDD()
.map(line -> {
return new SomeClass(line);
});
Dataset responseSet = sparkSession.createDataFrame(jRDD, SomeClass.class);
responseSet.write()
.format("text")
.save(path + "processed");
Whereas, If I want to read binary file(same size as text) it takes much more time.
String path = "/tmp/one.txt";
JavaRDD<SomeClass> jRDD = sparkContext
.binaryRecords(path, 10000, new Configuration())
.toJavaRDD()
.map(line -> {
return new SomeClass(line);
});
Dataset responseSet = spark.createDataFrame(jRDD, SomeClass.class);
responseSet.write()
.format("text")
.save(path + "processed");
Below is my configuration.
driver-memory 8g
executor-memory 6g
num-executors 16
Time taken by first code with 150 MB file is 1.30 mins.
Time taken by second code with 150 MB file is 4 mins.
Also, first code was able to run on all 16 executors whereas second uses only one.
ny suggestions why it is slow?
I found the issue. The textFile()method was creating 16 partitions(you can checknumOfPartitions using getNumPartitions() method on RDD) whereas binaryRecords() created only 1(Java binaryRecords doesn't provide overloaded method which specifies num of partitions to be created).
I increased numOfPartitions on RDD created by binaryRecords() by using repartition(NUM_OF_PARTITIONS) method on RDD.
I am trying to move data from a table in PostgreSQL table to a Hive table on HDFS. To do that, I came up with the following code:
val conf = new SparkConf().setAppName("Spark-JDBC").set("spark.executor.heartbeatInterval","120s").set("spark.network.timeout","12000s").set("spark.sql.inMemoryColumnarStorage.compressed", "true").set("spark.sql.orc.filterPushdown","true").set("spark.serializer", "org.apache.spark.serializer.KryoSerializer").set("spark.kryoserializer.buffer.max","512m").set("spark.serializer", classOf[org.apache.spark.serializer.KryoSerializer].getName).set("spark.streaming.stopGracefullyOnShutdown","true").set("spark.yarn.driver.memoryOverhead","7168").set("spark.yarn.executor.memoryOverhead","7168").set("spark.sql.shuffle.partitions", "61").set("spark.default.parallelism", "60").set("spark.memory.storageFraction","0.5").set("spark.memory.fraction","0.6").set("spark.memory.offHeap.enabled","true").set("spark.memory.offHeap.size","16g").set("spark.dynamicAllocation.enabled", "false").set("spark.dynamicAllocation.enabled","true").set("spark.shuffle.service.enabled","true")
val spark = SparkSession.builder().config(conf).master("yarn").enableHiveSupport().config("hive.exec.dynamic.partition", "true").config("hive.exec.dynamic.partition.mode", "nonstrict").getOrCreate()
def prepareFinalDF(splitColumns:List[String], textList: ListBuffer[String], allColumns:String, dataMapper:Map[String, String], partition_columns:Array[String], spark:SparkSession): DataFrame = {
val colList = allColumns.split(",").toList
val (partCols, npartCols) = colList.partition(p => partition_columns.contains(p.takeWhile(x => x != ' ')))
val queryCols = npartCols.mkString(",") + ", 0 as " + flagCol + "," + partCols.reverse.mkString(",")
val execQuery = s"select ${allColumns}, 0 as ${flagCol} from schema.tablename where period_year='2017' and period_num='12'"
val yearDF = spark.read.format("jdbc").option("url", connectionUrl).option("dbtable", s"(${execQuery}) as year2017")
.option("user", devUserName).option("password", devPassword)
.option("partitionColumn","cast_id")
.option("lowerBound", 1).option("upperBound", 100000)
.option("numPartitions",70).load()
val totalCols:List[String] = splitColumns ++ textList
val cdt = new ChangeDataTypes(totalCols, dataMapper)
hiveDataTypes = cdt.gpDetails()
val fc = prepareHiveTableSchema(hiveDataTypes, partition_columns)
val allColsOrdered = yearDF.columns.diff(partition_columns) ++ partition_columns
val allCols = allColsOrdered.map(colname => org.apache.spark.sql.functions.col(colname))
val resultDF = yearDF.select(allCols:_*)
val stringColumns = resultDF.schema.fields.filter(x => x.dataType == StringType).map(s => s.name)
val finalDF = stringColumns.foldLeft(resultDF) {
(tempDF, colName) => tempDF.withColumn(colName, regexp_replace(regexp_replace(col(colName), "[\r\n]+", " "), "[\t]+"," "))
}
finalDF
}
val dataDF = prepareFinalDF(splitColumns, textList, allColumns, dataMapper, partition_columns, spark)
val dataDFPart = dataDF.repartition(30)
dataDFPart.createOrReplaceTempView("preparedDF")
spark.sql("set hive.exec.dynamic.partition.mode=nonstrict")
spark.sql("set hive.exec.dynamic.partition=true")
spark.sql(s"INSERT OVERWRITE TABLE schema.hivetable PARTITION(${prtn_String_columns}) select * from preparedDF")
The data is inserted into the hive table dynamically partitioned based on prtn_String_columns: source_system_name, period_year, period_num
Spark-submit used:
SPARK_MAJOR_VERSION=2 spark-submit --conf spark.ui.port=4090 --driver-class-path /home/fdlhdpetl/jars/postgresql-42.1.4.jar --jars /home/fdlhdpetl/jars/postgresql-42.1.4.jar --num-executors 80 --executor-cores 5 --executor-memory 50G --driver-memory 20G --driver-cores 3 --class com.partition.source.YearPartition splinter_2.11-0.1.jar --master=yarn --deploy-mode=cluster --keytab /home/fdlhdpetl/fdlhdpetl.keytab --principal fdlhdpetl#FDLDEV.COM --files /usr/hdp/current/spark2-client/conf/hive-site.xml,testconnection.properties --name Splinter --conf spark.executor.extraClassPath=/home/fdlhdpetl/jars/postgresql-42.1.4.jar
The following error messages are generated in the executor logs:
Container exited with a non-zero exit code 143.
Killed by external signal
18/10/03 15:37:24 ERROR SparkUncaughtExceptionHandler: Uncaught exception in thread Thread[SIGTERM handler,9,system]
java.lang.OutOfMemoryError: Java heap space
at java.util.zip.InflaterInputStream.<init>(InflaterInputStream.java:88)
at java.util.zip.ZipFile$ZipFileInflaterInputStream.<init>(ZipFile.java:393)
at java.util.zip.ZipFile.getInputStream(ZipFile.java:374)
at java.util.jar.JarFile.getManifestFromReference(JarFile.java:199)
at java.util.jar.JarFile.getManifest(JarFile.java:180)
at sun.misc.URLClassPath$JarLoader$2.getManifest(URLClassPath.java:944)
at java.net.URLClassLoader.defineClass(URLClassLoader.java:450)
at java.net.URLClassLoader.access$100(URLClassLoader.java:73)
at java.net.URLClassLoader$1.run(URLClassLoader.java:368)
at java.net.URLClassLoader$1.run(URLClassLoader.java:362)
at java.security.AccessController.doPrivileged(Native Method)
at java.net.URLClassLoader.findClass(URLClassLoader.java:361)
at java.lang.ClassLoader.loadClass(ClassLoader.java:424)
at sun.misc.Launcher$AppClassLoader.loadClass(Launcher.java:331)
at java.lang.ClassLoader.loadClass(ClassLoader.java:357)
at org.apache.spark.util.SignalUtils$ActionHandler.handle(SignalUtils.scala:99)
at sun.misc.Signal$1.run(Signal.java:212)
at java.lang.Thread.run(Thread.java:745)
I see in the logs that the read is being executed properly with the given number of partitions as below:
Scan JDBCRelation((select column_names from schema.tablename where period_year='2017' and period_num='12') as year2017) [numPartitions=50]
Below is the state of executors in stages:
The data is not being partitioned properly. One partition is smaller while the other one becomes huge. There is a skew problem here.
While inserting the data into Hive table the job fails at the line:spark.sql(s"INSERT OVERWRITE TABLE schema.hivetable PARTITION(${prtn_String_columns}) select * from preparedDF") but I understand this is happening because of the data skew problem.
I tried to increase number of executors, increasing the executor memory, driver memory, tried to just save as csv file instead of saving the dataframe into a Hive table but nothing affects the execution from giving the exception:
java.lang.OutOfMemoryError: GC overhead limit exceeded
Is there anything in the code that I need to correct ? Could anyone let me know how can I fix this problem ?
Determine how many partitions you need given the amount of input data and your cluster resources. As a rule of thumb it is better to keep partition input under 1GB unless strictly necessary. and strictly smaller than the block size limit.
You've previously stated that you migrate 1TB of data values you use in different posts (5 - 70) are likely way to low to ensure smooth process.
Try to use value which won't require further repartitioning.
Know your data.
Analyze the columns available in the the dataset to determine if there any columns with high cardinality and uniform distribution to be distributed among desired number of partitions. These are good candidates for an import process. Additionally you should determine an exact range of values.
Aggregations with different centrality and skewness measure as well as histograms and basic counts-by-key are good exploration tools. For this part it is better to analyze data directly in the database, instead of fetching it to Spark.
Depending on the RDBMS you might be able to use width_bucket (PostgreSQL, Oracle) or equivalent function to get a decent idea how data will be distributed in Spark after loading with partitionColumn, lowerBound, upperBound, numPartitons.
s"""(SELECT width_bucket($partitionColum, $lowerBound, $upperBound, $numPartitons) AS bucket, COUNT(*)
FROM t
GROUP BY bucket) as tmp)"""
If there are no columns which satisfy above criteria consider:
Creating a custom one and exposing it via. a view. Hashes over multiple independent columns are usually good candidates. Please consult your database manual to determine functions that can be used here (DBMS_CRYPTO in Oracle, pgcrypto in PostgreSQL)*.
Using a set of independent columns which taken together provide high enough cardinality.
Optionally, if you're going to write to a partitioned Hive table, you should consider including Hive partitioning columns. It might limit the number of files generated later.
Prepare partitioning arguments
If column selected or created in the previous steps is numeric (or date / timestamp in Spark >= 2.4) provide it directly as the partitionColumn and use range values determined before to fill lowerBound and upperBound.
If bound values don't reflect the properties of data (min(col) for lowerBound, max(col) for upperBound) it can result in a significant data skew so thread carefully. In the worst case scenario, when bounds don't cover the range of data, all records will be fetched by a single machine, making it no better than no partitioning at all.
If column selected in the previous steps is categorical or is a set of columns generate a list of mutually exclusive predicates that fully cover the data, in a form that can be used in a SQL where clause.
For example if you have a column A with values {a1, a2, a3} and column B with values {b1, b2, b3}:
val predicates = for {
a <- Seq("a1", "a2", "a3")
b <- Seq("b1", "b2", "b3")
} yield s"A = $a AND B = $b"
Double check that conditions don't overlap and all combinations are covered. If these conditions are not satisfied you end up with duplicates or missing records respectively.
Pass data as predicates argument to jdbc call. Note that the number of partitions will be equal exactly to the number of predicates.
Put database in a read-only mode (any ongoing writes can cause data inconsistency. If possible you should lock database before you start the whole process, but if might be not possible, in your organization).
If the number of partitions matches the desired output load data without repartition and dump directly to the sink, if not you can try to repartition following the same rules as in the step 1.
If you still experience any problems make sure that you've properly configured Spark memory and GC options.
If none of the above works:
Consider dumping your data to a network / distributes storage using tools like COPY TO and read it directly from there.
Note that or standard database utilities you will typically need a POSIX compliant file system, so HDFS usually won't do.
The advantage of this approach is that you don't need to worry about the column properties, and there is no need for putting data in a read-only mode, to ensure consistency.
Using dedicated bulk transfer tools, like Apache Sqoop, and reshaping data afterwards.
* Don't use pseudocolumns - Pseudocolumn in Spark JDBC.
In my experience there are 4 kinds of memory settings which make a difference:
A) [1] Memory for storing data for processing reasons VS [2] Heap Space for holding the program stack
B) [1] Driver VS [2] executor memory
Up to now, I was always able to get my Spark jobs running successfully by increasing the appropriate kind of memory:
A2-B1 would therefor be the memory available on the driver to hold the program stack. Etc.
The property names are as follows:
A1-B1) executor-memory
A1-B2) driver-memory
A2-B1) spark.yarn.executor.memoryOverhead
A2-B2) spark.yarn.driver.memoryOverhead
Keep in mind that the sum of all *-B1 must be less than the available memory on your workers and the sum of all *-B2 must be less than the memory on your driver node.
My bet would be, that the culprit is one of the boldly marked heap settings.
There was an another question of yours routed here as duplicate
'How to avoid data skewing while reading huge datasets or tables into spark?
The data is not being partitioned properly. One partition is smaller while the
other one becomes huge on read.
I observed that one of the partition has nearly 2million rows and
while inserting there is a skew in partition. '
if the problem is to deal with data that is partitioned in a dataframe after read, Have you played around increasing the "numPartitions" value ?
.option("numPartitions",50)
lowerBound, upperBound form partition strides for generated WHERE clause expressions and numpartitions determines the number of split.
say for example, sometable has column - ID (we choose that as partitionColumn) ; value range we see in table for column-ID is from 1 to 1000 and we want to get all the records by running select * from sometable,
so we going with lowerbound = 1 & upperbound = 1000 and numpartition = 4
this will produce a dataframe of 4 partition with result of each Query by building sql based on our feed (lowerbound = 1 & upperbound = 1000 and numpartition = 4)
select * from sometable where ID < 250
select * from sometable where ID >= 250 and ID < 500
select * from sometable where ID >= 500 and ID < 750
select * from sometable where ID >= 750
what if most of the records in our table fall within the range of ID(500,750). that's the situation you are in to.
when we increase numpartition , the split happens even further and that reduce the volume of records in the same partition but this
is not a fine shot.
Instead of spark splitting the partitioncolumn based on boundaries we provide, if you think of feeding the split by yourself so, data can be evenly
splitted. you need to switch over to another JDBC method where instead of (lowerbound,upperbound & numpartition) we can provide
predicates directly.
def jdbc(url: String, table: String, predicates: Array[String], connectionProperties: Properties): DataFrame
Link