Performance Tuning, Cost Optimization / Internals, Research by Dmitry Tolpeko
By default, Spark EMR clusters have spark.dynamicAllocation.enabled set to true meaning that the cluster will dynamically allocate resources to scale the executors up and down whenever required.
But what is the initial number of executors when you start your Spark job?
Starting from EMR 5.32 and EMR 6.2 you can notice that Spark can launch much larger executors that you request in your job settings. For example, EMR created my cluster with the following default settings (it depends on the instance type and maximizeResourceAllocation classification option):
spark.executor.memory 18971M spark.executor.cores 4 spark.yarn.executor.memoryOverheadFactor 0.1875
But when I start a Spark session (pyspark command) I see the following:
I have a partitioned Hive table created by an open-source version of Hadoop that uses S3A scheme as the location for every partition. The table has more than 10,000 partitions and every partition has about 8,000 Parquet files:
$ hive -e "show partitions events"; ... dateint=20220419/hour=11 dateint=20220419/hour=12 dateint=20220419/hour=13 $ hive -e "describe formatted events partition (dateint=20220419, hour='11')" | grep Location Location: s3a://cloudsqale/events/dateint=20220419/hour=11 $ hive -e "describe formatted events partition (dateint=20220419, hour='12')" | grep Location Location: s3a://cloudsqale/events/dateint=20220419/hour=12 $ hive -e "describe formatted events partition (dateint=20220419, hour='13')" | grep Location Location: s3a://cloudsqale/events/dateint=20220419/hour=13
S3A:// is specified for every partition in this table.
Reading a Partition in Amazon EMR Spark
When I made an attempt to read data from a single partition using Spark SQL:
$ spark-sql --master yarn -e "select count(*) from events where dateint=20220419 and hour='11'"
The Spark driver failed with:
# java.lang.OutOfMemoryError: GC overhead limit exceeded # -XX:OnOutOfMemoryError="kill -9 %p" # Executing /bin/sh -c "kill -9 4847"...
Let’s review how EC2 vCPUs correspond to YARN vCores in Amazon EMR and Qubole Hadoop clusters. As an example, I will choose m4.4xlarge, r4.4xlarge and c4.4xlarge EC2 instance types.
EC2 vCPU is a thread of a CPU core (typically, there are two threads per core). Does it mean that YARN vCores should be equal to the number of EC2 vCPU? That’s not always the case.
I already wrote about recovering the ghost nodes in Amazon EMR clusters, but in this article I am going to extend this information for EMR clusters configured for auto-scaling.
In a Hadoop cluster besides Active nodes you may also have Unhealthy, Lost and Decommissioned nodes. Unhealthy nodes are running but just excluded from scheduling the tasks because they, for example, do not have enough disk space. Lost nodes are nodes that are not reachable anymore. The decommissioned nodes are nodes that successfully terminated and left the cluster.
But all these nodes are known to the Hadoop YARN cluster and you can see their details such as IP addresses, last health updates and so on.
At the same time there can be also Ghost nodes i.e. nodes that are running by Amazon EMR services but Hadoop itself does not know anything about their existence. Let’s see how you can find them.
Usually Hadoop is able to automatically recover cluster nodes from Unhealthy state by cleaning log and temporary directories. But sometimes nodes stay unhealthy for a long time and manual intervention is necessary to bring them back.
Amazon EMR allows you to define scale-out and scale-in rules to automatically add and remove instances based on the metrics you specify.
In this article I am going to explore the instance controller logs that can be very useful in monitoring the auto-scaling. The logs are located in /emr/instance-controller/log/ directory on the EMR master node.
When you run many Hadoop clusters it is useful to automatically collect metrics from all clusters in a single place (Hive table i.e.).
This allows you to perform any advanced and custom analysis of your clusters workload and not be limited to the features provided by Hadoop Administration UI tools that often offer only per cluster view so it is hard to see the whole picture of your data platform.
After creating an Amazon EMR cluster with Spark support, and running a spark application you can notice that the Spark job creates too many tasks to process even a very small data set.
For example, I have a small table country_iso_codes having 249 rows and stored in a comma-delimited text file with the length of 10,657 bytes.
When running the following application on Amazon EMR 5.7 cluster with Spark 2.1.1 with the default settings I can see the large number of partitions generated: