Amazon EMR Serverless lets you run open supply massive knowledge frameworks corresponding to Apache Spark and Apache Hive with out managing clusters and servers. With EMR Serverless, you possibly can run analytics workloads at any scale with computerized scaling that resizes sources in seconds to satisfy altering knowledge volumes and processing necessities.
We’ve launched job employee metrics in Amazon CloudWatch for EMR Serverless. This function lets you monitor vCPUs, reminiscence, ephemeral storage, and disk I/O allocation and utilization metrics at an mixture employee stage in your Spark and Hive jobs.
This submit is a part of a sequence about EMR Serverless observability. On this submit, we talk about methods to use these CloudWatch metrics to watch EMR Serverless staff in close to actual time.
CloudWatch metrics for EMR Serverless
On the per-Spark job stage, EMR Serverless emits the next new metrics to CloudWatch for each driver and executors. These metrics present granular insights into job efficiency, bottlenecks, and useful resource utilization.
| WorkerCpuAllocated | The whole numbers of vCPU cores allotted for staff in a job run |
| WorkerCpuUsed | The whole numbers of vCPU cores utilized by staff in a job run |
| WorkerMemoryAllocated | The whole reminiscence in GB allotted for staff in a job run |
| WorkerMemoryUsed | The whole reminiscence in GB utilized by staff in a job run |
| WorkerEphemeralStorageAllocated | The variety of bytes of ephemeral storage allotted for staff in a job run |
| WorkerEphemeralStorageUsed | The variety of bytes of ephemeral storage utilized by staff in a job run |
| WorkerStorageReadBytes | The variety of bytes learn from storage by staff in a job run |
| WorkerStorageWriteBytes | The variety of bytes written to storage from staff in a job run |
The next are the advantages of monitoring your EMR Serverless jobs with CloudWatch:
- Optimize useful resource utilization – You may achieve insights into useful resource utilization patterns and optimize your EMR Serverless configurations for higher effectivity and value financial savings. For instance, underutilization of vCPUs or reminiscence can reveal useful resource wastage, permitting you to optimize employee sizes to attain potential price financial savings.
- Diagnose frequent errors – You may determine root causes and mitigation for frequent errors with out log diving. For instance, you possibly can monitor the utilization of ephemeral storage and mitigate disk bottlenecks by preemptively allocating extra storage per employee.
- Acquire close to real-time insights – CloudWatch affords close to real-time monitoring capabilities, permitting you to trace the efficiency of your EMR Serverless jobs as and when they’re operating, for fast detection of any anomalies or efficiency points.
- Configure alerts and notifications – CloudWatch lets you arrange alarms utilizing Amazon Easy Notification Service (Amazon SNS) based mostly on predefined thresholds, permitting you to obtain notifications by e mail or textual content message when particular metrics attain important ranges.
- Conduct historic evaluation – CloudWatch shops historic knowledge, permitting you to research developments over time, determine patterns, and make knowledgeable selections for capability planning and workload optimization.
Answer overview
To additional improve this observability expertise, we’ve got created an answer that gathers all these metrics on a single CloudWatch dashboard for an EMR Serverless software. It’s worthwhile to launch one AWS CloudFormation template per EMR Serverless software. You may monitor all the roles submitted to a single EMR Serverless software utilizing the identical CloudWatch dashboard. To be taught extra about this dashboard and deploy this resolution into your individual account, discuss with the EMR Serverless CloudWatch Dashboard GitHub repository.
Within the following sections, we stroll you thru how you should use this dashboard to carry out the next actions:
- Optimize your useful resource utilization to avoid wasting prices with out impacting job efficiency
- Diagnose failures on account of frequent errors with out the necessity for log diving and resolve these errors optimally
Stipulations
To run the pattern jobs supplied on this submit, you’ll want to create an EMR Serverless software with default settings utilizing the AWS Administration Console or AWS Command Line Interface (AWS CLI), after which launch the CloudFormation template from the GitHub repo with the EMR Serverless software ID supplied because the enter to the template.
It’s worthwhile to submit all the roles on this submit to the identical EMR Serverless software. If you wish to monitor a special software, you possibly can deploy this template in your personal EMR Serverless software ID.
Optimize useful resource utilization
When operating Spark jobs, you usually begin with the default configurations. It may be difficult to optimize your workload with none visibility into precise useful resource utilization. Among the commonest configurations that we’ve seen clients regulate are spark.driver.cores, spark.driver.reminiscence, spark.executor.cores, and spark.executors.reminiscence.
For example how the newly added CloudWatch dashboard worker-level metrics may help you fine-tune your job configurations for higher price-performance and enhanced useful resource utilization, let’s run the next Spark job, which makes use of the NOAA Built-in Floor Database (ISD) dataset to run some transformations and aggregations.
Use the next command to run this job on EMR Serverless. Present your Amazon Easy Storage Service (Amazon S3) bucket and EMR Serverless software ID for which you launched the CloudFormation template. Be sure to make use of the identical software ID to submit all of the pattern jobs on this submit. Moreover, present an AWS Identification and Entry Administration (IAM) runtime function.
Now let’s examine the executor vCPUs and reminiscence from the CloudWatch dashboard.
This job was submitted with default EMR Serverless Spark configurations. From the Executor CPU Allotted metric within the previous screenshot, the job was allotted 396 vCPUs in complete (99 executors * 4 vCPUs per executor). Nonetheless, the job solely used a most of 110 vCPUs based mostly on Executor CPU Used. This means oversubscription of vCPU sources. Equally, the job was allotted 1,584 GB reminiscence in complete based mostly on Executor Reminiscence Allotted. Nonetheless, from the Executor Reminiscence Used metric, we see that the job solely used 176 GB of reminiscence through the job, indicating reminiscence oversubscription.
Now let’s rerun this job with the next adjusted configurations.
| Authentic Job (Default Configuration) | Rerun Job (Adjusted Configuration) | |
| spark.executor.reminiscence | 14 GB | 3 GB |
| spark.executor.cores | 4 | 2 |
| spark.dynamicAllocation.maxExecutors | 99 | 30 |
| Complete Useful resource Utilization | 6.521 vCPU-hours 26.084 memoryGB-hours 32.606 storageGB-hours | 1.739 vCPU-hours 3.688 memoryGB-hours 17.394 storageGB-hours |
| Billable Useful resource Utilization | 7.046 vCPU-hours 28.182 memoryGB-hours 0 storageGB-hours | 1.739 vCPU-hours 3.688 memoryGB-hours 0 storageGB-hours |
We use the next code:
Let’s examine the executor metrics from the CloudWatch dashboard once more for this job run.
Within the second job, we see decrease allocation of each vCPUs (396 vs. 60) and reminiscence (1,584 GB vs. 120 GB) as anticipated, leading to higher utilization of sources. The unique job ran for 4 minutes, 41 seconds. The second job took 4 minutes, 54 seconds. This reconfiguration has resulted in 79% decrease price financial savings with out affecting the job efficiency.
You should utilize these metrics to additional optimize your job by rising or reducing the variety of staff or the allotted sources.
Diagnose and resolve job failures
Utilizing the CloudWatch dashboard, you possibly can diagnose job failures on account of points associated to CPU, reminiscence, and storage corresponding to out of reminiscence or no house left on the system. This lets you determine and resolve frequent errors rapidly with out having to examine the logs or navigate by Spark Historical past Server. Moreover, as a result of you possibly can examine the useful resource utilization from the dashboard, you possibly can fine-tune the configurations by rising the required sources solely as a lot as wanted as an alternative of oversubscribing to the sources, which additional saves prices.
Driver errors
For example this use case, let’s run the next Spark job, which creates a big Spark knowledge body with a couple of million rows. Usually, this operation is finished by the Spark driver. Whereas submitting the job, we additionally configure spark.rpc.message.maxSize, as a result of it’s required for job serialization of knowledge frames with numerous columns.
After a couple of minutes, the job failed with the error message “Encountered errors when releasing containers,” as seen within the Job particulars part.
When encountering non-descriptive error messages, it turns into essential to analyze additional by analyzing the motive force and executor logs to troubleshoot additional. However earlier than additional log diving, let’s first examine the CloudWatch dashboard, particularly the motive force metrics, as a result of releasing containers is mostly carried out by the motive force.
We are able to see that the Driver CPU Used and Driver Storage Used are effectively inside their respective allotted values. Nonetheless, upon checking Driver Reminiscence Allotted and Driver Reminiscence Used, we will see that the motive force was utilizing the entire 16 GB reminiscence allotted to it. By default, EMR Serverless drivers are assigned 16 GB reminiscence.
Let’s rerun the job with extra driver reminiscence allotted. Let’s set driver reminiscence to 27 GB as the place to begin, as a result of spark.driver.reminiscence + spark.driver.memoryOverhead ought to be lower than 30 GB for the default employee sort. park.rpc.messsage.maxSize shall be unchanged.
The job succeeded this time round. Let’s examine the CloudWatch dashboard to look at driver reminiscence utilization.
As we will see, the allotted reminiscence is now 30 GB, however the precise driver reminiscence utilization didn’t exceed 21 GB through the job run. Subsequently, we will additional optimize prices right here by lowering the worth of spark.driver.reminiscence. We reran the identical job with spark.driver.reminiscence set to 22 GB, and the job nonetheless succeeded with higher driver reminiscence utilization.
Executor errors
Utilizing CloudWatch for observability is right for diagnosing driver-related points as a result of there is just one driver per job and driver sources used is the precise useful resource utilization of the only driver. However, executor metrics are aggregated throughout all the employees. Nonetheless, you should use this dashboard to supply solely an ample quantity of sources to make your job succeed, thereby avoiding oversubscription of sources.
For example, let’s run the next Spark job, which simulates uniform disk over-utilization throughout all staff by processing very massive NOAA datasets from a number of years. This job additionally transiently caches a really massive knowledge body on disk.
After a couple of minutes, we will see that the job failed with “No house left on system” error within the Job particulars part, which signifies that among the staff have run out of disk house.
Checking the Working Executors metric from the dashboard, we will determine that there have been 99 executor staff operating. Every employee comes with 20 GB storage by default.
As a result of it is a Spark job failure, let’s examine the Executor Storage Allotted and Executor Storage Used metrics from the dashboard (as a result of the motive force gained’t run any duties).
As we will see, the 99 executors have used up a complete of 1,940 GB from the whole allotted executor storage of two,126 GB. This contains each the info shuffled by the executors and the storage used for caching the info body. We don’t see the complete 2,126 GB being utilized from this graph as a result of there is perhaps a couple of executors out of the 99 executors that weren’t holding a lot knowledge when the job failed (earlier than these executors may begin processing duties and retailer the info body chunks).
Let’s rerun the identical job however with elevated executor disk measurement utilizing the parameter spark.emr-serverless.executor.disk. Let’s attempt with 40 GB disk per executor as a place to begin.
This time, the job ran efficiently. Let’s examine the Executor Storage Allotted and Executor Storage Used metrics.
Executor Storage Allotted is now 4,251 GB as a result of we’ve doubled the worth of spark.emr-serverless.executor.disk. Though there’s now twice as a lot aggregated executors’ storage, the job nonetheless used solely a most of 1,940 GB out of 4,251 GB. This means that our executors had been possible operating out of disk house solely by a couple of GBs. Subsequently, we will attempt to set spark.emr-serverless.executor.disk to a fair decrease worth like 25 GB or 30 GB as an alternative of 40 GB to avoid wasting storage prices as we did within the earlier situation. As well as, you possibly can monitor Executor Storage Learn Bytes and Executor Storage Write Bytes to see in case your job is I/O intensive. On this case, you should use the Shuffle-optimized disks function of EMR Serverless to additional improve your job’s I/O efficiency.
The dashboard can be helpful to seize details about transient storage used whereas caching or persisting the info frames, together with spill-to-disk situations. The Storage tab of Spark Historical past Server data any caching actions, as seen within the following screenshot. Nonetheless, this knowledge shall be misplaced from Spark Historical past Server after the cache is evicted or when the job finishes. Subsequently, Executor Storage Used can be utilized to do an evaluation of a failed job run on account of transient storage points.
On this explicit instance, the info was evenly distributed among the many executors. Nonetheless, when you’ve got a knowledge skew (for, instance just one–2 executors out of 99 course of essentially the most quantity of knowledge, and consequently, your job runs out of disk house), the CloudWatch dashboard gained’t precisely seize this situation as a result of the storage knowledge is aggregated throughout all of the executors for a job. For diagnosing points on the particular person executor stage, we have to observe per-executor-level metrics. We discover extra superior examples of how per-worker-level metrics may help you determine, mitigate, and resolve hard-to-find points by EMR Serverless integration with Amazon Managed Service for Prometheus.
Conclusion
On this submit, you discovered methods to successfully handle and optimize your EMR Serverless software utilizing a single CloudWatch dashboard with enhanced EMR Serverless metrics. These metrics can be found in all AWS Areas the place EMR Serverless is out there. For extra particulars about this function, discuss with Job-level monitoring.
Concerning the Authors
Kashif Khan is a Sr. Analytics Specialist Options Architect at AWS, specializing in massive knowledge companies like Amazon EMR, AWS Lake Formation, AWS Glue, Amazon Athena, and Amazon DataZone. With over a decade of expertise within the massive knowledge area, he possesses intensive experience in architecting scalable and sturdy options. His function includes offering architectural steering and collaborating carefully with clients to design tailor-made options utilizing AWS analytics companies to unlock the complete potential of their knowledge.
Veena Vasudevan is a Principal Associate Options Architect and Knowledge & AI specialist at AWS. She helps clients and companions construct extremely optimized, scalable, and safe options; modernize their architectures; and migrate their massive knowledge, analytics, and AI/ML workloads to AWS.