Apache Spark pool configurations in Azure Synapse Analytics
A Spark pool is a set of metadata that defines the compute resource requirements and associated behavior characteristics when a Spark instance is instantiated. These characteristics include but aren't limited to name, number of nodes, node size, scaling behavior, and time to live. A Spark pool in itself doesn't consume any resources. There are no costs incurred with creating Spark pools. Charges are only incurred once a Spark job is executed on the target Spark pool and the Spark instance is instantiated on demand.
You can read how to create a Spark pool and see all their properties here Get started with Spark pools in Synapse Analytics
The Isolated Compute option provides more security to Spark compute resources from untrusted services by dedicating the physical compute resource to a single customer. Isolated compute option is best suited for workloads that require a high degree of isolation from other customer's workloads for reasons that include meeting compliance and regulatory requirements. The Isolate Compute option is only available with the XXXLarge (80 vCPU / 504 GB) node size and only available in the following regions. The isolated compute option can be enabled or disabled after pool creation although the instance might need to be restarted. If you expect to enable this feature in the future, ensure that your Synapse workspace is created in an isolated compute supported region.
- East US
- West US 2
- South Central US
- US Gov Arizona
- US Gov Virginia
Apache Spark pool instance consists of one head node and two or more worker nodes with a minimum of three nodes in a Spark instance. The head node runs extra management services such as Livy, Yarn Resource Manager, Zookeeper, and the Spark driver. All nodes run services such as Node Agent and Yarn Node Manager. All worker nodes run the Spark Executor service.
A Spark pool can be defined with node sizes that range from a Small compute node with 4 vCore and 32 GB of memory up to a XXLarge compute node with 64 vCore and 432 GB of memory per node. Node sizes can be altered after pool creation although the instance might need to be restarted.
| Size | vCore | Memory |
|---|---|---|
| Small | 4 | 32 GB |
| Medium | 8 | 64 GB |
| Large | 16 | 128 GB |
| XLarge | 32 | 256 GB |
| XXLarge | 64 | 432 GB |
| XXX Large (Isolated Compute) | 80 | 504 GB |
Autoscale for Apache Spark pools allows automatic scale up and down of compute resources based on the amount of activity. When the autoscale feature is enabled, you set the minimum, and maximum number of nodes to scale. When the autoscale feature is disabled, the number of nodes set will remain fixed. This setting can be altered after pool creation although the instance might need to be restarted.
Apache Spark pools now support elastic pool storage. Elastic pool storage allows the Spark engine to monitor worker node temporary storage and attach extra disks if needed. Apache Spark pools utilize temporary disk storage while the pool is instantiated. Spark jobs write shuffle map outputs, shuffle data and spilled data to local VM disks. Examples of operations that could utilize local disk are sort, cache, and persist. When temporary VM disk space runs out, Spark jobs could fail due to “Out of Disk Space” error (java.io.IOException: No space left on device). With “Out of Disk Space” errors, much of the burden to prevent jobs from failing shifts to the customer to reconfigure the Spark jobs (for example, tweak the number of partitions) or clusters (for example, add more nodes to the cluster). These errors might not be consistent, and the user might end up experimenting heavily by running production jobs. This process can be expensive for the user in multiple dimensions:
- Wasted time. Customers are required to experiment heavily with job configurations via trial and error and are expected to understand Spark’s internal metrics to make the correct decision.
- Wasted resources. Since production jobs can process varying amount of data, Spark jobs can fail nondeterministically if resources aren't over-provisioned. For instance, consider the problem of data skew, which could result in a few nodes requiring more disk space than others. Currently in Synapse, each node in a cluster gets the same size of disk space and increasing disk space across all nodes isn't an ideal solution and leads to tremendous waste.
- Slowdown in job execution. In the hypothetical scenario where we solve the problem by autoscaling nodes (assuming costs aren't an issue to the end customer), adding a compute node is still expensive (takes a few minutes) as opposed to adding storage (takes a few seconds).
No action is required by you, plus you should see fewer job failures as a result.
Note
Azure Synapse Elastic pool storage is currently in Public Preview. During Public Preview there is no charge for use of Elastic pool storage.
The automatic pause feature releases resources after a set idle period, reducing the overall cost of an Apache Spark pool. The number of minutes of idle time can be set once this feature is enabled. The automatic pause feature is independent of the autoscale feature. Resources can be paused whether the autoscale is enabled or disabled. This setting can be altered after pool creation although active sessions will need to be restarted.