Impala auto-scaling on public clouds
Your Impala Virtual Warehouse in Cloudera Data Warehouse (CDW) Public Cloud has an auto-scaler process that works with coordinators and executors to make resources available for queued queries. This ensures that workload demand is met without wasting cloud resources.
Impala Virtual Warehouses use the following running processes to support auto-scaling:
- Coordinator processes: Handle all incoming queries, creating execution plans and handing off the query to executor processes for execution. By default, there are two coordinator processes to enable high availability and fault tolerance. The number of coordinator processes is controlled by the Enable Impala HA setting. This setting is enabled by default. When Enable Impala HA is enabled, there are two coordinator processes. If one coordinator process fails, a backup coordinator process takes over so there is no single point of failure. If Enable Impala HA is disabled, there is only one coordinator process available. Having the ability to toggle this setting on and off enables you to save on cloud resource consumption:
- Executor processes: Processes that execute query fragments. Executor processes run on executor nodes, the unit of sizing for Virtual Warehouses. Each executor node runs one executor:
Executor groups: A group of executors that can execute queries. Each executor group can run up to twelve queries. The size of executor groups is determined by the size you choose when you first create the Virtual Warehouse (XSMALL, SMALL, MEDIUM, or LARGE).
It is recommended that you select the size of a Virtual Warehouse based on the number of nodes you need to run a typical query for your workloads based on your normal query size and complexity. For example, if you expect that 20 nodes are needed to run a typical query in your workloads, you would create a medium-sized Virtual Warehouse, which by default has 20 executor nodes. However, if you have a large number of queries that must be run in your workloads concurrently, you can set the minimum node count to 40 and maximum node count to 80 by using the Nodes: Min: Max: setting. Then each executor group for this Virtual Warehouse would still contain the 20 executor nodes required to run an average query, but you would have two executor groups to handle the volume of queries in your workloads. If the total number of nodes auto-scales up to 80, then there will be 4 executor groups containing 20 nodes each.
The rationale behind selecting a size based on the normal size and complexity of your queries is that then the Virtual Warehouse can run all your normal queries in an acceptable time. Selecting size based on this criteria prevents you from "over-sizing" your Virtual Warehouse and as a result, unnecessarily incurring extra costs. On the other hand, keep in mind that under-sizing your Virtual Warehouse might result in queries spilling, which makes them run slower in spite of the fact that they may eventually complete correctly.
- Auto-scaler: A process that monitors Impala to determine when more or fewer resources are needed. When the auto-scaler detects an imbalance in resources, it sends a request to the Kubernetes framework to increase or decrease the number of executor groups in the Virtual Warehouse.
- When users connect to Impala Virtual Warehouses using SQL applications such as Hue, the Impala shell, or other SQL clients that use JDBC or ODBC, the query is handled by the coordinator process. First, the coordinator generates an execution plan for the query. The execution plan includes an estimate of the memory required to run the query.
- Then the coordinator locates an executor group that has enough available memory to run the query. Each executor group is limited by the number of queries and the memory available to run them. Currently, executor groups can handle up to 3 large queries. Executors can handle more queries that are simpler and that do not utilize concurrency on the executor. If there are no executor groups with enough available resources to handle a query, it is queued until resources become available.
- When the auto-scaler detects that there are queued queries, it adds executor groups to
the Virtual Warehouse to execute the queries. The auto-scaler starts scaling up and down
the number of executor groups according to what is set for the auto-scale settings:
- Scale Up Delay: Sets the length of time in seconds that the system waits before adding more executors when it detects queries waiting in the queue to execute. The time to auto-scale is affected by how the underlying Kubernetes system is configured.
- Scale Down Delay: Sets the length of time in seconds that the system waits before it removes executors when it detects idle executor groups. As with the Scale Up Delay setting, the time to auto-scale is affected by how the underlying Kubernetes system is configured.