Help Center/ MapReduce Service/ Component Operation Guide (ME-Abu Dhabi Region)/ Using CarbonData (for MRS 3.x or Later)/ CarbonData FAQ/ Why Does CarbonData Require Additional Executors Even Though the Parallelism Is Greater Than the Number of Blocks to Be Processed?
Updated on 2022-02-22 GMT+08:00
View PDF
Share

Why Does CarbonData Require Additional Executors Even Though the Parallelism Is Greater Than the Number of Blocks to Be Processed?

Question

Why does CarbonData require additional executors even though the parallelism is greater than the number of blocks to be processed?

Answer

CarbonData block distribution optimizes data processing as follows:

  1. Optimize data processing parallelism.
  2. Optimize parallel reading of block data.

To optimize parallel processing and parallel read, CarbonData requests executors based on the locality of blocks so that it can obtain executors on all nodes.

If you are using dynamic allocation, you need to configure the following properties:

  1. Set spark.dynamicAllocation.executorIdleTimeout to 15 minutes (or the average query time).
  2. Set spark.dynamicAllocation.maxExecutors correctly. The default value 2048 is not recommended. Otherwise, CarbonData will request the maximum number of executors.
  3. For a bigger cluster, set carbon.dynamicAllocation.schedulerTimeout to a value ranging from 10 to 15 seconds. The default value is 5 seconds.
  4. Set carbon.scheduler.minRegisteredResourcesRatio to a value ranging from 0.1 to 1.0. The default value is 0.8. Block distribution can be started as long as the value of carbon.scheduler.minRegisteredResourcesRatio is within the range.
Parent topic: CarbonData FAQ