Question
When Spark Streaming tasks are running, the data processing performance does not improve significantly as the number of executors increases. What should I pay attention to if I perform parameter optimization?
Answer
When the number of executor cores is 1, comply with the following rules to optimize Spark Streaming running parameters:
- The Spark task processing speed is related to the number of partitions in Kafka. When the number of partitions is less than the specified number of executors, the number of actually used executors is the same as the number of partitions, and other executors will be idle. Therefore, the number of executors must be less than or equal to the number of partitions.
- When data skew occurs on different partitions of Kafka, the executor corresponding to the partition with a large amount of data touches the glass ceiling of data processing. Therefore, when the Producer program is executed, data is sent to each partition on average to improve the processing speed.
- When partition data is evenly distributed, increasing the number of partitions and executors will improve the Spark processing speed. (When the number of partitions is the same as that of executors, the processing speed is the fastest.)
- When partition data is evenly distributed, ensure that the number of partitions is an integer multiple of the number of executors for proper allocation of resources.