Maximum length of the scan range. Interacts with the number of HDFS blocks in the table to determine how many CPU cores across the cluster are involved with the processing for a query. (Each core processes one scan range.)
Lowering the value can sometimes increase parallelism if you have unused CPU capacity, but a too-small value can limit query performance because each scan range involves extra overhead.
Only applicable to HDFS tables. Has no effect on Parquet tables. Unspecified or 0 indicates backend default, which is the same as the HDFS block size for each table.
Although the scan range can be arbitrarily long, Impala internally uses an 8 MB read buffer so that it can query tables with huge block sizes without allocating equivalent blocks of memory.
|<< MAX_IO_BUFFERS||MEM_LIMIT >>|