Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
In V3 format we read the whole blocklet at once to memory in order save IO time. But it turns out to be costlier in case of parallel reading of more carbondata files.
For example if we need to compact 50 segments then compactor need to open the readers on all the 50 segments to do merge sort. But the memory consumption is too high if each reader reads whole blocklet to the memory and there is high chances of going out of memory.
Solution:
In this type of scenarios we can introduce new readers for V3 to read the data page by page instead of reading whole blocklet at once to reduce the memory footprint.
