Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
The current (full) scan on LSM components is not very efficient, especially on hard disks, in two aspects:
- We often need to use a priority queue to merge results from multiple components. However, we only read a page at a time, which incurs a lot of random I/O overhead on hard disks.
- Full scan can often fill up (and clean up) the buffer cache. This problem is especially notable when we do merge. After a merge operation, the buffer cache would be filled up pages of old components, which would not be accessed by future queries.