It's observed that Velox backend uses more memory than we configured. Which is perhaps related to the untracked Velox global memory manager.

We should set a capacity according to Spark overhead size to here to limit that memory manager's usage.