Why plasma memory allocations are intentionally discontinuous to prevent dlmalloc to coalesce them?

Ray Core
1

I am learning the code in plasma of Ray. It uses dlmalloc as the default memory allocator and uses memory-mapping. There is a comment in the file malloc.h:

/// Gap between two consecutive mmap regions allocated by fake_mmap.
/// This ensures that the segments of memory returned by
/// fake_mmap are never contiguous and dlmalloc does not coalesce it
/// (in the client we cannot guarantee that these mmaps are contiguous).
constexpr int64_t kMmapRegionsGap = sizeof(size_t);

I wonder why memory allocations are intentionally discontinuous to prevent dlmalloc to coalesce them?
Thanks！！