Major Description
Maintenance Manager threads use memory in both read-side and write-side block decoders and encoders. In a server running a workload with a table with ~250 string columns, I see each thread using between 100-200MB of RAM. If a user isn't aware of this memory usage they may under-provision the heap.
We should probably reserve some amount of memory for the MM threads, potentially track them via a MemTracker, and warn if the total configured heap (cache + MM threads estimate) is too close to the soft limit, leaving no room for MRS/DMS.