Thanks to Lars' local transactions patch (
HBASE-5229), we can entertain the possibility of doing local transactions within META regions.
We need this mainly for region splits and merges. Clients scan the META concurrent to the split/merge operations, and to prevent the clients from seeing overlapping region boundaries or holes in META, we just through hoops. For more backgroun, see BlockingMetaScannerVisitor,
HBASE-5986, and my comments at https://reviews.apache.org/r/8716/.
Now, for the actual implementation options:
1. As outlined in http://hadoop-hbase.blogspot.com/2012_02_01_archive.html,
- We have to implement a Custom RegionSplitPolicy for the META regions to ensure that a table's regions are always co-located in the same META region. Then we can add MultiRowMutationEndpoint as a system level coprocessor, and use it for META operations.
2. Do smt like HBASE-7716, and expose local atomic multi-row operations as a native API.
3. Move META to zookeeper. Use zookeeper.multi.
Then we can change region split / merge logic to make use of atomic META operations.