meta table region assignment could reach to the 'FAILED_OPEN' state, which makes the region not available unless the target region server shutdown or manual resolution. This is undesirable state for meta tavle region.
Here is the sequence how this could happen (the code is in AssignmentManager#assign()):
Step 1: Master detects a region server (RS1) that hosts one meta table region is down, it changes the meta region state from 'online' to 'offline'
Step 2: In a loop (with configuable maximumAttempts count, default is 10, and minimal is 1), AssignmentManager tries to find a RS to host the meta table region. If there is no RS available, it would loop forver by resetting the loop count (BUG#1 from this logic - a small bug)
Step 3: Once a new RS is found (RS2), inside the same loop as Step 2, AssignmentManager tries to assign the meta region to RS2 (OFFLINE, RS1 => PENDING_OPEN, RS2). If for some reason that opening the region in RS2 failed (eg. the target RS2 is not ready to serve - ServerNotRunningYetException), AssignmentManager would change the state from (PENDING_OPEN, RS2) to (FAILED_OPEN, RS2). then it would retry (and even change the RS server to go to). The retry is up to maximumAttempts. Once the maximumAttempts is reached, the meta region will be in the 'FAILED_OPEN' state, unless either (1). RS2 shutdown to trigger region assignment again or (2). it is reassigned by an operator via HBase Shell.
Based on the document ( http://hbase.apache.org/book/regions.arch.html ), this is by design - "17. For regions in FAILED_OPEN or FAILED_CLOSE states , the master tries to close them again when they are reassigned by an operator via HBase Shell.".
However, this is bad design, espcially for meta table region (it is arguable that the design is good for regular table - for this ticket, I am more focus on fixing the meta region availablity issue).
I propose 2 possible fixes:
Fix#1 (band-aid change): in Step 3, just like Step 2, if the region is a meta table region, reset the loop count so that it would not leave the loop with meta table region in FAILED_OPEN state.
Fix#2 (more involved): if a region is in FAILED_OPEN state, we should provide a way to automatically trigger AssignmentManager::assign() after a short period of time (leaving any region in FAILED_OPEN state or other states like 'FAILED_CLOSE' is undesirable, should have some way to retrying and auto-heal the region).
I think at least for 1.0.0, Fix#1 is good enough. We can open a task-type of JIRA for Fix#2 in future release.