I would prefer the option 2 suggested by Andrew, SFT backported to 2.5 as
experimental. Regarding the original "thin client/server master do all the
fix " approach for hbck2 mentioned by Duo, that has already been relaxed to
include some logic containing functions for some break scenarios where
master didn't have a solution implemented yet
(see RegionInfoMismatchTool, FsRegionsMetaRecoverer,
MissingTableDescriptorGenerator,
for example). Clusters facing such issues would then require a whole hbase
upgrade to a version including the fix logic, which is not feasible when
these are production deployments.
