Details
-
Improvement
-
Status: Resolved
-
Critical
-
Resolution: Fixed
-
1.4.0
Description
With the advent of the HA NameNode in HDFS, the need for planned Hadoop outages has greatly diminished. I believe that Accumulo should follow suit and eventually allow for no-outage upgrades.
To support this feature, we would require wire compatibility between different versions so that the cluster can continue to function in a mixed state. There are already contingencies for master fail-over and tservers to leave/join the cluster, so I think this is a natural extension of that code.
To accomplish this for bugfix releases should be trivial. I think it is only necessary to remove (or relax) the version check that happens when a tserver attempts to join. There are rarely (never?) changes to the thrift protocol between bugfix releases, so theoretically a 1.4.1 tserver is already compatible with a 1.4.0 master and vice versa.
To maintain compatibility in the face of thrift protocol changes is certainly a harder challenge, which might be best reserved for a separate ticket.
Attachments
Issue Links
- is duplicated by
-
ACCUMULO-662 Accumulo instance with different versions not starting up
- Resolved