On Nov 16, 2011, at 3:02 PM, Doug Cutting wrote: > > > Another definition is that a major release permits incompatible changes, > either in APIs, wire-formats, on-disk formats, etc. This is more > objective measure. For example, one might in release X+1 deprecate > features of release X but still remain compatible with them, while in > X+2 we'd remove them. So every major release would make incompatible > changes, but only of things that had been deprecated two releases ago. > Often the reason for the incompatible changes is new primary APIs or > re-implementation of primary components, but those more subjective > measures would not be the justification for the major version, rather > any incompatible changes would.
This is mostly consistent with what is stated in wrt to API changes in HADOOP-5071 on "Hadoop Compatibility requirements" : https://issues.apache.org/jira/browse/HADOOP-5071. HADOOP-5071 was derived from a long series of email discussions and describes some of the subtle nuances for compatibility for API, on-disk format, wire etc. Some notes (see details there) 1) break in compatibility => major number change, but major number change does NOT => break in compatibility. 2) we routinely change on disk format on hdfs but do an automatic upgrade. That is okay and allowed without a major number change. > > Of course, we should work hard to never make incompatible changes... Agreed. Once things are in customer hands it is very hard to remove even deprecated methods (But once in a while we have to have to do it after sufficient time to upgrade to new APIs). Java for example does not remove deprecated methods. sanjay
