+1 I remember Andrew saying same thing about CoprocessorEnvironment#getTable functions during the review of HBASE-17732, that we should get rid of them and let CP do their own table resource management like all other things.
On Wed, Oct 18, 2017 at 2:45 PM, Stack <st...@duboce.net> wrote: > tl;dr I believe the original intent for TableWrapper and > CoprocessorHConnection can now be gotten elsewhere so we should purge these > classes. > > In base CoprocessorEnvironment, there are methods to return a Table > instance. There are none to return an Admin. In our code base, the only > user is the AccessController. Phoenix uses it twice in its Indexer > implementation. > > When you call CE#getTable, the default implementation in > BaseEnvironemnt calls HTableWrapper.createWrapper which takes a > BaseEnvironment List in which we keep all Table instances. On shutdown of > the coprocessor, the list is iterated and all tables are closed. > > Going via TableWrapper, the class comment says: > > * A wrapper for HTable. Can be used to restrict privilege. > * > * Currently it just helps to track tables opened by a Coprocessor and > * facilitate close of them if it is aborted. > * > * We also disallow row locking. > * > * There is nothing now that will stop a coprocessor from using HTable > * objects directly instead of this API, but in the future we intend to > * analyze coprocessor implementations as they are loaded and reject those > * which attempt to use objects and methods outside the Environment > * sandbox. > > > TableWrapper by my reading delegates all calls to a Table instance with no > interception (there is not rowlocking to override anymore). On open, we do > ensure the Table is up on a CoprocessorHConnection which does the > following: > > * Connection to an HTable from within a Coprocessor. We can do some > nice tricks since we know we > * are on a regionserver, for instance skipping the full > serialization/deserialization of objects > * when talking to the server. > > > The above 'trick' is now commonplace in servers as the Master and > RegionServer always make Connections that will short-circuit if an > opportunity. > > As I read TableWrapper and CoprocessorHConnection, they were written at > another time when Table construction was heavyweight and Table#close was > not expected of clients and before the introduction of the general > Server-side short-circuit Connection facility. > > Unless objection, I think we should purge them. > > Writing here in case I'm missing some key facility they provide. > > Thanks, > St.Ack > -- -- Appy
