----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/19383/ -----------------------------------------------------------
(Updated March 21, 2014, 12:56 a.m.) Review request for mesos, Benjamin Hindman and Vinod Kone. Changes ------- No functional change, refactored Promise to Future and updated comments and logging messages based on Vinod's feedback. Removed dependent reviews. Bugs: MESOS-764 https://issues.apache.org/jira/browse/MESOS-764 Repository: mesos-git Description ------- This implements the Registry-backed Master, with the following exceptions that will be addressed in follow up changes: -Note that the --registry_strict flag is enforced to be false in master/main.cpp. -Reconciliation remains unimplemented as before. -Improvements can be made to killTask, specifically we should add SlaveID to the message in order to drop fewer requests for unknown slaves. -Orthogonally, this does not address MESOS-682. I've updated 'deactivated' slaves to be a cache of SlaveIDs rather than UPIDs as this was the intent originally (we were concerned about the unbounded growth of the set, but cache<SlaveID, Nothing> keeps a fixed capacity). Diffs (updated) ----- src/master/constants.hpp cdaaad060d4ee777f8b0838b63c0fd031da861ea src/master/constants.cpp 18548834468243bef8ae090f70363e2b9f571ac5 src/master/master.hpp a37a2a2fa3713b8251eb318dfb45e0793cb344ff src/master/master.cpp 3b28f72be2016997e30649d2b12ed0e8f1a57dd5 src/messages/messages.proto c26a3d0e69bbbd447c859cf175c139ab8948fde2 src/slave/slave.cpp d8d3e0fa54972201d72b2650ec0ba922a4912d54 Diff: https://reviews.apache.org/r/19383/diff/ Testing ------- This change preserves the previous semantics and so all existing tests pass. This is because the Registrar can only operate in a "non-strict" manner. An unfortunate effect of this change is that many tests run slower due to the fact that messages are dropped while we're recovering, an alternative approach here would be to re-enqueue *all* incoming messages through recover(). However, this adds queuing delay to each message processed in the Master and the performance implications of this are not well understood for large clusters. Thanks, Ben Mahler