mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Neil Conway <neil.con...@gmail.com>
Subject Re: Review Request 53897: Changed how master represents "recovered" frameworks.
Date Wed, 07 Dec 2016 05:50:05 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/53897/
-----------------------------------------------------------

(Updated Dec. 7, 2016, 5:50 a.m.)


Review request for mesos and Vinod Kone.


Changes
-------

Rebase.


Bugs: MESOS-6419
    https://issues.apache.org/jira/browse/MESOS-6419


Repository: mesos


Description
-------

After master failover, the new master doesn't know which frameworks were
registered with the previous master (because this information is not
currently stored in the registry). In the period after the master fails
over but before the framework scheduler has re-registered, the master
learns about the frameworks in the cluster when agents re-register (an
agent reports the FrameworkInfo for all of the frameworks it is running
when it re-registers).

Such frameworks were previously represented separately from the normal
list of frameworks in the master: the master kept a collection of
`FrameworkInfo` for these "recovered" frameworks.

This commit removes this separate collection of "recovered" frameworks.
Instead, the master now treats recovered frameworks very similarly to
frameworks that are registered but currently disconnected. For example,
recovered frameworks will now have a `Framework` object which tracks the
tasks/executors running under that framework; recovered frameworks will
be reported via the normal "frameworks" key when querying HTTP
endpoints. Similarly, "teardown" operations on recovered frameworks will
now work correctly (MESOS-6419).

This means there is no longer a concept of "orphan tasks" [1]: if the
master knows about a task, the task will be running under a framework
(albeit the framework might be recovered or disconnected). A new
"recovered" key has been added to various HTTP endpoints/APIs to
determine if a framework hasn't yet re-registered after master failover.

Another behavioral change is that, previously, a framework
re-registering after master failover was allowed to make arbitrary
changes to its `FrameworkInfo`. The master now only applies updates to
FrameworkInfo fields that can be safely changed (as was already done
when a framework fails over to a master that hasn't failed over). The
previous behavior qualifies as a bug.

[1] The exception here is if the cluster contains Mesos agents older
than 1.0, because old Mesos agents don't report `FrameworkInfo`s when
they re-register.


Diffs (updated)
-----

  include/mesos/master/master.proto 966105cf14bf92ad15a38653cd5c7f3322821289 
  include/mesos/v1/master/master.proto 0e5a8674a76a46c6f2cc17e11cd735f756d94fd1 
  src/master/http.cpp ac560d1fdd219d0de0c5d987a32a7112e149602f 
  src/master/master.hpp b444b2352360fb4f7179acd97dffc0cd81cc7afa 
  src/master/master.cpp 67f32229470da4cf7953881d1c5dcb99393002de 
  src/tests/api_tests.cpp 7c70f284a9225ca23abc7049f48f3efba314b641 
  src/tests/fault_tolerance_tests.cpp 59f68f65fac9fca4a6941793b712bfe7bb30c880 
  src/tests/master_allocator_tests.cpp bb94e38d5bb472801366c172cfc036f2eecdcbcb 
  src/tests/master_authorization_tests.cpp 06c6e47250003cd467bf37e1daa8bd636ef8fef5 
  src/tests/master_tests.cpp dfedbbdf78e8054813872e9eeebccc7504097751 
  src/tests/partition_tests.cpp 5a0d4bd2de6a5aa0e9fdf0d34cd10d16fd4e34a1 
  src/tests/teardown_tests.cpp 125e1808e51da53fdf1bb1babc956ff062cbb102 

Diff: https://reviews.apache.org/r/53897/diff/


Testing
-------

`make check`


Thanks,

Neil Conway


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message