Gilbert Song created MESOS-6371:
-----------------------------------
Summary: Remove the 'recover()' interface in 'ContainerLogger'.
Key: MESOS-6371
URL: https://issues.apache.org/jira/browse/MESOS-6371
Project: Mesos
Issue Type: Improvement
Components: containerization
Reporter: Gilbert Song
Assignee: Gilbert Song
This issue arises from the nested container support in Mesos.
Currently, the container logger interface mainly contains `recover()` and
`prepare()` methods. The `prepare` will be called in containerizer::launch() to
launch a container, while `recover` will be called in containerizer::recover()
to recover containers. Both methods rely on 2 parameters: ExecutorInfo and
sandbox directory. The sandbox directory for nested containers can still be
passed to the logger. However, because of nested container support,
ExecutorInfo is no longer available for nested containers.
In logger prepare, the ExecutorInfo is used for deliver FrameworkID,
ExecutorID, and Label for custom metadata. In containerizer launch, we can
still pass the ExecutorInfo of a nested container's top level parent to the
logger, so that those information will not be lost.
In logger recover, since currently the logger is stateless, and most of the
logger modules are doing `noop` in logger::recover(). The recover interface
should exist together with `cleanup` method if the logger become stateful in
the future. To avoid adding tech debt in containerizer nested container
support, we should remove the `recover` in container logger for now (can add it
back together with `cleanup` in the future if the container logger become
stateful).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
