Kevin Klues created MESOS-7027:
----------------------------------
Summary: CommandExecutor ENV overwritten by Docker Image ENV in
Unified Containerizer
Key: MESOS-7027
URL: https://issues.apache.org/jira/browse/MESOS-7027
Project: Mesos
Issue Type: Bug
Reporter: Kevin Klues
Priority: Blocker
Using the unified containerizer, if a docker image is provisioned and has
environment variables set via the ENV directive, those environment variables
will be inherited by the {{mesos-executor}} process and overwrite similarly
named environment variables that otherwise would have been inherited from the
agent.
This causes problems (for example) in DC/OS when trying to launch tasks based
off the {{nvidia/cuda}} image. The {{nvidia/cuda}} image explicitly sets
{{LD_LIBRARY_PATH}} to its own value so that the proper nvidia libraries will
be available to whatever command is launched inside the container.
However, DC/OS relies on {{LD_LIBRARY_PATH}} to contain a path to
{{/opt/mesosphere/lib}} so that all of the mesosphere libraries are available
to the mesos binaries launched by the agent ({{mesos-containerizer}},
{{mesos-execute}}, etc.). This is necessary to make sure that any external
dependencies they might have (e.g. libssl.so) can be resolved at runtime.
By overwriting the executor's environment with the Docker Image environment,
{{LD_LIBRARY_PATH}} will not be set properly and {{mesos-execute}} will fail.
It seems to me, the Docker Image environment should *only* actually overwrite
the environment of the user process (not its executor). However, this can get
complicated, because the executor actually is the user process in the case of
launching a custom executor.
We need to rethink how the environment is inherited/overwritten through all the
various processes that get spawned while launching a container as well as how
to make it work for tasks launched by arbitrary executors.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
