David Greenberg created MESOS-2634:
--------------------------------------
Summary: Support late-bound subcontainerization/pods
Key: MESOS-2634
URL: https://issues.apache.org/jira/browse/MESOS-2634
Project: Mesos
Issue Type: Story
Components: containerization
Reporter: David Greenberg
The k8s pods abstraction is not the only way to run multiple containers on a
single host. In my use of Mesos, I have two use cases for running multiple
containers together:
1. I run a job that needs a "babysitter" process to run with it, such that it
can inspect the filesystem being written to by the primary job. The babysitter
verifies that the job is operating correctly, and takes automated
recovery/termination actions on failure.
2. I run a webserver that writes logs. I also run logstash in the same
container, in order to ship the logs to our main log storage system.
In both of these cases, I need to run common initialization code, and then I
start the various sub-tasks that need to be isolated. Rather than need my
framework's scheduler to be able to make decisions at launch time about how it
wants to carve up the process, I'd like my executor to be able to locally
create subgroups for CPU and Memory containerization, so that it can decide
on-the-fly whether to carve its container into smaller pieces.
I think that by triggering the splitting of containers in the executor, this
will enable more dynamic usage.
The downside to this is that the container splitting would have to be rebuilt
for every containerizer; the upside is the flexibility in creating temporary
subcontainers only when they're needed.
I think that this would only require for command-line utilities that understood
the Mesos environment (that the executor could shell out to), or perhaps it
would be done with an executor API call to the slave, to ensure the
subcontainers didn't break cleanup.
I've heard rumors that Aurora can do something like this; perhaps their
implementation could graduate to Mesos proper?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
