Re: Questions on LRPs and container networking

Amit Kumar Gupta

1. It gives it an identifier that it looks up in the BBS, so it's decoupled
from directly knowing about the droplet blobstore. When it gets the Task
or LRP definition from the blobstore, that will include a set of Actions to
run, which include a Download action to retrieve something from a URL. The
URL is provided by CC to the CC Bridge, and then CC Bridge puts that URL in
an appropriate Download step in the Actions it attaches to the LRP or Task
that it requests from Diego.

2. Each Cell has a colocated Garden, yes.

3. The Rep updates the BBS with the state of its running containers. The
route-emitter is a client of the BBS, and determines what's routable based
on that information, and emits messages (currently over NATS) to tell the
router what to do with its routing table.

4. There's all sorts of stuff. If the route is still in the route table,
the router will route to it, but if it fails to make a connection it's
smart enough to try another IP:port for that route. If a container dies,
the Rep notifies the BBS, which the route-emitter notices, and it updates
the router. If the process is alive but not responding, the
healthcheck/monitor co-process that runs with the container will notice,
and kill the container.


On Wed, Jul 20, 2016 at 9:12 AM, J K <falconwing(a)> wrote:

1.) After the auctioneer declares a winner for an auction, what does it
tell the rep that causes it to start doing the work? For example, does it
give it some kind of identifier that enables it to look up the appropriate
droplet in the blobstore?

2.) Does every cell that hosts application containers have its own
instance of Garden? When work is ready to be done by a cell, is it the rep
which tells Garden to make a container with a given droplet? (Are a cell's
own CPU, disk, and memory resources consumed during the process of building
a container?)

3.) After a cell has finished making a container, what happens to make
that container routable from the public Internet? What CF and cell
components does the request pass through or get inspected by before it
reaches the container?

4.) When a container dies, what stops the load balancer and/or router from
continuing to route requests to it? If a container is merely deadlocked in
some way but the process is still alive (just not responding), does the
answer change?

Join { to automatically receive all group messages.