#mcollective

should i copy agents files (.rb/.ddl) into both client and server agent/application folders ?

or i just need to copy in to client

mcollective appears to use hostname for "identity". Is there away to set the identity explcitly?

The folks that build/manage our systems are not very consistent in their naming-style.

identity = foo

in the server.cfg file I presume. Will that have any impact on existing certificates?

seems like mcollective is similiar to fabric, however it says its not a fancy ssh forloop

"it uses publish subscribe middleware" and "real time discovery of network resources"

hmm not sure what that means

is mcollective basically a non-ssh version of server orchestration utility

basically a fabric with using its own client/server, rather than ssh and sshd

brucelee: the best I can do is explain the pub/sub and discovery bits

basically, with mco, you install the "server" on each box you want to orchestrate, and the "client" on the box you'll be orchestrating from - the pub/sub middleware is a message queue that the client uses to make requests of the servers

so when you use the mco client to orchestrate something, it puts a message on the MQ - usually you'll constrain your actions with metadata, like facts - this is part of the message

so if i want to ping all the servers

how does the message queue work in that scenario?

the servers that are subscribed to that queue will receive the message, check for validity (this layer is pluggable, as are most)

now heres my "why" question... why do this versus pinging every server directly?

well, let me go back

from a scaling stand point theres two ways to go about "nodes"

there could be a list of nodes... this list is actively updated

or there could be nodes that are brought up with a specific "client"

walk up, hang tight

i guess this is the "autodiscovery" mechanims"

but lets say the service discovery part is not needed, theres other tools you have which tells you what nodes are on your network

then you can just go directly to each node and query for what you need

how does mcollective's pub/sub way work, and whats the advantages to that?

ok, so let me catch up on all of that - had a coworker walk up and ask a ?

k

brucelee: so, let's say that you don't use pubsub. That's going to mean opening a direct connection to each server that you want to orchestrate, right?

yeah

ie if your orchestration piece is managing 1000 nodes, it needs to use 1000 network connections to do so, correct? (Sure, it may batch, etc, but overall, that's what's going to happen)

so think through that for a moment, and realize that in the pubsub model, the orchestrator says what it wants, and nodes receive the message, and act on it if it applies to them

so what this comes down to, is about a pardigm shift (IMO)

You ditch the need for a source of truth - eligible nodes simply do the work

at the same time, you cut down on network traffic, which can slow you down quite a bit depending on rtt and such

is there a trick to restarting mcollective using mco? i tried it with the mco service command, but the daemon only stopped and didn't start again

shafer: it's not a thing that is generally very doable

mco waits for the response for a service restart

and of course the bounced mco wouldn't know to send it even if it /did/ come back up

that said, I think service restarts are implemented as stop; start

and well, once mco stops, you can't really tell it to start itself

I'm sure that's not a very satisfactory answer, but do you follow what I"m saying?

torrancew: i do

k

it doesn't issue a restart arg to the script so it doesn't finish

As I understand it

i expected a report of failure back to me after calling it, but i was hoping it had succeeded

but it's been about a year since I dug into that

you using the service agent?

yes

i can dig in to the code when i have some free time to find an answer. just thought maybe it was something someone had figured out already

i can use ansible to do it on the off chance i need it

it ultimately defers to puppet to restart services

which means it'll do what the puppet provider on that box does

torrancew: interesting, i wouldn't have guessed that

ya

alot of the PL agents defer to puppet to do the work

gets cross-platform compatibility for basically free

makes sense, then it doesn't have to implement all of the service abstraction stuff for itself

right