Opinionated / sensible / recommended default tools for charms

[Kapil Thangavelu]

Excerpts from Adam Gandelman's message of Tue Nov 08 14:37:33 -0500 2011:
> On 11/08/2011 08:29 AM, Michael Nelson wrote:
> > On Tue, Nov 1, 2011 at 5:29 PM, Michael Nelson
> > <michael.nelson at canonical.com>  wrote:
> >> Just for anyone interested, here's part 1 of my juju+puppet experiment
> >> (with a 2min demo):
> >>
> >> http://micknelson.wordpress.com/2011/11/08/experimenting-with-juju-and-puppet/
> >>
> >> It's mostly pretty obvious - I just use the install/config-changed
> >> hooks to defer to puppet... but it helped me learn much more about
> >> puppet itself and how I can use it together with juju, as well as
> >> start thinking how I could provide an *optional* puppet-master (given
> >> that this will be nearly always private).
> >>
> >> -Michael
> >>
> 
> * WARNING: this ended up much longer than I had anticipated.   *
> 
> Michael-
> 
> Cool!  I'm glad there are other people scratching their heads about this 
> one.  I've played around with a few different ways of doing this (mostly 
> exactly like yours) and ran into a couple of concerns.
> 
> 1. Stuffing puppet manifests into charms and calling them in hooks via 
> 'puppet apply' loses the ability to maintain and share puppet types and 
> resources.  Unless you've set a constraint to only depend on puppet 
> primitives built into puppet-core, you cannot make use of Puppet's 
> shared name space for modules, types and resources(one of the features 
> that makes Puppet so powerful)  For example, lets say I've put together 
> a Puppet class that incorporates other custom types, providers and 
> modules.  When instantiated with a couple of parameters, it propagates a 
> huge chunk of system configuration that would otherwise be a nightmare 
> to deploy using shell scripts.   Now I want to make use of that class 
> from manifests contained within a Juju charm.  There is no way to do 
> that unless I also ship the custom module and all of its custom 
> dependencies within that charm alongside the manifests that called via 
> 'puppet apply'  If I want to use this in multiple juju charms, I need to 
> include the module and its dependencies in every Juju charm that uses 
> it.  This quickly becomes a nightmare to maintain and impossible to 
> avoid divergence.
> 
> 2. Calling 'puppet apply' at hook execution only enforces that 
> configuration when the relation's state changes. Ie, I've deployed our 
> massive application stack on a Friday and gone on vacation for 2 weeks. 
> On Monday  Jr. SysAdmin deleted the wrong config file on our web 
> server.  With a puppetmaster in place, his mistake is corrected in ~30 
> minutes (or the next time the agent checks into the puppet master).  
> Without, I'd have to remove and re-add Juju relations to resync system 
> configuration with the Juju's view of the world.

I think the juju world view (ie. service level) would be the machine is 
transient, if the service on the unit isn't functional anymore, spin up a new 
unit. That sort of presupposes a monitoring co-location charm and 
notification/event processing trigger against the juju api to allow for 
dynamically responding to the outage event, but that's the road we want to march 
down.

> 
> I'm not sure what the solution is. I've got a few ideas, but they all 
> mostly end up reproducing work across multiple domains.  What I 
> initially brainstormed:
> 
> a. All data gathered from config-get is stuffed custom facts which are 
> persistent and updated every time the corresponding hooks are executed 
> (ie, $juju-relation-db_user, $juju-relation-db_passwd)  Hooks do the 
> bare minimum configuration necessary to get the service up and running 
> (seems to be the story of most charms currently)

I'd rather this was out of band of juju core. A colo puppet could cron the 
inspection of service config and service relation data into facts prior to the 
puppet agent run to get this information.

> b. Units also get a custom fact that describes their service-unit, to be 
> used for node classification.
> c. The charm (or, possibly, a colocated puppet-agent charm), 
> authenticates the agent with a central puppet master and runs the puppet 
> agent in a typical polling configuration.

better is the colo puppet-agent charm, else we have charms assuming a puppet 
master.

> d. The puppet master classifies nodes based on service unit, and 
> enforces the basic configuration (what is stored in facts) as well as 
> any other environment-specific configuration admins may want.
> 
> This configuration ends up being maintained over time until the relation 
> changes, at which point the facts are updated.  The corresponding puppet 
> code on the master may be something as simple as using Augeas to manage 
> individual settings in a config file, or some more involved resources 
> that deploy large chunks of config that are too cumbersome to express 
> and maintain in the juju charm.

If the charm isn't functional without a puppet master setup, then its probably 
broken imo unless it explicitly declares a puppet master relation dependency.

> 
> However, saving information about the environment on disk (like relation 
> data) is a bad idea.  It would be much better if there were some way for 
> units to access that data outside of hook execution, so that puppet 
> could inspect it client-side dynamically via custom providers 
> (eliminating point a above)  It would also be great if Juju could be 
> used as an external node classifier (eliminating b above).   Then, we're 
> no longer coupled with Puppet and facter and the remaining items (c. and 
> d.) could be expressed in Chef-speak just as easily.

i don't see the problem with using local disk for state, but a cli api to 
inspect the juju unit data should suffice for the extraction use case. For non 
relation data this is already available, ie. config-get works outside of unit 
execution if you know the magic args.

an external node classifier to me sounds pretty reasonable, but the question is 
does that requires it to map to already extant definitions in the config 
managmenet system or can it define new roles/classes?

> 
> I think its important to think about the problem in more general terms.  
> To me, this isn't really about "how can we integrate this with Puppet," 
> but "how does configuration management fit into the Juju stack."  To 
> many people, there is a gap that needs to be filled before they'll 
> consider Juju.  Personally, I'd love to see Juju driving configuration 
> management (chef, puppet, whatever) similar to how it drives machine 
> providers (ec2, orchestra, whatever) There was some talk about this 
> topic at UDS and I expect there will be more as Juju gains a wider 
> audience.


In addition to populating the config management system with juju data, A 
co-located charm on the puppet or chef nodes that drives their respective apis 
to create classes/roles might suffice, but as i noted elsewhere in the thread 
juju and most configuration management systems have very different models. 

Clint's concern that we're just reintroducing the old paradigm on the new 
mechanism, without understanding the new paradigm rings true. The key aspect to 
an integration is preserving the functionality of both, which for 
juju is orchestration and encapsulation/reuse.

cheers,

Kapil