Rapid deployment capability is table stakes when we are talking about a PaaS solution. Every vendor touts it, and to be frank, every user simply expects it to be there. While I think it is interesting to talk about rapid deployment and perhaps compare speed of one solution to that of another, I think it is infinitely more interesting to talk about the mechanics of deployment for a particular solution. That is, I think the more interesting and important question is ‘What deployment style does a particular solution take?’
At a very high, black and white level, I think two primary deployment styles permeate the landscape of PaaS today: contained and referential. I want to compare each approach, but before that, let me use a few words to describe each style.
– Contained: In the contained deployment model, PaaS solutions deploy environments based on packages that contain most, if not all, of the desired configuration as well as the logic to apply that configuration. For instance, if a solution were to deploy a virtual image in the contained model, the virtual machine would have the necessary information and logic embedded to configure itself upon start up. It would not necessarily need to contact external systems or wait for instructions from other actors.
– Referential: In the referential deployment model, PaaS solutions deploy environments using a minimal base package. At some point during the deployment process, the deployed environment communicates with a third party in some fashion to procure the necessary configuration information. Going back to the example above, if a virtual image were deployed in the referential model, the virtual machine would start up and then communicate with a third party service (either by initiating a request or waiting for instructions). This third party service would send down the configuration information and instructions for the environment hosted within the virtual machine.
When comparing the two approaches, it is helpful to understand the advantages and drawbacks of each. A closer look at the contained model reveals an obvious benefit: speed. In this model, the deployment package contains most of what it will require in order to fully configure itself. It does not rely on contacting an external service and having to pull down the necessary binaries and configuration information.
This advantage comes with an obvious drawback: management burden. By building more and more into the deployment package, you are increasing the amount of content that must be maintained and updated in said package. While it is not a huge concern if you only have a handful of discrete packages, you may not be able to rely on that luxury. You may expect that after some amount of time, the number of permutations to support necessitate spending an inordinate amount of time updating deployment packages. If this is the case, you can easily end up in a situation where the benefits of rapid deployment are negated by increased administrative costs.
The referential approach avoids the above pitfall. In this model, the deployment package remains fairly skeletal. Instead of packing in all of the content like in the contained model, the deployment packages in the referential model know just enough to integrate with an external system to get the right configuration information (think Chef and Puppet). This means that you only need to update and maintain configuration data and configuration actions in a single location instead of in each and every deployment package. As the number of different required environments increase, this approach can mean a significant reduction in management burden.
There is a flip side to this coin of course. The referential approach typically results in longer deployment times – dependent on the time required to install and configure content for your environments of course. Since the deployment packages contain very little content at deploy-time, they must pull or otherwise receive that data at some point during the deployment. This may or may not be a big issue for your particular use case, but it is a potential drawback worth considering.
So which approach is better? It is my opinion, one derived from numerous user experiences, that there is no way to generalize the answer to that question. In cases where content is infrequently updated and the number of environmental permutations is fairly well constrained, the contained deployment model can be extremely effective and efficient. On the other hand, in cases where content is dynamic and ever-changing, the referential deployment model is a virtual requirement. From a user’s standpoint, I strongly suggest pursuing solutions that support both kinds of deployment models. Tools should not dictate your approach. Your requirements for a particular usage scenario should!