The Theory behind AMP

Why AMP?

Building and deploying applications in the cloud computing era has changed many things. Provision a bare server just-in-time, and use automated tools to install an application. Use APIs to add the server to a load balancer. When load goes up, provision another server; when load drops, kill a server to save money.

Many new tools have appeared that take advantage of this new era. However each of them only solve part of the problem and don’t consider the big picture. For example, configuration management tools such as Chef can, in a single command, provision a new cloud server then install and configure an application – but they require extra programming to reconfigure an load balancer whenever the pool of web servers changes. Amazon Auto Scaling can provision new servers and update load balancers, but it is dependent on CloudWatch – this means either using proxy metrics such as average response time, or writing more code to expose an application’s real metrics. A dedicated monitoring tool may be able to easily monitor the key metrics with little effort, but its alerts will need to be integrated it into the server provisioning process.

So all the tools are there to to create and manage a cloud-scale application that can adapt to demand to meet user expectations without wasting money on superfluous services - but you will need several such tools and it is up to you to integrate them into your deployment plan. Some of these tools – such as the Amazon Web Services web of EC2, CloudWatch, AutoScaling and CloudFormation – mean that you may suffer from lock-in. Related projects in OpenStack (Heat, Ceilometer, Murano, Solum, etc) provide similar functionality but again for a restricted target. The most common policies (such as minimising request latency) may be easy, but less common policies such as follow-the-sun and follow-the-moon may be up to you to implement. Your scaling policies may understand that “high demand = add another server”, but may not understand requirements such as some clustered services requiring an odd number of instances to prevent voting deadlocks.

How AMP Can Help

In this context the advantage of AMP becomes apparent: a single tool is able to manage provisioning and application deployment, monitor an application’s health and metrics, understand the dependencies between components (such as knowing that adding a new web server means that the load balancer needs reconfiguration) and apply complex policies to manage the application. The tool provides a REST API and a GUI, and allows the autonomic blueprints to be treated as an integral part of the application. With AMP, these policies become modular components which can be reused and easily added to blueprints.

AMP is about deploying and managing applications: composing a full stack for an application; deploying to cloud and non-cloud targets; using monitoring tools to collect key health/performance metrics; responding to situations such as a failing node; and adding or removing capacity to match demand.

Blueprints

An AMP blueprint defines an application, using a declarative YAML syntax supporting JVM plugins. A basic blueprint might comprise a single process, such as a web-application server running a WAR file or a SQL database and its associated DDL scripts. More complex blueprints encompass combinations of processes across multiple machines and services, such as a load-balancing HTTP server or SDN controller fronting a cluster of J2EE application servers, in turn connected to a resilient cluster of SQL database servers. Even larger clustered application running in multiple regions can be described, with features such as message buses with resilient brokers, cacheing tiers of NoSQL key-value store servers, a high-availability database cluster and multiple application components connected across these layers.

One main benefit of these blueprints is that they are composable: best-practice blueprints for one process or pattern (e.g. a Cassandra cluster) can be incorporated in other blueprints (e.g. an application with a Cassandra cluster as one component). Another major benefit is that the blueprints can be treated as source code as part of an applications codebase: tested, tracked, versioned, and hardened as an integral part of the devops process. In some ways, AMP is to run-time what Maven is to build-time.

Blueprints Turn into Deployments

AMP knows about Chef, Salt, and similar tools, and APT and Yum and plain old shell scripts, for deploying application components. Blueprints are built from a mixture of both off-the-shelf packages such as Tomcat, MySQL, Cassandra, and many more from our library; and components that are bespoke to individual applications; together with policies that allow the application to be self-managing.

AMP is built for the cloud, and will take a blueprint and deploy it to one of many supported clouds or even to multiple different clouds, or to private infrastructure (bring-your-own-node), or to other platforms. It will dynamically configure and connect all the different components of an application, e.g. so load balancers know where the web servers are and the web applications know where the database is.

AMP collects key metrics to monitor the health of applications; for example, by sending a request and measuring latency, or installing monitoring tools and using those to read a server’s management interface to determine the request queue length. These metrics can be fed into policies, which automatically take actions such as restarting a failed node, or scaling out the web tier if user demand exceeds capacity. This allows an application to be self-managing: to recover itself from simple failures, to scale out when demand increases and meet capacity; then scale in as demand drops and stop paying for spare capacity.

In short, AMP blueprints allow the best practices for deploying and managing complex software to be codified as part of the software development process.

Agile and Flexible

AMP is a product built from the ground up for application agility. This includes portability across non-cloud, cloud, and PaaS targets; devops-style infrastructure-as-code applied to applications; and real-time autonomic management based on promise theory. Some introductions to these concepts, associated tools, and open specifications may be useful.

Cloud computing at its core is about provisioning resources on-demand. The most widely known aspect is IaaS (infrastructure-as-a-service) such as Amazon EC2, Softlayer, Google Cloud Platform, Apache CloudStack, or OpenStack. By leveraging the Apache jclouds project (and contributing heavily to it), the AMP project is able to work with a large number of such providers. Higher up the stack, however, there is an increasingly diverse set of platform targets, from PaaS (platform-as-a-service) such as Cloud Foundry and Apache Stratos, through to myriad containers and runtime fabrics such as LXC/Docker, Apache Mesos, Apache Karaf, Apache Hadoop, and Apache Usergrid and other backend-as-a-service environments. AMP is based on the premise that applications may need to run in any or all of these, and the model must be flexible and open enough to support this.

The buzzword-compliant trends of agile and devops have reinforced many important lessons:

• The truth is in the code (not any ancillary documents)
• If it isn’t tested then assume it isn’t working
• Toolchain integration and APIs are key to a project’s success
• Even more critical is empowering all stakeholders to a project
• AMP’s focus on blueprinting and modeling as code and APIs serves these principles.

Autonomic Computing

Another major influence on the design of AMP are the ideas of autonomic computing and promise theory. It is not necessary to have a thorough understanding of these to use AMP, but contributors tend to become versed in these ideas quickly. Essentially, autonomics is based on the concept of components looking after themselves where possible (self-healing, self-optimizing, etc), and exposing a sensor (data outputs) and effector (operations) API where they may need to controlled by another element. Promise theory extends this approach by introducing the idea that communicating intent (through promises) is a more reliable basis for complex cooperating systems than obligation-based effectors. Tools such as CF Engine, Chef, Puppet, Ansible, and Salt apply promise theory to files and processes on machines; AMP can leverage all of these tools, complementing it with an application-oriented model.

Standards

Finally we note some emerging standards in this area. OASIS CAMP (Cloud Application Management for Platforms) and TOSCA (Topology and Orchestration Specification for Cloud Applications) both define YAML application models similar to AMP’s. CAMP focuses on the REST API for interacting with such a management layer, and TOSCA focuses on declarative support for more sophisticated orchestration. Currently AMP uses a YAML which complies with CAMP’s syntax and exposes many of the CAMP REST API endpoints. We would like to support the hot-off-the-press TOSCA YAML and expand the CAMP REST API coverage.