Difference between revisions of "Sahara/WhyNotHeat"

Revision as of 13:30, 30 July 2013

1. The first question is “Why doesn’t Savanna use Heat to provision VMs?”

Generally using Heat underneath for infrastructure provisioning looks reasonable. In a tactic perspective there are few factors making Heat usage underneath Savanna problematic:

• Heat stability for Grizzly release. Savanna currently maintains

Grizzly+ compatibility.

• Installation of large Hadoop clusters (100+ nodes). Will be

addressed by proposed architecture changes.

• Anti-affinity support for HDFS redundancy in cloud environment
• Circular dependencies - we should generate ‘/etc/hosts’ for all

instances in provisioned cluster. We can’t use cloud init for this directly. There are a couple possible solutions using Heat, but none of them looks like a straightforward solution.

• Level of complexity. We try to keep things as simple as possible.

Adding extra layer will increase overall complexity of the solution. In addition both Savanna and Heat under active development changing lots of internals and even APIs and will require extra effort to coordinate.

Once Heat fulfills all the requirements we will be able and should use Heat for VM provisioning.

2. Let’s answer the second question - why we need Savanna? Can’t we use Heat to do what Savanna does?

• Savanna provides bunch of Hadoop-specific features. It’ll be hard to

provide them as Heat plugin

• Savanna provides Hadoop-specific APIs and functionality. Heat use

cases are mostly around provisioning/deployment.

• Savanna provides integration with various Hadoop distributions

through pluggable mechanism

Now, more details on each item. Hadoop specific features:

• Tight Swift integration. Hadoop can read and write from/to Swift

object storage. Savanna provides required configs for the Hadoop cluster.

• Usage of anti-affinity to preserve data-redundancy of HDFS nodes

Hadoop-specific APIs and functionality:

• Hadoop cluster scaling
• Elastic Data Processing: https://wiki.openstack.org/wiki/Savanna/EDP

Integration with Hadoop distributions through pluggable mechanism: - Usually Hadoop cluster deployment is a multi-step operation. First step is to install management console (for instance Apache Ambari). Second step is to communicate with management console through REST API to provision Hadoop on the cluster. Savanna wraps all this operations under well-defined API.

I hope all the items above explain why we need Savanna as a separate OpenStack service.

3. Why can’t Savanna be used as a plugin for Heat?

It should be and it will be someday.