Skip to content

Latest commit

 

History

History
291 lines (191 loc) · 9.33 KB

FEATURES.rst

File metadata and controls

291 lines (191 loc) · 9.33 KB

Senlin Feature Request Pipeline

This document records the feature requests the developer team has received and considered. This document SHOULD NOT be treated as a replacement of the blueprints (or specs) which already accompanied with a design. The feature requests here are meant to be a pipeline for mid-term goals that Senlin should strive to achieve. Whenever a feature can be implemented with a practical design, the feature should be moved to a blueprint (and/or specs) review.

This document SHOULD NOT be treated as a replacement of the TODO file the development team is maintaining. The TODO file records actionable work items that can be picked up by any developer who is willing to do it, while this document records more general requirements that needs at least a draft design before being worked on.

High Priority

API versioning support

This support can help to improve the stability and compatibility of Senlin API interface. We can learn some experience from what Nova has done for providing API version control.

Container Clustering

[Austin Discussion: https://etherpad.openstack.org/p/newton-senlin-container] We need to add a simple/generic container profile (which can be rebased on some other projects later) so that we can experiment with issues unique to container cluster management:

  • Scheduling
  • Networking/Storage
  • APIs
  • Security issues
  • Dependencies

Event Listener

To make Senlin responsive to events published by other OpenStack services, an event subscriber is needed so that Senlin can receive notifications from sources such as Ceilometer, Nova, or Zaqar.

This is of a high priority because Senlin needs it as one of its HA solutions.

Scavenger Process

Senlin needs a scavenger process that runs as a background daemon. It is tasked with cleansing database for old data, e.g. event records. Its behavior must be customizable because users may want the old records to be removed or to be archived in a certain way.

Fault Tolerance

Senlin in most cases will be managing clusters with nodes distributed somewhere. One problems inherent to such a distributed architecture is about partial failures, communication latencies, concurrency, consistency etc. There are hardware/software failures expected. Senlin must remain operational in the face of such failures.

Access Permission Control

Currently, all access to Senlin objects like cluster, profile are project_safe by default. This is for preventing user manipulating resources belong to other users. However, sharing resource between different users/projects with limited privilege(e.g. read-only, read-write) is also a very reasonable demand in many cases. Therefore, we may need to provide access permission control in Senlin to support this kind of requirement.

Middle Priority

Support Scheduled Actions

This is a request to trigger some actions at a specified time. One typical use case is to scale up a cluster before weekend or promotion season as a preparation for the coming burst of workloads.

Blue-Green Deployment

Support to deploy environments using blue-green deployment pattern. http://martinfowler.com/bliki/BlueGreenDeployment.html

Multi-cloud Support

In some case, user could have the demand to create/scale cluster cross different clouds. Therefore, Senlin is supposed to have the ability to manage nodes which span cross multiple clouds within the same cluster. Support from both profile and policy layers are necessary for providing this ability.

User Defined Actions

Actions in Senlin are mostly built-in ones at present. There are requirements to incorporate Shell scripts and/or other structured software configuration tools into the whole picture. One of the option is to provide a easy way for Senlin to work with Ansible, for example.

Event Notification

Event notification is a feature that enables an external tool to subscribe to events sent from Senlin when interesting things happen. One option is to use the messaging service provided by the Zaqar project.

Customizable Batch Processing

An important non-functional requirement for Senlin is the scale of clusters it can handle. We will strive to make it handle large scale ones, however that indicates that we need to improve DB accesses in case of heavy loads. One potential tradeoff is to introduce an option for users to customize the size of batches when large number of DB requests pouring in.

Container/Docker support

Using Senlin to manage Container/Docker cluster is possible. We should provide related support for users who have this requirement.

Versioned Object Support

oslo.versionedobjects is a lib provides a generic versioned object model that is RPC-friendly, with inbuilt serialization, field typing, and remotable method calls. We can use it to promote the upgrading capatibility of Senlin service.

Low Priority

Define and Enforce Quotas

There is a potential request to limit how many clusters a user can create, how large a cluster can become.

Use Barbican to Store Secrets

Currently, Senlin uses the cryptography package for data encryption and decryption. There should be support for users to store credentials using the Barbican service, in addition to the current solution.

Use VPNaaS to Build Cross-Region/Cross-Cloud

When buidling clusters that span more than one region or cloud, there are requirements to place all cluster nodes on the same VPN so that workloads can be distributed to the nodes as if they sit on the same network.

Vertical Scaling

Though Senlin is mainly concerns about the horizontal scaling in/out support, there are possibilities/requirements to scale nodes in the vertical direction. Vertical scaling means automatically adding compute/storage/network resources to cluster nodes. Depending on the support from corresponding services, this could be explored.

Replace Green Threads with Python Threading

Senlin is now using green threads (eventlets) for async executions. The eventlets execution model is not making the use of multi-processing platforms in an efficient way. Senlin needs a scalable execution engine, so native multi-threading is needed.

Metrics Collection

Senlin needs to support metric collections about the clusters and nodes it manages. These metrics should be collectable by the ceilometer service, for example.

AWS Compatible API

There are requirements for Senlin to provide a AWS compatible API layer so that existing workloads can be deployed to Senlin and AWS without needing to change a lot of code or configurations.

Integration with Mistral

There are cases where the (automated) operations on clusters and nodes form a workflow. For example, an event triggers some actions to be executed in sequence and those actions in turn triggers other actions to be executed.

Support to Suspend/Resume Operations

A user may want to suspend/resume a cluster or an individual node. Senlin needs to provide a generic definition of 'suspend' and 'resume'. It needs to be aware of whether the profile and the driver support such operations.

Interaction with Congress

This is of low priority because Senlin needs a notification mechanism in place before it can talk to Congress. The reason to interact with Congress is that there could be enterprise level policy enforcement that Senlin has to comply to.

Investigation of Tooz

There is requirement to manage multiple senlin-engine instances in a distributed way. Or, we can use a variant of DLM to manage cluster membership. E.g. use redis/zookeeper to build clusters in their sense so that when the cluster membership changes, we may possibly receive a notification. This would be helpful for cluster health management.

Tooz is the promised focal point in this field, generalizing the many backends that we don't want to care about. This TODO item is about two things:

  1. Whether Tooz does provide a reliable membership management infra?
  2. Is there a comparison between zookeeper and redis for example.

Run Senlin API under Apache

Add support to have Senlin API run under Apache.

Baremetal Support

Managing baremetal cluster is a very common requirement from user. It is reasonable for Senlin to support it by talking with service like Ironic.

TOSCA support

Provide TOSCA support in Senlin(maybe reuse heat-translator/tosca-parser?)

Dynamic Plugin Loading

Design and implement dynamic plugin loading mechanism that allows loading plugins from any paths.

Scaling to Existing Nodes

[Conclusion from Austin: https://etherpad.openstack.org/p/newton-senlin-as]

Senlin can improve scale-out operation so that it can add existing nodes to a cluster when doing scale-out. We are not intended to scale to nodes not created by Senlin.