Working with Roadmaps

[ English | Indonesia | Deutsch | 日本語 ]

Working with Roadmaps

The good news: OpenStack has unprecedented transparency when it comes to providing information about what’s coming up. The bad news: each release moves very quickly. The purpose of this appendix is to highlight some of the useful pages to track, and take an educated guess at what is coming up in the next release and perhaps further afield.

OpenStack follows a six month release cycle, typically releasing in April/May and October/November each year. At the start of each cycle, the community gathers in a single location for a Project Teams Gathering (PTG). At the PTG, the features for the coming releases are discussed, prioritized, and planned. The below figure shows an example release cycle, with dates showing milestone releases, code freeze, and string freeze dates, along with an example of when the summit occurs. Milestones are interim releases within the cycle that are available as packages for download and testing. Code freeze is putting a stop to adding new features to the release. String freeze is putting a stop to changing any strings within the source code.

_images/osog_ac01.png

Influencing the Roadmap

OpenStack truly welcomes your ideas (and contributions) and highly values feedback from real-world users of the software. By learning a little about the process that drives feature development, you can participate and perhaps get the additions you desire.

Feature requests typically start their life in Etherpad, a collaborative editing tool, which is used to take coordinating notes at a design summit session specific to the feature. This then leads to the creation of a blueprint on the Launchpad site for the particular project, which is used to describe the feature more formally. Blueprints are then approved by project team members, and development can begin.

Therefore, the fastest way to get your feature request up for consideration is to create an Etherpad with your ideas and propose a session to the PTG. If the PTG has already passed, you may also create a blueprint directly. Read this blog post about how to work with blueprints the perspective of Victoria Martínez, a developer intern.

The roadmap for the next release as it is developed can be seen at Releases.

To determine the potential features going in to future releases, or to look at features implemented previously, take a look at the existing blueprints such as OpenStack Compute (nova) Blueprints, OpenStack Identity (keystone) Blueprints, and release notes.

Aside from the direct-to-blueprint pathway, there is another very well-regarded mechanism to influence the development roadmap: the user survey. Found at OpenStack User Survey, it allows you to provide details of your deployments and needs, anonymously by default. Each cycle, the user committee analyzes the results and produces a report, including providing specific information to the technical committee and project team leads.

Aspects to Watch

You want to keep an eye on the areas improving within OpenStack. The best way to “watch” roadmaps for each project is to look at the blueprints that are being approved for work on milestone releases. You can also learn from PTL webinars that follow the OpenStack summits twice a year.

Driver Quality Improvements

A major quality push has occurred across drivers and plug-ins in Block Storage, Compute, and Networking. Particularly, developers of Compute and Networking drivers that require proprietary or hardware products are now required to provide an automated external testing system for use during the development process.

Easier Upgrades

One of the most requested features since OpenStack began (for components other than Object Storage, which tends to “just work”): easier upgrades. In all recent releases internal messaging communication is versioned, meaning services can theoretically drop back to backward-compatible behavior. This allows you to run later versions of some components, while keeping older versions of others.

In addition, database migrations are now tested with the Turbo Hipster tool. This tool tests database migration performance on copies of real-world user databases.

These changes have facilitated the first proper OpenStack upgrade guide, found in Upgrades, and will continue to improve in the next release.

Deprecation of Nova Network

With the introduction of the full software-defined networking stack provided by OpenStack Networking (neutron) in the Folsom release, development effort on the initial networking code that remains part of the Compute component has gradually lessened. While many still use nova-network in production, there has been a long-term plan to remove the code in favor of the more flexible and full-featured OpenStack Networking.

An attempt was made to deprecate nova-network during the Havana release, which was aborted due to the lack of equivalent functionality (such as the FlatDHCP multi-host high-availability mode mentioned in this guide), lack of a migration path between versions, insufficient testing, and simplicity when used for the more straightforward use cases nova-network traditionally supported. Though significant effort has been made to address these concerns, nova-network was not be deprecated in the Juno release. In addition, to a limited degree, patches to nova-network have again begin to be accepted, such as adding a per-network settings feature and SR-IOV support in Juno.

This leaves you with an important point of decision when designing your cloud. OpenStack Networking is robust enough to use with a small number of limitations (performance issues in some scenarios, only basic high availability of layer 3 systems) and provides many more features than nova-network. However, if you do not have the more complex use cases that can benefit from fuller software-defined networking capabilities, or are uncomfortable with the new concepts introduced, nova-network may continue to be a viable option for the next 12 months.

Similarly, if you have an existing cloud and are looking to upgrade from nova-network to OpenStack Networking, you should have the option to delay the upgrade for this period of time. However, each release of OpenStack brings significant new innovation, and regardless of your use of networking methodology, it is likely best to begin planning for an upgrade within a reasonable timeframe of each release.

As mentioned, there’s currently no way to cleanly migrate from nova-network to neutron. We recommend that you keep a migration in mind and what that process might involve for when a proper migration path is released.

Distributed Virtual Router

One of the long-time complaints surrounding OpenStack Networking was the lack of high availability for the layer 3 components. The Juno release introduced Distributed Virtual Router (DVR), which aims to solve this problem.

Early indications are that it does do this well for a base set of scenarios, such as using the ML2 plug-in with Open vSwitch, one flat external network and VXLAN tenant networks. However, it does appear that there are problems with the use of VLANs, IPv6, Floating IPs, high north-south traffic scenarios and large numbers of compute nodes. It is expected these will improve significantly with the next release, but bug reports on specific issues are highly desirable.

Replacement of Open vSwitch Plug-in with Modular Layer 2

The Modular Layer 2 plug-in is a framework allowing OpenStack Networking to simultaneously utilize the variety of layer-2 networking technologies found in complex real-world data centers. It currently works with the existing Open vSwitch, Linux Bridge, and Hyper-V L2 agents and is intended to replace and deprecate the monolithic plug-ins associated with those L2 agents.

New API Versions

The third version of the Compute API was broadly discussed and worked on during the Havana and Icehouse release cycles. Current discussions indicate that the V2 API will remain for many releases, and the next iteration of the API will be denoted v2.1 and have similar properties to the existing v2.0, rather than an entirely new v3 API. This is a great time to evaluate all API and provide comments while the next generation APIs are being defined. A new working group was formed specifically to improve OpenStack APIs and create design guidelines, which you are welcome to join.

OpenStack on OpenStack (TripleO)

This project continues to improve and you may consider using it for greenfield deployments, though according to the latest user survey results it remains to see widespread uptake.

Data processing service for OpenStack (sahara)

A much-requested answer to big data problems, a dedicated team has been making solid progress on a Hadoop-as-a-Service project.

Bare metal Deployment (ironic)

The bare-metal deployment has been widely lauded, and development continues. The Juno release brought the OpenStack Bare metal drive into the Compute project, and it was aimed to deprecate the existing bare-metal driver in Kilo. If you are a current user of the bare metal driver, a particular blueprint to follow is Deprecate the bare metal driver

Database as a Service (trove)

The OpenStack community has had a database-as-a-service tool in development for some time, and we saw the first integrated release of it in Icehouse. From its release it was able to deploy database servers out of the box in a highly available way, initially supporting only MySQL. Juno introduced support for Mongo (including clustering), PostgreSQL and Couchbase, in addition to replication functionality for MySQL. In Kilo, more advanced clustering capability was delivered, in addition to better integration with other OpenStack components such as Networking.

Message Service (zaqar)

A service to provide queues of messages and notifications was released.

DNS service (designate)

A long requested service, to provide the ability to manipulate DNS entries associated with OpenStack resources has gathered a following. The designate project was also released.

Scheduler Improvements

Both Compute and Block Storage rely on schedulers to determine where to place virtual machines or volumes. In Havana, the Compute scheduler underwent significant improvement, while in Icehouse it was the scheduler in Block Storage that received a boost. Further down the track, an effort started this cycle that aims to create a holistic scheduler covering both will come to fruition. Some of the work that was done in Kilo can be found under the Gantt project.

Block Storage Improvements

Block Storage is considered a stable project, with wide uptake and a long track record of quality drivers. The team has discussed many areas of work at the summits, including better error reporting, automated discovery, and thin provisioning features.

Toward a Python SDK

Though many successfully use the various python-*client code as an effective SDK for interacting with OpenStack, consistency between the projects and documentation availability waxes and wanes. To combat this, an effort to improve the experience has started. Cross-project development efforts in OpenStack have a checkered history, such as the unified client project having several false starts. However, the early signs for the SDK project are promising, and we expect to see results during the Juno cycle.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.