Backwards compatibility and TripleO
TripleO has run with good but not perfect backwards compatibility since creation. It's time to formalise this in a documentable and testable fashion.
TripleO will follow Semantic Versioning (aka semver) for versioning all releases. We will strive to avoid breaking backwards compatibility at all, and if we have to it will be because of extenuating circumstances such as security fixes with no other way to fix things.
TripleO has historically run with an unspoken backwards compatibility policy but we now have too many people making changes - we need to build a single clear policy or else our contributors will have to rework things when one reviewer asks for backwards compat when they thought it was not needed (or vice versa do the work to be backwards compatible when it isn't needed.
Secondly, because we haven't marked any of our projects as 1.0.0 there is no way for users or developers to tell when and where backwards compatibility is needed / appropriate.
Adopt the following high level heuristics for identifying backwards incompatible changes:
- Making changes that break user code that scripts or uses a public interface.
- Becoming unable to install something we could previously.
- Being unable to install something because someone else has altered things -e.g. being unable to install F20 if it no longer exists on the internet is not an incompatible change - if it were returned to the net, we'd be able to install it again. If we remove the code to support this thing, then we're making an incompatible change. The one exception here is unsupported projects - e.g. unsupported releases of OpenStack, or Fedora, or Ubuntu. Because unsupported releases are security issues, and we expect most of our dependencies to do releases, and stop supporting things, we will not treat cleaning up code only needed to support such an unsupported release as backwards compatible. For instance, breaking the ability to deploy a previous still supported OpenStack release where we had previously been able to deploy it is a backwards incompatible change, but breaking the ability to deploy an unsupported OpenStack release is not.
Corollaries to these principles:
- Breaking a public API (network or Python). The public API of a project is any released API (e.g. not explicitly marked alpha/beta/rc) in a version that is >= 1.0.0. For Python projects, a _ prefix marks a namespace as non-public e.g. in
quux is not public because it's in a non-public namespace. For our projects that accept environment variables, if the variable is documented (in the README.md/user documentation) then the variable is part of the public interface. Otherwise it is not.
- Increasing the set of required parameters to Heat templates. This breaks scripts that use TripleO to deploy. Note that adding new parameters which need to be set when deploying new things is fine because the user is doing more than just pulling in updated code.
- Decreasing the set of accepted parameters to Heat templates. Likewise, this breaks scripts using the Heat templates to do deploys. If the parameters are no longer accepted because they are for no longer supported versions of OpenStack then that is covered by the carve-out above.
- Increasing the required metadata to use an element except when both Tuskar and tripleo-heat-templates have been updated to use it. There is a bi-directional dependency from t-i-e to t-h-t and back - when we change signals in the templates we have to update t-i-e first, and when we change parameters to elements we have to alter t-h-t first. We could choose to make t-h-t and t-i-e completely independent, but don't believe that is a sensible use of time - they are closely connected, even though loosely coupled. Instead we're treating them a single unit: at any point in time t-h-t can only guarantee to deploy images built from some minimum version of t-i-e, and t-i-e can only guarantee to be deployed with some minimum version of t-h-t. The public API here is t-h-t's parameters, and the link to t-i-e is equivalent to the dependency on a helper library for a Python library/program: requiring new minor versions of the helper library is not generally considered to be an API break of the calling code. Upgrades will still work with this constraint - machines will get a new image at the same time as new metadata, with a rebuild in the middle. Downgrades / rollback may require switching to an older template at the same time, but that was already the case.
- Decreasing the accepted metadata for an element if that would result in an error or misbehaviour.
Other sorts of changes may also be backwards incompatible, and if identified will be treated as such - that is, this list is not comprehensive.
We don't consider the internal structure of Heat templates to be an API, nor any test code within the TripleO codebases (whether it may appear to be public or not).
TripleO's incubator is not released and has no backwards compatibility guarantees - but a point in time incubator snapshot interacts with ongoing releases of other components - and they will be following semver, which means that a user wanting stability can get that as long as they don't change the incubator.
TripleO will promote all its component projects to 1.0 within one OpenStack release cycle of them being created. Projects may not become dependencies of a project with a 1.0 or greater version until they are at 1.0 themselves. This restriction serves to prevent version locking (makes upgrades impossible) by the depending version, or breakage (breaks users) if the pre 1.0 project breaks compatibility. Adding new projects will involve creating test jobs that test the desired interactions before the dependency is added, so that the API can be validated before the new project has reached 1.0.
Adopt the following rule on when we are willing to [deliberately] break backwards compatibility:
- When all known uses of the code are for no longer supported OpenStack releases.
- If the PTL signs off on the break. E.g. a high impact security fix for which
we cannot figure out a backwards compatible way to deliver it to our users and distributors.
We also need to:
- Set a timeline for new codebases to become mature (one cycle). Existing codebases will have the clock start when this specification is approved.
- Set rules for allowing anyone to depend on new codebases (codebase must be 1.0.0).
- Document what backwards compatible means in the context of heat templates and elements.
- Add an explicit test job for deploying Icehouse from trunk, because that will tell us about our ability to deploy currently supported OpenStack versions which we could previously deploy - that failing would indicate the proposed patch is backwards incompatible.
- If needed either fix Icehouse, or take a consensus decision to exclude Icehouse support from this policy.
- Commit to preserving backwards compatibility.
When we need alternate codepaths to support backwards compatibility we will mark them clearly to facilitate future cleanup:
# Backwards compatibility: <....>
- We could say that we don't do backwards compatibility and release like the OpenStack API services do, but this makes working with us really difficult and it also forces folk with stable support desires to work from separate branches rather than being able to collaborate on a single codebase.
- We could treat tripleo-heat-templates and tripleo-image-elements separately to the individual components and run them under different rules - e.g. using stable branches rather than semver. But there have been so few times that backwards compatibility would be hard for us that this doesn't seem worth doing.
Keeping code around longer may have security considerations, but this is a well known interaction.
Other End User Impact
End users will love us.
None anticipated. Images will be a marginally larger due to carrying backwards compat code around.
Other Deployer Impact
Deployers will appreciate not having to rework things. Not that they have had to, but still.
Developers will have clear expectations set about backwards compatibility which will help them avoid being asked to rework things. They and reviewers will need to look out for backward incompatible changes and special case handling of them to deliver the compatibility we aspire to.
- Primary assignee:
- Draft this spec.
- Get consensus around it.
- Release all our non-incubator projects as 1.0.0.
- Add Icehouse deploy test job. (Because we could install Icehouse at the start of Juno, and if we get in fast we can keep being able to do so).
None. An argument could be made for doing a quick cleanup of stuff, but the reality is that it's not such a burden we've had to clean it up yet.
To ensure we don't accidentally break backwards compatibility we should look at the oslo cross-project matrix eventually - e.g. run os-refresh-config against older releases of os-apply-config to ensure we're not breaking compatibility. Our general policy of building releases of things and using those goes a long way to giving us good confidence though - we can be fairly sure of no single-step regressions (but will still have to watch out for N-step regressions unless some mechanism is put in place).
The users manual and developer guides should reflect this.