Clone this wiki locally
(proposal only.. still under discussion -warner 16-Aug-2012)
Managing The Changelog
Each release (or "train", in our process) may change behaviors that affect different groups: features are added or deprecated, runtime requirements are modified, installations must be upgraded or changed. It should be easy for the affected parties to learn what they must do when they deploy the new release.
Requirements of the changelog-management process:
- easy for developers to quickly record significant behavioral changes
- easy for users (IdPs, RPs, dev/prod ops, QA) to learn what affects them
- play nicely with DVCS, pull-requests
- minimize workload of Release Manager who produces a NEWS announcement
- easy for reviewers to enforce
This proposal is based upon the experience of the Twisted (Python) developer community, as described here:
Each pull-request should add files to the
docs/changes/ directory. The files should be named
6543.ops. Patches which don't have specific issue numbers should use some other probably-unique identifier (a few words that summarize the issue) instead of the Issue number.
These snippets should contain a few sentences explaining what specific audiences need to know about the effects of the patch. If an audience doesn't need to know about the changes, then the patch should not include a snippet for that audience. Many bugfixes will have no snippets at all.
The various audiences, type suffixes, and what they care about, are:
- IdPs (
123.idp): Identity Providers care about format changes to the
/.well-known/browseridfile and the public keys inside it, changes to the certificate-signing process that they must implement, and the
navigator.idAPIs used by their authentication and provisioning pages.
- RPs (
123.rp): Relying Parties care about changes to the URLs of
include.jsand the verifier service, API changes of
navigator.idand the verifier service, and new features their users may be able to take advantage of.
- devops/prodops (
123.ops): Operations personnel care about changes to the way the browserid services are deployed on Mozilla servers: new processes (like the "router" and "keysigner"), new outbound network requests (firewall rules), dependencies on third-party libraries and tools (including node.js), database schema changes, significant changes to expected load or database access patterns. They should also review code that adds new database queries for efficiency.
- QA (
123.qa): QA staff care about new features that need test coverage, behavioral changes which affect QA methodology, and dialog/UI changes that require "bidpom" model updates (e.g. CSS selectors).
We may add other audiences in the future.
Reviewers should check for snippets in the pull-requests they review, and should feel free to ask patch authors to write them when necessary.
These snippets should be addressed to the audience in question, and cite published documentation (if available), like this:
RPs can now include a `siteLogo` URL in their `navigator.id.request()` call, which should point to a site-relative path of a .png image file that will be displayed in the sign-in dialog. See XYZ for details.
The Release Process
Periodically, just before a train is branched, the Release Manager (i.e. Lloyd) will land a commit which deletes all the snippet files and merges their contents into more-nicely-formatted NEWS or ChangeLog file. The RM will exercise editorial judgement on what needs to be visible (sometimes an IdP-oriented change needs to be announced to RPs too), and gets to add detail or delete low-level entries entirely.
If the prose quality of the snippets is high enough, this process may be done mechanically some day.
By putting the snippets in separate files, they will not cause merge conflicts. By recording the snippets in the same branch as the behavioral change, the presence of a snippet will reliably indicate whether the described changes have taken effect or not. (this means that a change which adds the implementation of a new feature, but leaves it turned off, should not have the snippet, but a subsequent change which "flips the switch" does get the snippet).
By merging together the snippets periodically, we avoid an unbounded proliferation of tiny hard-to-read files, and get nice human-readable release notes for each train.