Changeset Evolution
Contents
Changeset Evolution is a set of features to gracefully handle history rework operations.
It offers a safe and simple way to refine changesets locally and propagate those changes to other repositories.
It can automatically detect and handle the complex issues that can arise from exchanging draft changesets.
It makes it even possible for multiple developers to rewrite the same part of the history in a distributed way.
It fully respects the Phases concept so unsure users will only rewrite parts of the history that are safe to be changed. Phases have been part of Mercurial since early 2012.
You can read the documentation or watch the talk at FOSDEM 2013 to learn more. (Slides).
1. Status
While well on the way, the full implementation of the changeset evolution concept is still in progress. Core Mercurial already supports many of the associated features, but for now they are still disabled by default. The current implementation has been usable for multiple years already, and some parts of it are used in production in multiple projects and companies (including the Mercurial project itself, Facebook, Google, etc…).
However, there are still some areas were the current implementation has gaps. This means some use cases or performance issues are not handled as well as they currently are without evolution. Mercurial has been around for a long time and is strongly committed to backward compatibility, and turning evolution on by default nowadays could regress the experience of some of our current users. The feature will be enabled by default at the point where users who do not use or care about the new features added by evolution won't be impacted by it.
2. Using Evolution
Using evolution is safe and no data loss/corruption is to be expected. Once you turn evolution on, all commands from core Mercurial will use it. In addition, one can enable evolution locally and still use an "old" server. You won't "poison" the server (but won't be able to use the new feature with that server).
Testing the concept early is useful for us, core developers. It provides us with data and use cases that we need to iron things out before enabling it by default. This is especially true if you can test evolve in a context that involves client/server/distributed workflows.
Here are the recommended steps to try evolution:
Subscribe to the evolve beta tester mailing list,
Install and use the EvolveExtension
- Enjoy changeset evolution
The evolve extensions will take care of enabling all the appropriate features in core and add a small layer with the latest state of the new commands and algorithms. The extension is developed by core Mercurial developers and its code collaborates closely with Mercurial's core internals. It is left outside of core to support older versions of Mercurial (extending the tester base) and gain flexibility when experimenting with new algorithms.
As the implementation is still in progress, some command behavior might change. Subscribe to the evolve beta tester mailing list to make sure you stay aware of new releases and the changes they might introduce.
3. Overview
You can also check the evolve documentation.
3.1. Rewriting history
Mercurial offers multiple commands to rewrite history:
hg commit --amend: can add more changes into a commit
hg rebase: can move changesets around in your graph (requires the RebaseExtension)
hg histedit: can perform rewrite operations on some of your changesets (requires the HisteditExtension)
The experimental EvolveExtension adds more commands, which will eventually be moved into core:
hg uncommit: can remove changes from a commit and put them back in your working directory
hg fold: can squash multiple changesets together as a single new commit
hg prune: can remove changesets from your history
hg split: can split a changeset into smaller ones
All these operations are very safe to use, even for Mercurial rookies. Mercurial will actively prevent you from rewriting parts of history which are not safe to rewrite. Read about the Phases concept for details.
3.2. Tracking and sharing rewriting
Obsolescence markers make it possible to mark changesets that have been deleted or superseded by a new version of the changeset.
Unlike the previous way of handling such changes (which stripped the old changesets from the repository), obsolescence markers can be propagated between repositories. This allows for a safe and simple way of exchanging mutable history and altering it after the fact. Changeset phases are respected, such that only draft and secret changesets can be altered (see hg phases for details).
Obsolescence is tracked using "obsolescence markers", a piece of metadata that tracks which changesets have been made obsolete, potential successors for a given changeset, the moment the changeset was marked as obsolete, and the user who performed the rewriting operation. The markers are stored separately from standard changeset data and can be exchanged without any of the precursor changesets, preventing unnecessary exchange of obsolete data.
The complete set of obsolescence markers describes a history of changeset modifications that is orthogonal to the repository history of file modifications. This changeset history allows for detection and automatic resolution of edge cases arising from multiple users rewriting the same part of history concurrently.
3.3. Automatic detection and resolution arising troubles
Exchanging mutable changesets has inherent issues that we must be prepare to deal with. Most people will never run into them but Mercurial is able to detect and solve them automatically. Please note: there was a renaming of the terms used here; what used to be called "troubled" (encompassing the three kinds of issues below) is now called "unstable". You may still encounter documentation using the old terms, please point it out or contribute an update if you notice any.
There are three kinds of unstable changesets:
- In some situations you may have non-obsolete changesets descending from obsolete changesets. Such changesets are said to be "orphaned" (this used to be called 'unstable').
In some other situations you may have successors for changesets which are now immutable. In such case the obsolescence marker does not apply and the unlucky successors are said to be "phase-divergent" (this used to be called 'bumped').
- Finally when multiple changesets claim to be the successors of changesets they are said to be "content-divergent" (the old name for this was simply 'divergent').
When Mercurial detect such unstable changesets, it will warn the user and prevent push by default. You can use the hg evolve command to automatically resolve them.
This command is partially implemented in the EvolveExtension.
4. Current implementation state
As of January 2017, the following areas are covered:
- All commands know how to create, store and read obsolescence information,
- obsolete changesets are properly excluded from the user's view and not exchanged between repositories,
- the obsolescence markers are properly exchanged between repositories, propagating relevant obsolescence information alongside the changesets,
- hg log and summary will display some information related to evolution,
- commands like hg update, hg pull and hg rebase can use obsolescence information to make smarter decisions,
- all evolution issues are properly detected and diagnosed,
the evolve command can automatically resolve the most common cases of such issues.
And the following areas need improvement: (get on the tester mailing list if you get stuck on one of them)
- the most rare/advance case of evolution issues are not automatically handled yet,
- we need more barriers to prevent the user from shooting themselves in the foot in some corner cases
- raw obsolescence data is available to the user, but we lack good tools to visualise the obsolescence history,
- bringing an obsolete changeset to life is harder than it should be,
- support for other advance maintenance operations (like bundling or stripping obsmarkers) is missing,
- during exchange, the discovery protocol for markers can be very slow in some case (we have a second iteration to be implemented),
- we need more scaling work for large/old installations, for update and bookmarks.
5. Older materials
Initial presentation at Copenhagen: http://public.octopoid.net/talk.pdf
First tutorial, written by Peter Arrenbrecht: http://arrenbrecht.ch/mercurial/evolution/
Blog post by Arne Babenhauserheide (Test of the hg evolve extension for easier upstreaming): http://draketo.de/light/english/mercurial/hg-evolve-2013-01-12