Merge Tracking Requirements and Use Cases

This document details Subversion's merge tracking requirements and their supporting use cases, the majority of which are driven by Subversion's Developer and Merge Meister users. A few outliers are driven by SCM automation.

Summary

Merging

Repeated Merge

Track which changesets have been applied where, so users can repeatedly merge branch A to branch B without having to remember the last range of revisions ported. This would also track changeset cherry picking done by users, so we don't accidentally re-merge changesets that are already applied. This is the problem that svk and arch claim to have already solved, what they're calling star-merge.

Cherry Picking

Merge of one or more individual changes from branch A into branch B. This sometimes involves manual application of the changes from rN in branch A (e.g. not using svn merge), or manual adjustment of a change merged into a WC before it's committed to the repository. Regardless of the merge method used, Subversion must provide a way to indicate that the change(s) have been merged into branch B.

Additionally, it's important to be able to cherry pick changes in multiple different directions. For example, if you create a release branch B by copying the trunk you should be able to both forward port changes made on B into trunk and backport changes made on trunk into B without confusing the merge tracking algorithm.

Use may be predicated on information from the Show Change Sets Available for Merge feature (Auditing section).

Track Renames in Merge

svn merge needs to handle renames better. This requires true rename support.

Edit foo.c on branch A. Rename foo.c to bar.c on branch B.

  1. Try merging the branch A edit into a working copy of branchB: svn merge will skip the file, because it can't find it.
  2. Conversely, try merging branch B rename to branch A: svn merge will delete the 'newer' version of foo.c and add bar.c, which has the older text.

Problem #2 stems from the fact that we don't have true renames, just copies (with history) and deletes. That's not fixable without a FS schema change, and (probably) a libsvn_wc rewrite.

It's not clear what it would take to solve problem #1. See the discussion about our rename woes, and the relationship to merge tracking.

Record Manual Merge

Allow change sets to be marked as merged, effectively a way to manipulate the merge memory on a source and destination. This is related to the revision blocking concept.

Fundamentally, the use case is to support merge tracking of change set which is sufficiently different when ported to a different branch that use of svn merge is no longer appropriate. Examples scenarios include:

  • The actual change you want to apply to branch has no overlap with its incarnation on the source branch, yet is conceptually equivalent.
  • Only a subset of a change set warrants application.
  • The branch content has drifted far enough apart to make automatic merging impossible.

David James cites a specific example from Subversion's own development:

"When I merge from trunk to 1.3.x, I often run into conflicts, so I create a temporary merge branch. If I tried to merge from the temporary merge branch to 1.3.x, the merge metadata would be lost because svnmerge.py does not track indirect metadata. To correct this metadata, I'd use [this feature]."

Use may be predicated on information from the Show Change Sets Available for Merge feature (Auditing section).

Rollback Merge

Undo a merge and/or associated tracking meta data. Necessary for both automated and manual merges.

Use may be predicated on information from the Show Change Sets Available for Merge and Show Change Sets Already Merged features (Auditing section).

Block/Unblock Change Set

Protect revisions which should never be merged from accidental merging.

Properties

Changes to arbitrary properties set on a versioned resource should be mergable exactly as entries within a directory (e.g. add, deleted, etc.), or content in a file (see repeated merge).

Subversion's revision properties (sometimes referred to as revprops) need not be handled.

Merge Previews

Merge previews are dry runs that show conflicts and "non-trivial, non-conflicting" (NTNC) changes in advance. These previews should be exportable and parseable. (From the distributable resolution item in the summit summary.)

SCM Automated Merge

The ability to automate merges (e.g. from a stable branch to a development branch), including interfaces for resolving conflicts and handling other errors, is important. Merge Meisters who do multi-thousand file merges stress this.

Distributability of Merge Resolution

The mechanism for resolving conflicts in a merge should be distributable across N developers. Meeting this requirement might mean a departure for Subversion, since it implies that merge results are not just stored in one working copy, but are somehow available on the server side. (From the distributable resolution item in the summit summary.)

Auditing

Merge tracking must be audit-friendly, supporting some basic forms of reporting which allow for discovery of following types of information:

  • If you merge rN into some destination (e.g. branch B), it should be possible to query rN itself to ask what destinations it has been merged to, and the answer set should contain B.
  • If you merge rN into a branch B, and rN was committed by author A, then svn blame should show the changed lines in B as last touched by A, even if the merge was committed by you and you are not A. This must also work when merging a range of revisions with different authors.
  • If you've blocked some set of revisions from being merged from branch B into some destination (e.g. trunk), you should be able to discover which revisions have been blocked.
  • If you merge rN into a branch B, and rN was committed by author A, then svn log -rN should show both the original author A, and the author who merged the change.
  • It should be possible to query any path (file, directory, or symlink) to find out what changes (revisions) have been merged under it. For files, "under" just means "into".
  • It should be possible to get a 'log' of changes to merge tracking information. This 'log' should list both when the merges were applied, and when the merge-tracking information was updated.
  • It should be easy to discover all the paths at which a particular node revision (i.e., unique versioned file entity) exists, especially in a given revision. In other words, this is the "what branches does this exact version of this file exist in" problem, often requested by so-called enterprise-level users.
  • Merge records should be transitive. Often we merge a bunch of changes to a backport branch, tweak them there, then later merge the branch into a release line. Later queries of the release line should show that the original revisions are present, and queries of the original revisions should show that they went to the release line as well as the backport branch.

Miscellaneous

Allow for Dump/Load

Whatever solution is chosen must play well with svnadmin dump and svnadmin load. For example, the metadata used to store merge tracking history must not be stored in terms of some filesystem backend implementation detail (e.g. "node-revision-ids") unless, perhaps, those IDs are present for all items in the dump as a sort of "soft data" (which would allow them to be used for "translating" the merge tracking data at load time, where those IDs would be otherwise irrelevant). See issue 1525 about user-visible entity IDs.

User Interface Ease in Common Cases

(This was one of the points made at the EuroOSCON 2005 Version Control BOF session.)

The interface for common-case merges should be easy. Currently it is not. For example, a very common case is merging all previously unmerged changes from trunk to branch (more formally, from a source line to a descendant destination line). Today, one must type svn merge -rX:Y URL WC. But why can't the dest just remember what has been merged from that src before and do the right thing? Then one could type svn merge SRC DST. Or better yet, branches could remember where they come from, and one could just type svn merge SRC, or svn merge DST, depending on whether one wants a push or pull interface. (It was pointed out that SVK does remember these things; if someone familiar with the SVK interface wants to put an example transcript here, that would be great.)

Here is the svk command transcript:

A branch creation by: svk copy //project/trunk //project/branch-B (or mirroring an existing subversion repository containing such branch)

To merge from trunk to branch-B: svk smerge //project/trunk //project/branch-B, or svk smerge --to //project/branch-B, or svk pull //project/branch-B

To merge from branch-B back to trunk: svk smerge //project/trunk //project/branch-B, or svk smerge --from //project/branch-B, or svk push //project/branch-B

There is also a -I flag in smerge to merge changes revision-by-revision. Push is by default with the option, and pull is not.

Another common case is porting a single change from one line to another. This currently requires -r<X-1>:X syntax, but Subversion 1.4 will include the -c option (introduced in r17054), so users will no longer need to perform this menial arithmetic. However, Subversion 1.4 will still require the URL of the merge source to be specified; a merge tracking solution that eases common cases would obviate the need for the user to supply the URL when a single change is ported from a branch's ancestor line.

svk merge also supports -c N flag which is -rN-1:N.

Compatibility with older Subversion clients

  • Older Subversion clients should still be able to commit merges on new repositories.
  • Repository administrators should be able to control whether older clients (which do not support merge tracking) can commit changes to specific (e.g. branch or tag) subdirectories.

Versioning

Merge tracking information should be maintained for all versions of files in a repository. Thus, if you create a copy of an old version of a branch, the merge tracking information will still be accurate.

Changes to merge tracking information should be versioned, therefore allowing tracking and auditing of changes to merge tracking data.

Authorization

Repository administrators should be able to control whether merge tracking information can be committed by specific users or to specific directories.

Repository administrators should be able to control whether changes can be committed without associated merge-tracking information. Using this feature, a repository administrator could enforce (for example) that all changes must be committed to trunk before they are merged to a release branch.