SVN Branch Description Format

(CCing everyone from the "Approaches to SVN to Git conversion" thread)

A recent discussion of the remote helper for Subversion turned into a plan for separating out revision->commit mapper projects. Just as content import is split into three parts (svn-fe, fast-import protocol, git-fast-import), r2c mapping can be split into SVN export, protocol, and git import. I've included an initial draft protocol below.

My approximate roadmap from here is to add test cases and a reference implementation to the protocol, then create an SVN exporter from my existing proof-of-concept, then finally write a git importer. My proof-of-concept work is at https://github.com/andrew-sayers/Proof-of-concept-History-Converter

All comments gladly welcomed, but I would particularly appreciate suggestions about how to specify the format more clearly, example SVN use cases that can't be readily expressed with the language as specified, and suggestions about whether forward compatibility should be built in.

	- Andrew

SVN Branch Description Format v0.1 ================================== Andrew Sayers <andrew-svn@pileofstuff.org>

This file specifies a simple format for describing how SVN revisions apply to branches. The goals of this format are to be a communication protocol for programs operating on SVN history and to provide a common language for developers discussing unusual SVN use cases.

Overview --------

The SVN Branch Description Format (``BDF'') is a line-based ASCII format, where each line specifies one action in the SVN history. An SVN history file contains two sections: the first is the ``header'' section that provides information about the file as a whole, the second is the ``body'' section that provides information about things that happened in specific revisions.

Any line in a BDF file that begins with a hash or semicolon character, or that contains only whitespace (including empty lines) is considered to be a comment, and must be ignored. Clients should use a leading '#' to create ordinary comments to be read by users, and a leading ';' for commented-out actions the client wishes to suggest to a user. This makes it easier for a user to accept/reject actions by search-and-replace.

Here is an example file:

    # the file must begin by specifying the revision of the format being used:
    This is a version 0.1 SVN Branch Description file
    Body:
    In r1, create branch "trunk"
    In r10, create branch "branches/1.0" from "trunk" r9
    In r20, create tag "tags/version_1" from "branches/1.0" r19
    In r20, deactivate "tags/version_1"
    # User intervention is required to confirm this action:
    ; In r25, merge "trunk" r24 into "branches/1.0"

Header section --------------

The header section begins with a version identifier, continues with any number of private actions, and ends with the header-body boundary marker.

Any unrecognised action in the header should be treated as a fatal error.

Version identifier ~~~~~~~~~~~~~~~~~~

The first action in this section must exactly match the following:

    This is a version 0.1 SVN Branch Description file

Later versions of the format will use a different identifier here. This might be another number (e.g. ``version 3''), a non-numeric identifier (e.g. ``experimental version''), a different format name (e.g. ``SVN-BDF file''), or anything else. Clients should treat anything other than the exact string above as a fatal error.

Private actions ~~~~~~~~~~~~~~~

Private actions begin with an open bracket and end with a close bracket. For example:

    (my-great-parser will write debugging information to "debug.log")

These actions are intended for internal use by clients using BDF as a storage format. Clients must begin any private action they create with a client-specific identifier (`my-great-parser` in the above example), and must ignore any private action that does not begin with their identifier. These requirements are designed to ensure that clients do not use private actions to communicate with other clients - please send such messages out-of-band (e.g. through arguments to a command-line program), or propose a revision to the format so that all clients can communicate using a standard language.

Header-body boundary marker ~~~~~~~~~~~~~~~~~~~~~~~~~~~

The last action in this section must exactly match the following:

    Body:

This serves only to indicate that the header is finishing, and the body beginning. The remainder of the file must be treated as the body section.

Body section ------------

The body section contains zero or more actions as described below. The following components are widely used in these actions:

Revision identifiers::
  These strings begin with the letter `r`, followed by a number in the
  range 1-9, then zero or more numbers in the range 0-9.  So `r1`,
  `r10` and `r999` are valid revision identifiers, but `1`, `r01` and
  `revision 999` are not.  Revision identifiers are indicated with
  `<revision>` in the definition of an action.

String identifiers::

  These strings begin with a double quote, then contain zero or more
  valid characters, then another double quote.  Valid characters
  include a backslash followed by `\`, `r`, `n`, or `"`, or any
  character other than backslash, carriage return, newline, or double
  quote.  So `"foo"`, `"foo\""` and `"foo\\"` are valid identifiers,
  but `"foo\"`, `"foo""` and `"foo\t"` are not.  Clients must unescape
  string identifiers by converting `\"`, `\\`, `\r` and `\n` to double
  quote, backslash, carriage return and newline respectively.  String
  identifiers are indicated with `<string>` in the definition of an
  action.

These are the valid actions in the body section:

    In <revision>, create branch <string>
    In <revision>, create branch <string> as <string>
    In <revision>, create branch <string> from <string> <revision>
    In <revision>, create branch <string> as <string> from <string> <revision>

    In <revision>, create tag <string>
    In <revision>, create tag <string> as <string>
    In <revision>, create tag <string> from <string> <revision>
    In <revision>, create tag <string> as <string> from <string> <revision>

    In <revision>, deactivate <string>
    In <revision>, delete <string>

    In <revision>, merge <string> up to <revision> into <string>
    In <revision>, cherry-pick <string> <revision> into <string>
    In <revision>, cherry-pick <string> <revision> to <revision> into <string>
    In <revision>, revert <string> <revision> from <string>
    In <revision>, revert <string> <revision> to <revision> from <string>

    In <revision>, ignore <string>
    In <revision>, amend <string>, keeping the old log message
    In <revision>, amend <string>, keeping the new log message
    In <revision>, amend <string>, keeping both log messages

All actions begin with `In <revision>,`. This identifies the revision the action applies to. Each action must refer to a revision greater than or equal to the previous action, except the first action which must be greater than or equal to revision 1. Clients should treat revision numbers that are too low as fatal errors. Clients may treat revision numbers as errors if they are not present in the repository (i.e. higher than the maximum revision or not part of the sub-repository being examined)

Create actions ~~~~~~~~~~~~~~

The `create branch` and `create tag` actions identify a branch or tag being created in the specified revision. The first string identifier indicates the SVN directory name associated with the branch. The `as <string>` identifier indicates the name the user intended for the branch or tag (if unspecified, clients should assume the user intended the branch to have the same name as the directory). The `from <string> <revision>` identifiers indicate the directory associated with a previously-named branch, and the revision number used as the parent for this branch or tag (if unspecified, clients should assume the user intended this to be a trunk). Here are some examples:

    # Create a trunk branch named "trunk" from directory "trunk":
    In r1, create branch "trunk"

    # Create a branch named "branches/foo" from directory "branches/foo",
    # whose parent is the directory "trunk" as it was in revision 5:
    In r10, create branch "branches/foo" as "foo" from "trunk" r5

    # Create a tag named "1.0" from directory "tags/1.0",
    # whose parent is the directory "branches/foo" as it was in revision 15:
    In r20, create tag "tags/1.0" as "1.0" from "branches/foo" r15

Clients should treat it as a fatal error if the `from` revision is greater than the current revision, or if the `from` branch was not active in the specified revision (i.e. it hadn't been created, had been deactivated or had been deleted).

If the `from` branch was active but not changed in the specified revision, clients should behave as if the user specified the last revision in which the branch was edited prior to the specified revision. If the `from` revision is equal to the current revision, clients should warn the user if the `from` branch was edited in that revision. These requirements make it significantly easier for users to manually edit a BDF file.

Clients should treat it as a fatal error if the name specified for a branch or tag is currently in use. A name is currently in use if it was previously declared with one of the `create` actions, and has not yet been deleted with the `delete` action (the `deactivate` action must not be treated as removing a branch name). Clients that check for names currently in use must treat branch and tag names to be in different namespaces - in other words, it is legal to have a branch named `foo` at the same time as a tag named `foo`.

Delete actions ~~~~~~~~~~~~~~

The `deactivate` and `delete` actions identify a branch or tag that should no longer be considered active. The string identifier indicates the directory name associated with the branch.

The `deactivate` action indicates that a branch or tag is still of historical interest, but should no longer be updated when changes are made. For example, a tag might be deactivated after it is created, to indicate that the tag should be considered immutable even though changes were accidentally made later on.

The `delete` action indicates that a branch or tag is no longer of any interest, and can be ignored completely. For example, a branch might be deleted if it had been fully merged into another branch before the directory was removed.

Merge actions ~~~~~~~~~~~~~

The `merge`, `cherry-pick` and `revert` actions identify changes in the relationship between two branches. The first string identifier indicates the directory name for the action's source. The last string identifier indicates the directory name for the action's destination. The revision identifier(s) indicates the revision(s) for the action's source.

The `merge` action indicates that all revisions up to the specified point have been applied from the source to the destination. The `cherry-pick` and `revert` actions identify an inclusive set of revisions that have been applied from the source to the destination. The `merge` and `cherry-pick` actions indicate that revisions that previously had not been merged now have been merged. The `revert` action indicates that revisions which had previously been merged have no longer been merged.

If two revisions were specified, clients must treat it as a fatal error if the second revision is less than or equal to the first.

Clients must treat it as a fatal error if the `from` branch was not active in the specified revision. If two revisions are specified, clients must treat it as a fatal error unless the `from` branch was active in both revisions.

For the `merge` action, if the `from` branch was active but not changed in the specified revision, clients must behave as if the user specified the highest revision before the specified revision in which the branch was changed. If the `from` revision is equal to the current revision, clients should warn the user if the from branch was changed in that revision.

For the two-revision `revert` and `cherry-pick` actions, clients must treat the second revision as in the previous paragraph. If the branch was not changed in the first revision, clients must behave as if the user specified the lowest revision after the specified revision in which the branch was changed.

For all `revert` and `cherry-pick` actions, clients must treat it as a fatal error if no revisions in the specified range changed the branch.

Note: the above requirements for altering revision numbers make it
significantly easier for users to manually edit an SVN Branch
Description file.

If a `merge` action specifies a revision less than or equal to any earlier merge action that has not been reverted, clients must treat it as a fatal error. Clients should not treat it as an error if the merge includes revisions that have previously been cherry-picked.

If a `cherry-pick` action includes the first revision in which the branch was changed after any earlier merge action, or if the destination branch was originally created from the source branch and the `cherry-pick` action includes the next revision in which the branch was changed, clients should warn the user that they probably meant to merge.

If a `revert` action specifies a revision that has not been cherry-picked, or is greater than any earlier merge action that has not been reverted, clients should treat it as a fatal error.

Clients may warn about any merge actions they feel are unusual, but should not treat anything as a fatal error unless specified above.

Note: there is no `unmerge` action.  See the `merge` examples in the
library for how unmerging is achieved.

Edit actions ~~~~~~~~~~~~

The `ignore` and `amend` actions identify a directory and revision which should not create a new revision in the history. For example, an accidental `svn commit` followed immediately by an `svn commit` reverting it might simply be ignored. The string identifier indicates the name of the directory to be edited.

The `ignore` action indicates that the client must act as if no changes were made to the directory in the specified revision, whether or not changes were actually made. Note that this only includes changes made in the SVN repository itself, not those indicated by actions in the SVN history file.

The `amend` actions indicate that the state of the branch in the current revision should overwrite the most recent revision for that branch. Clients may warn, but should not treat it as a fatal error, if the branch was not actually changed in the current revision. When overwriting the most recent revision, clients must retain the revision log from the previous revison if the action specifies `keeping the old log message`, replace the revision log entirely with the log message for the current revision if the action specifies `keeping the new log message`, or concatenate the new log message to the end of the old one if the action specifies `keeping both log messages`. Clients may reformat log messages when keeping both, but are reminded of the need for messages to look sensible when long chains of amendments are created.

Clients should treat it as a fatal error if an edit action is applied to a branch in the same revision as the branch is created.

Consider .SVN_BDF or .SVN.BDF instead of .BDF

I worry about a branch or tag containing a " Can subversion contain a "?

Steve On Mar 11, 2012, at 3:59 AM, Andrew Sayers wrote:

On 18/03/12 23:18, Steven Michalske wrote:

I had a look into valid paths during the week, but hadn't checked quote marks specifically. Thankfully it seems to work:

mkdir test_dir cd test_dir svnadmin create repo mkdir checkout svn checkout file://$(pwd)/repo/ checkout/ cd checkout/ mkdir '"' svn ci -m 'Created directory "'

I'll add this to the test suite :)

Of course, it wouldn't do for SVN to make things that simple. [1] seems to be the official definition of what a valid path looks like, and I've skipped over some important requirements in the spec. Most importantly, redundant '/'s are allowed at the end of a path, and multiple '/'s are collapsed down to one in SVN, so it seems prudent to import that little eccentricity into this format.

I could be persuaded about making '.svn-bdf' the recommended extension, but I'd also be happy to go with a more TLA-friendly name for the whole thing. "SVN Branch Format" would lend itself neatly to a three-letter extension (.sbf) that doesn't appear in Wikipedia's list of file formats[2], although it still encourages the RAS syndrome[3] I've repeatedly stumbled over while writing the spec. "SVN Branching Language" might work, and unlike "BDF" or "SBF", "SBL" doesn't sound at all like "PDF" when mumbled indistinctly. Alternative suggestions are welcome, with the obvious proviso that this is largely subjective and I'm going to pick whatever sounds best to my ear :)

I'm approaching a natural break in defining the format, so I'll paste a new version next week. After that I'll probably pause the format and work on the SVN exporter for a bit, so I'll have a structure to continue building the tests and reference implementation against.

	- Andrew

[1]http://subversion.apache.org/docs/api/latest/group__svn__fs__directories.html#details [2]http://en.wikipedia.org/wiki/List_of_file_formats [3]http://en.wikipedia.org/wiki/RAS_syndrome

(CCing Semen Vadishev as I'd like to know if the SubGit project has any opinion about this)

If you haven't been following the thread - this is a discussion of a new format for describing SVN branching/merging/tagging behaviour. Among other things, this would let people plug different SVN exporters into different Git importers.

I'd like advice/opinions from the community about licensing the specification and reference implementation for this format, because it seems like establishing an open standard is a bit different to promoting open source. I've always thought of copyleft as the tool I use to promote sharing, but standards get more sharing by abandoning copyleft and relying on the network effect - forking a standard makes your product less valuable, unless you're not allowed to use the standard or you have so much market share you don't have to care about standards in the first place.

I'm planning to release the spec under a Creative Commons Attribution-NoDerivs license (i.e. commercial use allowed, changes have to go through me) and the reference implementation under an MIT license (i.e. blatant theft of the exact recommended behaviour is encouraged). This should minimise the barriers for people wanting to implement the format as specified, and maximise the barriers for people wanting to subvert the format. The downside is that it makes life difficult for everyone if I'm hit by a bus, and makes me less inclined to put some of the more complex algorithms into the non-copyleft reference implementation.

Just to be clear, the format is one of three parts involved in getting SVN branches/merges/tags into git. I plan to release an SVN exporter and git importer under the GPL, but expect to make a special case for the format.

So the big question - would you be more inclined to use/contribute to the SVN Branch Description Format if it had a different license?

	- Andrew

Hi Andrew,

Andrew Sayers wrote:

[...]

Yes. By the way, I think fear of forking/discussion of potential improvements/translation into other languages in the context of standards is misguided. If you would like legal protection for your standard, that is what trademark law is for.

Kind regards, Jonathan

On 19/03/12 01:34, Jonathan Nieder wrote:

Could you expand on this? A quick tour of the git codebase suggests your objection is just to the "no derivatives" bit for documentation, and not to the MIT license for code?

It sounds like you're saying that forking isn't a big real-world problem, which I guess makes sense - it'll all work out in the end as long as a single standard is in everybody's interests. So the CC-BY license is my favourite for now.

	- Andrew

On Mon, Mar 19, 2012 at 08:31:41PM +0000, Andrew Sayers wrote:

I think the problem is that there are two levels of forking. You want people to be able to build off of your standard for a number of legitimate reasons. Perhaps they are publishing a draft proposal of enhancements. Perhaps they are adapting parts of the content of your standard to a different domain. Perhaps the original author has become unresponsive and somebody else wants to pick up maintainership. Those are all things we do with code, and they help the ecosystem.

What _isn't_ good is somebody modifying your standard and then claiming that their implementation is "the real SVN History Description" format.

I think Jonathan's point is that CC-BY-ND doesn't allow the legitimate things in the top paragraph. And your real problem (in the second paragraph) is not derivatives, but derivatives claiming to be something they are not (the official standard). And trademarks are the legal tool for avoiding confusion like that.

In practice, I don't think this kind of name-hijacking is a big deal. There are many forks of git, and somebody could make a derivative git that is buggy, interacts badly with existing repository formats, or interacts badly with other git clients via the network protocol. But people are usually kind enough not to call their other implementations "git", and it just works out in practice. So you could probably get by with just a regular source code license (but I am far from an expert, so take the appropriate grain of salt).

-Peff

This is the second draft of the SVN branching language. If you're interested in how SVN revisions map to git commits, please read it and let me know what you think. You might also be interested in the reference implementation and nascent collection of tests[1], although these wouldn't yet withstand a thorough review.

I plan to pause language work and concentrate on SVN export for a while now, because it will be easier to e.g. write tests once I have some code to test.

Here's a little background for people new to the list:

I'm interested in ways to map SVN revisions to git commits, but have found it surprisingly hard to grasp the problem. Having trouble grasping a problem is generally a good sign that your solution is too small, so I split the problem in three:

1. A language to describe SVN branching and merging
2. A program to export such a file from an SVN repository
3. A program to import such a file into a git repository

This has let me concentrate on manageable parts of the problem. For example, Sam Vilain recently pointed out[2] so-called "piecemeal merges" that split a merge across several revisions. There's no way I could have figured out a plan to solve that in one go, but between us we were able to find a representation and a way of importing it into git. I still don't know how to detect and export it from Subversion, but I can worry about that another day.

This is the second draft I've brought to the git mailing list for discussion. The list has a long and proud tradition of including patches in e-mails so people can review code from their e-mail client, but in this case the patch would be so large that it's better simply to include the new text. I've deliberately left the reference implementation and tests out of this e-mail because they're not ready for review.

	- Andrew

[1] https://github.com/andrew-sayers/SVN-Branching-Language [2] http://article.gmane.org/gmane.comp.version-control.git/192418

SVN Branching Language v0.1 =========================== Andrew Sayers <andrew-sbl@pileofstuff.org>

This file specifies a simple language for describing how SVN revisions relate to branches. The goals of this language are to be a communication protocol for programs operating on SVN history and to provide a lingua franca for developers discussing unusual SVN use cases.

Overview --------

The SVN Branching Language (``SBL'') is a line-based UTF-8 format, where each line specifies one action in the SVN history. An SBL file contains two sections: the first is the ``header'' section that provides information about the file as a whole, the second is the ``body'' section that provides information about things that happened in specific revisions.

Any line in an SBL file that begins with a hash or semicolon character, or that contains only whitespace (including empty lines) is considered to be a comment, and must be ignored. Clients should use a leading '#' to create ordinary comments to be read by users, and a leading ';' for commented-out actions the client wishes to suggest to a user. This makes it easier for a user to accept/reject actions by search-and-replace.

Here is an example file:

    # the file must begin by specifying the revision of the language being used:
    This is a version 0.1 SVN Branching Language file
    Body:
    In r1, create branch "trunk"
    In r10, create branch "branches/1.0" from "trunk" r9
    In r20, create tag "tags/version_1" from "branches/1.0" r19
    In r20, deactivate "tags/version_1"
    # User intervention is required to confirm this action:
    ; In r25, merge "trunk" up to r24 into "branches/1.0"

Terminology -----------

The following words should be interpreted as specified below:

Directory::
  A string representing a directory in the SVN tree.  A directory can
  have several flows (i.e. it can be created and deleted multiple
  times), some of which may have names and others not.  The exact
  rules for valid directory names are discussed below, but should be
  as close as possible to the rules for a valid SVN directory name.

Name::
  A string representing a branch or tag.  This is the conventional
  name for a tag as described by users, and is not represented
  anywhere in SVN.  For example, a directory "branches/v1.x/v1.0"
  might have the name "v1.0".  Note that although tags and branches
  are functionally identical, clients must treat tag names and branch
  names as being in different namespaces - for example, a branch name
  "v1.0" and a tag name "v1.0" can both exist at the same time.

Exist::
  A directory is said to exist in a given revision if the directory
  would appear in an `svn ls` command for that revision.  Clients can
  fully implement this language without tracking whether directories
  exist - the term is defined here purely to make discussion easier.

Active::
  A directory is said to be "active" if it its most recent "create"
  action occurred more recently than its most recent "delete" or
  "deactivate" action.  A branch will usually be active if it exists
  and has a name, but there are special cases - for example, an
  existent directory associated with a tag might be deactivated as a
  precautionary measure in case of accidental commits.  Note that only
  directories are said to be "active" - branches and tags are
  instead "accessible" (under a slightly different set of
  circumstances).

Accessible::
  A branch or tag is said to be "accessible" if it has been created
  more recently than it has been deleted (but not deactivated).  Note
  that only branches and tags are said to be "accessible".
  Directories are instead "active" (under a slightly different set of
  circumstances).

Trunk::
  A directory, branch or tag is said to be a "trunk" if it has no
  parent directory, branch or tag.  Trunk tags are possible but
  exceedingly rare.

Header section --------------

The header section begins with a version identifier, continues with any number of private actions, and ends with the header-body boundary marker.

Any unrecognised action in the header should be treated as a fatal error.

Version identifier ~~~~~~~~~~~~~~~~~~

The first action in this section (not including comments) must exactly match the following:

    This is a version 0.1 SVN Branching Language file

Later versions of the language will use a different identifier here. This might be another number (e.g. ``version 3''), a non-numeric identifier (e.g. ``experimental version''), a different name (e.g. ``SBL file''), or anything else. Clients must treat anything other than the exact string above as a fatal error.

Private actions ~~~~~~~~~~~~~~~

Private actions begin with an open bracket and end with a close bracket. For example:

    (my-great-parser will write debugging information to "debug.log")

These actions are intended for internal use by clients using SBL as a storage format. Clients must begin any private action they create with a client-specific identifier followed by a space (`my-great-parser` in the above example), and must ignore any private action that does not begin with their identifier. Client-specific identifiers must contain one or more characters and must not include the space, carriage return or newline characters. These requirements are designed to ensure that clients do not use private actions to communicate with other clients - please send such messages out-of-band (e.g. through arguments to a command-line program), or propose a revision to the language so that all clients can communicate using a standard language.

Header-body boundary marker ~~~~~~~~~~~~~~~~~~~~~~~~~~~

The last action in this section must exactly match the following:

    Body:

This serves only to indicate that the header is finishing, and the body beginning. The remainder of the file must be treated as the body section. Note that this action significantly reduces the complexity of writing an SBL parser, because clients can use this to switch between ``header'' and ``body'' parsing modes.

Body section ------------

The body section contains zero or more actions as described below. Any unrecognised action in the header should be treated as a fatal error.

The following components are widely used in body actions:

Revision identifiers::
  These strings begin with the letter `r`, followed by a number in the
  range 1-9, then zero or more numbers in the range 0-9.  So `r1`,
  `r10` and `r999` are valid revision identifiers, but `1`, `r01` and
  `revision 999` are not.  Revision identifiers are indicated with
  `<revision>` in the definition of an action.

String identifiers::

  These strings begin with a double quote, then contain zero or more
  valid characters, then another double quote.  Valid characters
  include a backslash followed by `\`, `r`, `n`, or `"`, or any
  character other than backslash, carriage return, newline, double
  quote, or the null character (U+0000).  So `"foo"`, `"foo\""` and
  `"foo\\"` are valid identifiers, but `"foo\"`, `"foo""` and
  `"foo\t"` are not.  Clients must unescape string identifiers by
  removing the leading/trailing quotes and converting `\"`, `\\`, `\r`
  and `\n` to double quote, backslash, carriage return and newline
  respectively.

Directory identifiers::
  These string identifiers indicate a directory.  As well as the
  requirements of a string identifier, valid directory identifiers
  must be valid directories when unescaped.  A valid directory must be
  a sequence of zero or more directory entry names, separated by slash
  characters `/`, and possibly ending with slash characters.  Each
  directory entry name can contain any Unicode character except the
  null character and the slash character `/`. No directory entry may
  be named `.`, `..`, or the empty string.  When unescaping a
  directory, clients must first do the unescaping for a string
  identifier, then convert the identifier to Unicode canonical
  decomposition (NFD) form, then collapse every series of '/'
  characters to a single character (e.g. `foo//bar` should become
  `foo/bar`), and finally remove any trailing `/` character.
  Directory identifiers are indicated with `<directory>` in the
  definition of an action.
  Note that although this definition is based on SVN's own
  documentation, SBL clients must follow the rules above instead of
  the SVN rules in the event of any contradiction between the two.
  SVN's documentation is available from their source repository:
  http://subversion.apache.org/docs/api/latest/group__svn__fs__directories.html#details

Name identifiers::
  These string identifiers indicate a branch or tag name.  As well as
  the requirements for a string identifier, valid name identifiers
  must be valid names when unescaped (see the "terminology" section
  for details).  As well as the rules for string identifers, the empty
  string is not a valid name identifier.

These are the valid actions in the body section:

    In <revision>, create branch <directory>
    In <revision>, create branch <directory> as <name>
    In <revision>, create branch <directory> from <directory> <revision>
    In <revision>, create branch <directory> as <name> from <directory> <revision>

    In <revision>, create tag <directory>
    In <revision>, create tag <directory> as <name>
    In <revision>, create tag <directory> from <directory> <revision>
    In <revision>, create tag <directory> as <name> from <directory> <revision>

    In <revision>, deactivate <directory>
    In <revision>, delete <directory>
    In <revision>, delete branch <name>
    In <revision>, delete tag <name>

    In <revision>, merge <directory> up to <revision> into <directory>
    In <revision>, cherry-pick <directory> <revision> into <directory>
    In <revision>, cherry-pick <directory> <revision> to <revision> into <directory>
    In <revision>, revert <directory> <revision> from <directory>
    In <revision>, revert <directory> <revision> to <revision> from <directory>
    
    In <revision>, ignore <directory>
    In <revision>, amend <directory>, keeping the old log message
    In <revision>, amend <directory>, keeping the new log message
    In <revision>, amend <directory>, keeping both log messages

All actions begin with `In <revision>,`. This identifies the revision the action applies to. Each action must refer to a revision greater than or equal to the previous action, except the first action which must be greater than or equal to revision 1. Clients should treat revision numbers that are too low as fatal errors.

Create actions ~~~~~~~~~~~~~~

The `create branch` and `create tag` actions identify a directory being made active and a branch/tag being made accessible in the specified revision. The first string identifier indicates the SVN directory name associated with the branch/tag. The `as <name>` identifier indicates the name the user intended for the branch/tag. If the `as <name>` identifier is not specified, clients must use the directory name as the identifier, unless the directory name is the empty string (root directory), in which case they should treat it as a fatal error (because the empty string is a valid directory but not a valid name). The `from <directory> <revision>` identifiers indicate the directory and revision number to be used as the parent for this branch/tag (if unspecified, clients should assume the user intended this to be a trunk). Here are some examples:

    # Create a trunk branch named "trunk" from directory "trunk":
    In r1, create branch "trunk"

    # Create a branch named "branches/foo" from directory "branches/foo",
    # whose parent is the directory "trunk" as it was in revision 5:
    In r10, create branch "branches/foo" as "foo" from "trunk" r5

    # Create a tag named "1.0" from directory "tags/1.0",
    # whose parent is the directory "branches/foo" as it was in revision 15:
    In r20, create tag "tags/1.0" as "1.0" from "branches/foo" r15

Clients should treat it as a fatal error if the directory was already active in the current revision.

Clients should treat it as a fatal error if the name was already accessible in the current revision (remember that branches and tags must be treated as being in different namespaces).

Clients should treat it as a fatal error if the `from` revision is greater than the current revision, or if the `from` directory was not accessible in the specified revision.

If the `from` directory was active but not changed in the specified revision, clients should behave as if the user specified the last revision in which the directory was changed prior to the specified revision. If the `from` revision is equal to the current revision, and the `from` directory was changed in the current revision, then clients should warn the user, but should not treat it as a fatal error. These requirements make it significantly easier for users to manually edit SBL files.

Delete actions ~~~~~~~~~~~~~~

The `deactivate` and `delete` actions identify a name that should become inaccessible, and/or a directory that should become inactive. The string identifier indicates the directory or name to be modified.

The `deactivate` action indicates that a directory should become inactive, but that the associated name should remain accessible. This is commonly used for a branch or tag that is still of historical interest, but should no longer be updated when changes are made. For example, a tag might be deactivated after it is created, to indicate that the tag should be considered immutable even though changes were accidentally committed to the directory later on.

The `delete` action indicates that a directory should become inactive, and the associated name should become inaccessible. This is commonly used when a branch or tag is no longer of any interest, and can be ignored completely. For example, a branch might be deleted if it had been fully merged into another branch before the directory was removed.

The `delete branch` and `delete tag` actions indicate that a name should become inaccessible, and the associated directory should become inactive if it is still active. For example, the branch associated with a deactivated directory might be deleted when the user wants to create a new branch with the same name.

Clients should only generate `delete branch` and `delete tag` actions if it would be unsafe to generate a `delete` action because the directory has already been deactivated, but should expect users to manually edit files and add any type of action.

Actions that deactivate a directory should treat it as a fatal error if the directory is not currently active. This includes the `delete` action, which behave in ways users would not exect if the directory is renamed or replaced in SVN before the action occurs.

Actions that make a name inaccessible should treat it as a fatal error if the name is not currently accessible.

Merge actions ~~~~~~~~~~~~~

The `merge`, `cherry-pick` and `revert` actions identify changes in the relationship between two directories. The first string identifier indicates the directory name for the action's source. The last string identifier indicates the directory name for the action's destination. The revision identifier(s) indicates the revision(s) for the action's source.

The `merge` action indicates that all revisions up to the specified point have been applied from the source to the destination. The `cherry-pick` and `revert` actions identify an inclusive set of revisions that have been applied from the source to the destination. The `merge` and `cherry-pick` actions indicate that revisions that previously had not been applied now have been applied. The `revert` action indicates that revisions which had previously been applied have no longer been applied.

If two revisions were specified, clients must treat it as a fatal error if the first revision is greater than the second.

Clients must treat it as a fatal error if the `from` directory was not active in the specified revision. If two revisions are specified, clients must treat it as a fatal error if the `from` directory was inactive in either revision, or was deactivated after the first revision and reactivated before the last revision.

For the `merge` action, if the `from` directory was active but not changed in the specified revision, clients must behave as if the user specified the highest revision in which the directory was changed before the revision they actually specified. If the `from` revision is equal to the current revision, clients should warn the user if the from directory was changed in that revision.

For the two-revision `revert` and `cherry-pick` actions, clients must treat the second revision as in the previous paragraph. If the from directory was not changed in the first revision, clients must behave as if the user specified the lowest revision in which the directory was changed after the specified revision.

For all `revert` and `cherry-pick` actions, clients must treat it as a fatal error if no revisions in the specified range changed the directory.

Note: the above requirements for altering revision numbers make it
significantly easier for users to manually edit an SBL file.

If a `merge` action specifies a revision less than or equal to any earlier merge action that has not been reverted, clients should treat it as a fatal error. Clients should not treat it as a fatal error if the merge includes revisions that have previously been cherry-picked.

If a `cherry-pick` action includes the first revision in which the directory was changed after any earlier merge action, or if the destination directory was originally created from the source directory and the `cherry-pick` action includes the next revision in which the directory was changed, clients should warn the user that they probably meant to merge.

If a `revert` action specifies a revision that has not been cherry-picked, or is greater than any earlier merge action that has not been reverted, clients should treat it as a fatal error.

Clients may warn about any merge actions they feel are unusual, but should not treat anything as a fatal error unless specified above.

Note: there is no `unmerge` action.  See `t/basic/unmerge.sh` for
how unmerging is achieved.

Edit actions ~~~~~~~~~~~~

The `ignore` and `amend` actions identify a directory and revision which should not create a new revision in the history. For example, an accidental `svn commit` followed immediately by an `svn commit` reverting it might simply be ignored. The string identifier indicates the name of the directory to be edited.

The `ignore` action indicates that the client must act as if no changes were made to the directory in the specified revision, whether or not changes were actually made. Note that this only includes changes made in the SVN repository itself, not those indicated by actions in the SBL file.

The `amend` actions indicate that the state of the directory in the current revision should overwrite the most recent revision for that directory. When overwriting the most recent revision, clients must retain the revision log from the previous revision if the action specifies `keeping the old log message`, replace the revision log entirely with the log message for the current revision if the action specifies `keeping the new log message`, or concatenate the new log message to the end of the old one if the action specifies `keeping both log messages`. Clients may reformat log messages when keeping both, but are reminded of the need for messages to look sensible when long chains of amendments are created.

Clients may warn, but should not treat it as a fatal error, if an edit action is specified for a directory that was not changed in the current revision.

Clients should treat it as a fatal error if an edit action is applied to a directory in the same revision as the directory becomes active.

Limitations -----------

SBL only allows a directory to be associated with a maximum of one branch or one tag at once. There is no clear use case for allowing a user to e.g. associate a tag with a directory that is already a branch, and disallowing it enables better detection of user errors. Future versions of this language might allow multiple branches/tags per directory if a use case is found.

Although not explicitly forbidden, clients are not required to support recursive branches. For example, if ``trunk'' is a branch, clients may assume that ``trunk/foo'' is neither a branch nor a tag.

See `t/advanced/subproject_branch.sh` for a common edge case for parsers that disallow recursive branches.

License -------

SVN Branching Language by Andrew Sayers <andrew-sbl@pileofstuff.org> is licensed under a Creative Commons Attribution 3.0 Unported License.

The full license is available here: http://creativecommons.org/licenses/by/3.0/legalcode A human-readable summary is available here: http://creativecommons.org/licenses/by/3.0/

The following explanation must have no bearing on the meaning of the above license:

My primary goal in drafting this language was to provide a single format that everyone could use to describe the behaviour of SVN repositories. I felt this goal was best served by allowing commercial use of the language, and by allowing other dialects to be formed but discouraged because of the network effect of a common language.

Hi,

Andrew Sayers wrote:

I found this pretty interesting. Doesn't it duplicate some of the functionality of reposurgeon [1] though?

[1]: http://esr.ibiblio.org/?p=4071

    Ram

On 30/03/12 05:06, Ramkumar Ramachandra wrote:

Yes, I've been procrastinating all week instead of reading up on reposurgeon and contacting ESR about possibile collaboration.

I think you need something a bit more expressive than reposurgeon's format to do SVN<->Git conversion well, and I think you need something a bit more accessible in order to document SVN edge cases. For example, I don't see how reposurgeon could represent all the madness around SVN cherry-picks that become merges when you manually add information from revision logs, then become cherry-picks again when you find a revert coming in from another branch. Having said that, a (lossy) conversion between SBL and reposurgeon format would probably be useful and not that hard.

The link above put it very well that most people leave an embarrassed “to be done” comment and disappear when they realise how much of a nightmare the mapping is. What it doesn't mention is that everyone experiences a slightly different part of the nightmare, and that we can only really tackle the problem by getting everyone's freaky edge cases written up in one language in one place. The test suite[1] isn't that impressive right now, but in the long-term I'm really keen to get implementers to pool their knowledge so we can all benefit. SBL is designed to let people open the relevant test without reading the spec and say "oh right I understand what a piecemeal merge is now. I'll go implement that in my project".

I'm currently working on code to read an SVN dump and write to SBL. This will definitely overlap with reposurgeon's SVN export functionality, but without seeing the final code I can't say how much. That's fine though - as I say, the only way to get a good solution is for multiple implementers to investigate the problem and share the edge cases they find.

	- Andrew

[1]https://github.com/andrew-sayers/SVN-Branching-Language/tree/master/t