Re: Semantics of Move
From: Julian Foad <julianfoad_at_btopenworld.com>
Date: Tue, 3 Sep 2013 16:59:04 +0100 (BST)
On 2013-08-16, Blair Zajac wrote:
In discussion yesterday I realized that the (node-id, copy-id) pair from the filesystem is still not what I want to refer to in this definition of semantics, because of lazy copies. When we copy a directory, we don't assign a new copy-id to each child right away, but only if and when the child is modified. I want to define the move semantics in terms of some (conceptual) id that is always new for each copied path, including all the unmodified children of a copied directory.
Therefore I've reworked the definition and also expanded and rewritten some other parts of it. The result is at <http://wiki.apache.org/subversion/MoveDev/MoveDev#Move_Semantics> and repeated in plain text below.
This section specifies the logical semantics of the versioned move operation that is the basis of move tracking, independent of any implementation.
== Node-Line-Id ==
Assume that each versioned path in a revision has an identifier that we will call its node-line-id. The node-line-id need not physically exist: it is a concept used in the definition of moves but not necessarily in the implementation.
The node-line-id is preserved when the content of a node is modified. A new node-line-id is assigned to every new node that is created by addition or by copying, including when it replaces a previous node at the same path. This new node-line-id is unique within the whole repository. Within any given revision, each node-line-id is unique among all the paths.
The node-line-id is similar to the (node-id, copy-id) tuple in the existing Subversion filesystems, except that the lazy-copy mechanism does not assign a new copy-id to a child of a copy until that child (or one of its descendants) is modified. Therefore an unmodified child of a copy has the same (node-id, copy-id) as the corresponding child path of the copy source, whereas (by definition) it has a new node-line-id.
Let the term node-line refer to the set of PATH_at_REV locations that have a given node-line-id.
== Definition of Move ==
Given a node-line N and two revisions rX and rY (X < Y), the definition of N being moved in rY with respect to rX is:
* Same node-line. A node-line with node-line-id N exists at path PN,X in rX and at path PN,Y in rY. It is the same node-line, and so has the same node kind. Its content may differ.
* Move and/or rename. The node-line N has either or both of: a different name (base name) in rX than in rY; and/or a different parent (parent directory node-line-id) in rX than in rY. Thus, the paths PN,X and PN,Y typically differ but may be the same.
* No gap. There cannot be a gap in the range of revisions: node-line-id N exists in every revision rX, rX+1, ..., rY-1, rY. Contrast with copy-and-delete, where there can be a gap between the delete and the copy. The possible "resurrection" extension to the move semantics would permit a gap.
* Children follow. If N is a directory, each child (recursively) of N in rX remains a child of N in rY, with the same name, unless it is separately moved or deleted. Any or all of the children can be separately moved within or outside the subtree at N, at the same time as N is moved.
== Properties of Move ==
Some properties of the move relationship are:
* Unique. move(A_at_X,B_at_Y) and move(A_at_X,C_at_Y) cannot both be true.
* Transitive. move(A_at_X,B_at_Y) followed by move(B_at_Y,C_at_Z) collapses to move(A_at_X,C_at_Z).
* Reversible. move(A_at_X,B_at_Y) followed by move(B_at_Y,A_at_Z) collapses to no-move.
* Time-symmetric. move(A_at_X,B_at_Y) is symmetric with the time-reversed relationship move(B_at_Y,A_at_X).
* No null move. An attempted move which does not change the node's name or its parent node, with or without a modification, is not distinguished from a normal succession of history.
In the notation A_at_X, A represents a parent directory node-line-id and child name (rather than a full path), X represents a revision number, and A != B, and X != Y, etc.
== Move versus Rename ==
In the versioned data model semantics, "move" refers to a change of parent node and/or a change of name.
At a higher level of semantics, for example when resolving conflicts during merge, it can be useful to distinguish between renaming and moving to a different parent node.
== Can't Move a Child of a Copy ==
Moving a child of a copy, within the same revision, is not tracked: it is an unversioned operation.
A versioned "move" takes a node that existed in the previous revision and places it in a new location. A copy, however, always creates new nodes, conceptually, even if the internal representation is a "lazy copy" pointer to the old node. Moving a child node therefore is a rearrangement of the new content. It is semantically the same as deleting the child node and creating a copy of it somewhere else. Compare with copying a node and then moving that copy somewhere else.
If we perform a copy and then move a child of it, either in a WC or in a repository, this should create a copy with a deleted child, and then another copy somewhere else which is the "moved" child in its new location. We can describe the relationship between the initial and final states perfectly well without saying "move", in the form "the subtree at path P_at_X is copied to path P1_at_Y, except for its child C which is copied to path P2/C_at_Y instead". Thus there is no loss of semantic information despite the absence of "move" in the result.
== Resurrecting a Deleted Node ==
A possible extension to the move semantics would be to allow a previously deleted node to be resurrected at the same location or at a different location.
The mental model is this. When a node is deleted, it is unlinked from the versioned tree, but its content continues to exist in the repository. When the node is resurrected, a link to that last version of its content is put into the versioned tree, like undoing the delete. The new link can be made anywhere in the versioned tree, like undoing the delete and moving the node at the same time.
This is not merely a blue-sky curiosity: it may necessary in order to ensure logical completeness. For example, let's say the node initially at path A/foo is moved to B/foo and then back to A/foo. If we create a branch C from A, and are continuously merging all changes under A into C, then C/foo will be deleted and will later be recreated. Or if 'svnsync' is replicating only subtree 'A' of repository R1 into repository R2, then in repository R2 we will see A/foo disappear and then reappear. In cases such as these, the 'reappearance' should be modeled as a resurrection; if it is modeled as a plain add or as a copy, it will not have the correct semantics of being the 'same' node that it was before.
This is an archived mail posted to the Subversion Dev mailing list.