Mercurial's design

The guts of Mercurial.

Data structures

Mercurial uses a few fundamental objects.

Nodeids

Nodeids are unique ids that represent the contents of a file and its position in the project history. For now they are computed using the SHA1 hash function, which generates 160 bits (40 hex digits). If you modify a file, commit the change, and then modify it to restore the original contents, the contents are the same but the history is different, so the file will get a new nodeid. This history-sensitivity is obtained by calculating the nodeid from the concatentation of the parent nodeids with the file's contents.

Revlogs

A revlog, for example .hgdatasomefile.d, is the most important data structure and represents all versions of a file. Each version is stored compressed in its entirety or stored as a compressed binary delta (difference) relative to the preceeding version in the revlog. Whether to store a full version is decided by how much data would be needed to reconstruct the file. This system ensures that Mercurial does not need huge amounts of data to reconstruct any version of a file, no matter how many versions are stored.

The reconstruction requires a single read, if Mercurial knows when and where to read. Each revlog therefore has an index, for example .hgdatasomefile.i, containing one fixed-size record for each version. The record contains:

• the nodeid of the file version
• the nodeids of its parents
• how many bytes to read from the revlog
• the offset in the revlog saying where to begin reading

With one read of the index to fetch the record and then one read of the revlog, Mercurial can in time proportional to the file size reconstruct any version of a file.

So that adding a new version requires only O(1) seeks, the revlogs and their indices are append-only.

Revlogs are also used for manifests and changesets.

Manifests

A manifest describes the state of a project by listing each file and its nodeid to specify which version. Recreating a particular state means simply looking up its manifest and reconstructing the listed file versions from their revlogs. The manifest is conceptually a file. All of its versions, which collectively represent the entire project history -- are stored in a revlog (see the file .hg00manifest.d) and an associated index (.hg00manfesti.i).

Changesets

A changeset lists all files changed in a checkin along with a change description and metadata like user and date. It also contains a nodeid of the resulting manifest. Given enough history information, a changeset can be converted into a manifest. Given enough history information and the metadata, a manifest can be converted into a changeset. So manifests are reundant, but maintaining both representations of the project history speeds up Mercurial.

Merging

These data structures are designed for merging, the fundamental operation in a distributed SCM that encourages branching.

Graph merging

Merging a pair of directed acyclic graphs (DAGs) -- the family tree of the file history -- requires determining whether nodes in different graphs correspond. Comparing the node contents (or hashes of the contents) is incorrect because it ignores the history.

However, using the nodeid avoids this error because the nodeid describes the file's contents and its graph position relative to the root. A merge simply checks whether each nodeid in graph A is in graph B and vice versa (for example using a hash table), and Mercurial adds the new nodes to the append-only revlog.

Branching and merging

Every working directory is potentially a branch and every user effectively works in their own branch. When Mercurial checks out a branch, it remembers the changeset that directly led to it so that the next checkin will have the correct parent.

To merge two branches, you check out their heads into the same working directory, which also performs a merge, and then check in (commit) the result once you're happy with the merge. The resulting checkin has two parents.

Mercurial decides when a merge is necessary by first determining whether the working directory contains uncommitted changes. This determination effectively turns the working directory into a branch of the checked-in version on which it is based. If the working directory is a direct ancestor or descendent of the second version that we're attempting to checkout, Mercurial replaces the working-directory version with the new version. Otherwise it merges the two versions.

Merging manifests

To merge manifests, first compare them and decide which files need to be added, deleted, and merged.

For each file to be merged, perform a graph merge and resolve conflicts as above. It's important to merge files using per-file DAGs rather than just changeset-level DAGs as this diagram illustrates:

 M   M1   M2

   AB
    |`-------v     M2 clones M (mainline)
   aB       AB     file A is change in mainline
    |`---v  AB'    file B is changed in M2
    |   aB / |     M1 clones M
    |   ab/  |     M1 changes B
    |   ab'  |     M1 merges from M2, changes to B conflict
    |    |  A'B'   M2 changes A
     `---+--.|
         |  a'B'   M2 merges from mainline, changes to A conflict
         `--.|
            ???    depending on which ancestor we choose, we will
                   have to redo A hand-merge, B hand-merge, or both
                   but if we look at the files independently, the
                   merge is easy

The result is a merged version in the working directory, waiting for checkin.

Rollback

When committing or merging, Mercurial adds the changeset entry last. Mercurial keeps a transaction log of the name of each file touched and its length prior to the transaction. On abort, it truncates each file to its prior length. This simplicity is one benefit of making revlogs append-only. The transaction journal also allows an undo operation.

Merging between repositories

A key feature of Mercurial is its ability to merge between independent repositories in a decentralized fashion. Each repository can act as a read-only server or as a client. A client pulls from the server all branches that it has not seen and adds them to its graph. This pull is done in two steps:

1. Searching for new roots. This part begins by finding all new heads and searching backwards from those heads to the first unknown nodes in their respective branches. These nodes are the roots used to calculate the changegroup: the set of all changesets starting at those roots. Mercurial takes pains to make this search efficient in both bandwidth and round-trips.

2. Pulling a changegroup. Once the roots are found, the changegroup can be transferred as a single streaming transfer. This transfer is organized as an ordered set of deltas for changesets, manifests, and files. Large chunks of deltas can be directly added to the repository without unpacking so the pull is quick.