Fetching data

Overview of a fetch

Inside bzr, a typical fetch happens like this:

  • a user runs a command like bzr branch or bzr pull

  • Repository.fetch is called (by a higher-level method such as ControlDir.sprout, Branch.fetch, etc).

  • An InterRepository object is created. The exact implementation of InterRepository chosen depends on the format/capabilities of the source and target repos.

  • The source and target repositories are compared to determine which data needs to be transferred.

  • The repository data is copied. Often this is done by creating a StreamSource and StreamSink from the source and target repositories and feeding the stream from the source into the sink, but some InterRepository implementations do differently.

How objects to be transferred are determined

See InterRepository._walk_to_common_revisions. The basic idea is to do a breadth-first search in the source repository’s revision graph (starting from the head or heads the caller asked for), and look in the target repository to see if those revisions are already present. Eventually this will find the common ancestors in both graphs, and thus the set of revisions to be copied has been identified.

All inventories for the copied revisions need to be present (and all parent inventories at the stacking boundary too, to support stacking).

All texts versions introduced by those inventories need to be transferred (but see also stacking constraints).

Fetch specs

The most fetch methods accept a fetch_spec parameter. This is how the caller controls what is fetched: e.g. all revisions for a given head (that aren’t already present in the target), the full ancestry for one or more heads, or even the full contents of the source repository.

The fetch_spec parameter is an object that implements the interface defined by AbstractSearchResult in bzrlib.graph. It describes which keys should be fetched. Current implementations are SearchResult, PendingAncestryResult, EmptySearchResult, and EverythingResult. Some have options controlling if missing revisions cause errors or not, etc.

There are also some “search” objects, which can be used to conveniently construct a search result for common cases: EverythingNotInOther and NotInOtherForRevs. They provide an execute method that performs the search and returns a search result.

Also, Graph._make_breadth_first_searcher returns an object with a get_result method that returns a search result.


A stream is an iterable of (substream type, substream) pairs. The substream type is a str that will be one of texts, inventories, inventory-deltas, chk_bytes, revisions or signatures. A substream is a record stream. The format of those records depends on the repository format being streamed, except for inventory-deltas records which are format-independent.

A stream source can be constructed with repo._get_source(to_format), and it provides a get_stream(search) method (among others). A stream sink can be constructed with repo._get_sink(), and provides an insert_stream(stream, src_format, resume_tokens) method (among others).

Stacking constraints

In short the rule is: “repositories must hold revisions’ parent inventories and their new texts (or else all texts for those revisions).”

This is sometimes called “the stacking invariant.”

Why that rule?

A stacked repository needs to be capable of generating a complete stream for the revisions it does hold without access to its fallback repositories 1. “Complete” here means that the stream for a revision (or set of revisions) can be inserted into a repository that already contains the parent(s) of that revision, and that repository will have a fully usable copy of that revision: a working tree can be built for that revision, etc.

Assuming for a moment the stream has the necessary inventory, signature and CHK records to have a usable revision, what texts are required to have a usable revision? The simple way to satisfy the requirement is to have every text for every revision at the stacking boundary. Thus the revisions at the stacking boundary and all their descendants have their texts present and so can be fully reconstructed. But this is expensive: it implies each stacked repository much contain O(tree) data even for a single revision of a 1-line change, and also implies transferring O(tree) data to fetch that revision.

Because the goal is a usable revision when added to a repository with the parent revision(s) most of those texts will be redundant. The minimal set that is needed is just those texts that are new in the revisions in our repository. However, we need enough inventory data to be able to determine that set of texts. So to make this possible every revision must have its parent inventories present so that the inventory delta between revisions can be calculated, and of course the CHK pages associated with that delta. In fact the entire inventory does not need to be present, just enough of it to find the delta (assuming a repository format, like 2a, that allows only part of an inventory to be stored). Thus the stacked repository can contain only O(changes) data 2 and still deliver complete streams of that data.

What about revisions at the stacking boundary with more than one parent? All of the parent inventories must be present, as a client may ask for a stream up to any parent, not just the left-hand parent. If any parent is absent then all texts must be present instead. Otherwise there will be the strange situation where some fetches of a revision will succeed and others fail depending the precise details of the fetch.

Implications for fetching

Fetches must retrieve the records necessary to satisfy that rule. The stream source will attempt to send the necessary records, and the stream sink will check for any missing records and make a second fetch for just those missing records before committing the write group.

Our repository implementations check this constraint is satisfied before committing a write group, to prevent a bad stream from creating a corrupt repository. So a fetch from a bad source (e.g. a damaged repository, or a buggy foreign-format import) may trigger BzrCheckError during commit_write_group.

To fetch from a stacked repository via a smart server, the smart client:

  • first fetches a stream of as many of the requested revisions as possible from the initial repository,

  • then while there are still missing revisions and untried fallback repositories fetches the outstanding revisions from the next fallback until either all revisions have been found (success) or the list of fallbacks has been exhausted (failure).


This is not just a theoretical concern. The smart server always opens repositories without opening fallbacks, as it cannot assume it can access the fallbacks that the client can.


Actually O(changes) isn’t quite right in practice. In the current implementation the fulltext of a changed file must be transferred, not just a delta, so a 1-line change to a 10MB file will still transfer 10MB of text data. This is because current formats require records’ compression parents to be present in the same repository.