Repositories

Status

Date:2007-07-08

This document describes the services repositories offer and need to offer within brlib.

Motivation

To provide clarity to API and performance tradeoff decisions by centralising the requirements placed upon repositories.

Terminology

A repository is a store of historical data for bzr.

Command Requirements

Command Needed services
Add None
Annotate Annotated file texts, revision details
Branch Fetch, Revision parents, Inventory contents, All file texts
Bundle Maximally compact diffs (file and inventory), Revision graph difference, Revision texts.
Commit Insert new texts, insert new inventory via delta, insert revision, insert signature
Fetching Revision graph difference, ghost identification, stream data introduced by a set of revisions in some cheap form, insert data from a stream, validate data during insertion.
Garbage Collection Exclusive lock the repository preventing readers.
Revert Delta from working tree to historical tree, and then arbitrary file access to obtain the texts of differing files.
Uncommit Revision graph access.
Status Revision graph access, revision text access, file fingerprint information, inventory differencing.
Diff As status but also file text access.
Merge As diff but needs up to twice as many file texts - base and other for each changed file. Also an initial fetch is needed.
Log Revision graph (entire at the moment) access, sometimes status between adjacent revisions. Log of a file needs per-file-graph. Dominator caching or similar tools may be needed to prevent entire graph access.
Missing Revision graph access, and revision texts to show output.
Update As for merge, but twice.

Data access patterns

Ideally we can make our data access for commands such as branch to dovetail well with the native storage in the repository, in the common case. Doing this may require choosing the behaviour of some commands to allow us to have a smaller range of access patterns which we can optimise more heavily. Alternatively if each command is very predicable in its data access pattern we may be able to hint to the low level layers which pattern is needed on a per command basis to get efficient behaviour.

Command Data access pattern
Annotate-cached Find text name in an inventory, Recreate one text, recreate annotation regions
Annotate-on demand Find file id from name, then breadth-first pre-order traversal of versions-of-the-file until the annotation is complete.
Branch Fetch, possibly taking a copy of any file present in a nominated revision when it is validated during fetch.
Bundle Revision-graph as for fetch; then inventories for selected revision_ids to determine file texts, then mp-parent deltas for all determined file texts.
Commit Something like basis-inventories read to determine per-file graphs, insertion of new texts (which may be delta compressed), generation of annotation regions if the repository is configured to do so, finalisation of the inventory pointing at all the new texts and finally a revision and possibly signature.
Fetching Revision-graph searching to find the graph difference. Scan the inventory data introduced during the selected revisions, and grab the on disk data for the found file texts, annotation region data, per-file-graph data, piling all this into a stream.
Garbage Collection Basically a mass fetch of all the revisions which branches point at, then a bait and switch with the old repository thus removing unreferenced data.
Revert Revision graph access for the revision being reverted to, inventory extraction of that revision, dirblock-order file text extract for files that were different.
Uncommit Revision graph access to synthesise pending-merges linear access down left-hand-side, with is_ancestor checks between all the found non-left-hand-side parents.
Status Lookup the revisions added by pending merges and their commit messages. Then an inventory difference between the trees involved, which may include a working tree. If there is a working tree involved then the file fingerprint for cache-misses on files will be needed. Note that dirstate caches most of this making repository performance largely irrelevant: but if it was fast enough dirstate might be able to be simpler/
Diff As status but also file text access for every file that is different - either one text (working tree diff) or a diff of two (revision to revision diff).
Merge As diff but needs up to twice as many file texts - base and other for each changed file. Also an initial fetch is needed. Note that the access pattern is probably id-based at the moment, but that may be 'fixed' with the iter_changes based merge. Also note that while the texts from OTHER are the ones accessed, this is equivalent to the newest form of each text changed from BASE to OTHER. And as the repository looks at when data is introduced, this should be the pattern we focus on for merge.
Log Revision graph (entire at the moment) access, log of a file wants a per-file-graph. Log -v will want newest-first inventory deltas between revisions.
Missing Revision graph access, breadth-first pre-order.
Update As for merge, but twice.

Patterns used

Note that these are able to be changed by changing what we store. For instance if the repository satisfies mpdiff requests, then bundle can be defined in terms of mpdiff lookups rather than file text lookups appropriate to create mpdiffs. If the repository satisfies full text requests only, then you need the topological access to build up the desired mpdiffs.

Pattern Commands
Single file text annotate, diff
Files present in one revision branch
Newest form of files altered by revisions merge, update?
Topological access to file versions/deltas annotate-uncached
Stream all data required to recreate revs branch (lightweight)
Stream file texts in topological order bundle
Write full versions of files, inv, rev, sig commit
Write deltas of files, inv for one tree commit
Stream all data introduced by revs fetch
Regenerate/combine deltas of many trees fetch, pack
Reconstruct all texts and validate trees check, fetch
Revision graph walk fetch, pack, uncommit, annotate-uncached, merge, log, missing
Top down access multiple invs concurrently status, diff, merge?, update?
Concurrent access to N file texts diff, merge
Iteration of inventory deltas log -v, fetch?

Facilities to scale well

Indices

We want < linear access to all data in the repository. This suggests everything is indexed to some degree.

Often we know the kind of data we are accessing; which allows us to partition our indices if that will help (e.g. by reducing the total index size for queries that only care about the revision graph).

Indices that support our data access patterns will usually display increased locality of reference, reducing the impact of a large indices without needing careful page size management or other tricks.

We need repository wide indices. For the current repositories this is achieved by dividing the keyspace (revisions, signatures, inventories, per-fileid) and then having an append only index within each keyspace. For pack based repositories we will want some means to query the index of each component pack, presumably as a single logical index.

It would be nice if indexing was made cleanly separate from storage. So that suggests indices don't know the meaning of the lookup; indices which offer particular ordering, or graph walking facilities will clearly need that information, but perhaps they don't need to know the semantics ?

Index size

Smaller indexes are good. We could go with one big index, or a different index for different operation styles. As multiple indices will occupy more space in total we should consider carefully about adding indices.

Index ordering

Looking at the data access patterns some operations such as graph walking can clearly be made more efficient by offering direct iteration rather than repeated reentry into the index - so having indices that support iteration in such a style would be useful eventually.

Changing our current indexes

We can consider introducing cleaner indices in advance of a full pack based repository.

There are many possibilities for this, but I've chosen one that seems ok to me for illustration.

A key element is to consider when indices are updated. I think that the update style proposed for pack based repositories - write once, then when we group data again rewrite a new single index - is sufficent.

Replace .kndx

We could discard the per-knit .kndx by writing a new index at the end of every bzr transaction indexing the new data introduced by the bzr operation. e.g. at the end of fetch. This can be based on the new GraphIndex index type.

Encoding a knit entry into a GraphIndex can be done as follows:

  • Change the key to include a prefix of the knit name, to allow filtering out of data from different knits.
  • Encode the parents from the knit as the zeroth node reference list.
  • If the knit hunk was delta compressed encode the node it was delta compressed against as the 1st node reference list (otherwise the 1st node reference list will be empty to indicate no compression parents).
  • For the value encode similarly to the current knit format the byte offset for the data record in the knit, the byte length for the data record in the knit and the no-end-of-line flag.

Its important to note that knit repositories cannot be regenerated by scanning .knits, so a mapped index is still irreplaceable and must be transmitted on push/pull.

A potential improvement exists by specialising this further to not record data that is not needed - e.g. an index of revisions does not need to support a pointer to a parent compressed text as revisions.knit is not delta-compressed ever. Likewise signatures do not need the parent pointers at all as there is no 'signature graph'.

Moving to pack based repositories

We have a number of challenges to solve.

Naming of files

As long as the file name is unique it does not really matter. It might be interesting to have it be deterministic based on content, but there are no specific problems we have solved by doing that, and doing so would require hashing the full file. OTOH hashing the full file is a cheap way to detect bit-errors in transfer (such as windows corruption).

Discovery of files

With non-listable transports how should the collection of pack/index files be found ? Initially record a list of all the pack/index files from write actions. (Require writable transports to be listable). We can then use a heuristic to statically combine pack/index files later.

Merging data on push

A trivial implementation would be to make a pack which has just the data needed for the push, then send that. More sophisticated things would be streaming single-pass creation, and also using this as an opportunity to increase the packedness of the local repo.

Caching and writeing of data

Repositories try to provide a consistent view of the data within them within a 'lock context'.

Locks

Locks come in two flavours - read locks and write locks. Read locks allow data to be read from the repository. Write locks allow data to be read and signal that you intend to write data at some point. The actual writing of data must take place within a Write Group.

Write locks provide a cache of repository data during the period of the write lock, and allow write_groups to be acquired. For some repositories the presence of a write lock is exclusive to a single client, for others which are lock free or use server side locks (e.g. svn), the write lock simply provides the cache context.

Write Groups

Write groups are the only allowed means for inserting data into a repository. These are created by start_write_group, and concluded by either commit_write_group or abort_write_group. A write lock must be held on the repository for the entire duration. At most one write group can be active on a repository at a time.

Write groups signal to the repository the window during which data is actively being inserted. Several write groups could be committed during a single lock.

There is no guarantee that data inserted during a write group will be invisible in the repository if the write group is not committed. Specifically repositories without atomic insertion facilities will be writing data as it is inserted within the write group, and may not be able to revert that data - e.g. in the event of a dropped SFTP connection in a knit repository, inserted file data will be visible in the repository. Some repositories have an atomic insertion facility, and for those all-or-nothing will apply.

The precise meaning of a write group is format specific. For instance a knit based repository treats the write group methods as dummy calls, simply meeting the api that clients will use. A pack based repository will open a new pack container at the start of a write group, and rename it into place at commit time.