Bazaar Performance Roadmap

Contents

1   About the performance roadmap

1.1   What should be in the roadmap?

A good roadmap provides a place for contributors to look for tasks, it provides users with a sense of when we will fix things that are affecting them, and it also allows us all to agree about where we are headed. So the roadmap should contain enough things to let all this happen.

I think that it needs to contain the analysis work which is required, a list of the use cases to be optimised, the disk changes required, and the broad sense of the api changes required. It also needs to list the inter-dependencies between these things: we should aim for a large surface area of 'ready to be worked on' items, that makes it easy to improve performance without having to work in lockstep with other developers.

Clearly the analysis step is an immediate bottleneck - we cannot tell if an optimisation for use case A is a pessimism for use case B until we have analysed both A and B. I propose that we complete the analysis of say a dozen core use cases end to end during the upcoming sprint in London. We should then be able to fork() for much of the detailed design work and regroup with disk and api changes shortly thereafter.

I suspect that clarity of layering will make a big difference to developer parallelism, so another proposal I have is for us to look at the APIs for Branch and Repository in London in the light of what we have learnt over the last years.

1.2   What should the final system look like, how is it different to what we have today?

One of the things I like the most about bzr is its rich library API, and I've heard this from numerous other folk. So anything that will remove that should be considered a last resort.

Similarly our relatively excellent cross platform support is critical for projects that are themselves cross platform, and thats a considerable number these days.

And of course, our focus on doing the right thing is what differentiates us from some of the other VCS's, so we should be focusing on doing the right thing quickly :).

What we have today though has grown organically in response to us identifying bottlenecks over several iterations of back end storage, branch metadata and the local tree representation. I think we are largely past that and able to describe the ideal characteristics of the major actors in the system - primarily Tree, Branch, Repository - based on what we have learnt.

1.3   What use cases should be covered?

My list of use cases is probably not complete - its just the ones I happen to see a lot :). I think each should be analysed comprehensively so we dont need to say 'push over the network' - its implied in the scaling analysis that both semantic and file operation latency will be considered.

These use cases are ordered by roughly the ease of benchmarking, and the frequency of use. This ordering is so that when people are comparing bzr they are going to get use cases we have optimised; and so that as we speed things up our existing users will have the things they do the most optimised.

  • status tree
  • status subtree
  • commit
  • commit to a bound branch
  • incremental push/pull
  • log
  • log path
  • add
  • initial push or pull [both to a new repo and an existing repo with different data in it]
  • diff tree
  • diff subtree
  • revert tree
  • revert subtree
  • merge from a branch
  • merge from a bundle
  • annotate
  • create a bundle against a branch
  • uncommit
  • missing
  • update
  • cbranch

1.4   How is development on the roadmap coordinated?

I think we should hold regular get-togethers (on IRC) to coordinate on our progress, because this is a big task and its a lot easier to start helping out some area which is having trouble if we have kept in contact about each areas progress. This might be weekly or fortnightly or some such.

we need a shared space to record the results of the analysis and the roadmap as we go forward. Given that we'll need to update these as new features are considered, I propose that we use doc/design as a working space, and as we analyse use cases we include them in there - including the normal review process for each patch. We also need documentation about doing performance tuning - not the minutiae, though that is needed, but about how to effective choose things to optimise which will give the best return on time spent - that is what the roadmap should help with, but this looks to be a large project and an overview will be of great assistance I think. We want to help everyone that wishes to contribute to performance to do so effectively.

Finally, its important to note that coding is not the only contribution - testing, giving feedback on current performance, helping with the analysis are all extremely important tasks too and we probably want to have clear markers of where that should be done to encourage such contributions.

1.5   Planned changes to the bzr core

Delivering the best possible performance requires changing the bzr core design from that present in 0.16. Some of these changes are incremental and can be done with no impact on disk format. Many of them however do require changes to the disk format, and these can be broken into two sets of changes, those which are sufficiently close to the model bzr uses today to interoperate with the 0.16 disk formats, and those that are not able to interoperate with the 0.16 disk formats - specifically some planned changes may result in data which cannot be exported to bzr 0.16's disk formats and then imported back to the new format without losing critical information. If/when this takes place it will be essentially a migration for users to switch from their bzr 0.16 repository to a bzr that supports them. We plan to batch all such changes into one large 'experimental' repository format, which will be complete stable and usable before we migrate it to become a supported format. Getting new versions of bzr in widespread use at that time will be very important, otherwise the user base may be split in two - users that have upgraded and users that have not.

The following changes are grouped according to their compatability impact: library only, disk format but interoperable, disk format interoperability unknown, and disk format, not interoperable.

1.5.1   Library changes

These changes will change bzrlib's API but will not affect the disk format and thus do not pose a significant migration issue.

  • For our 20 core use cases, we plan to add targeted API's to bzrlib that are repository-representation agnostic. These will instead reflect the shape of data access most optimal for that case.
  • Deprecate 'versioned files' as a library concept. Instead of asking for information about a file-over-time as a special case, we will move to an API that assumes less coupling between the historical information and the ability to obtain texts/deltas etc. Specifically, we need to remove all API's that act in terms of on disk representation except those within a given repository implementation.
  • Create a validator for revisions that is more amenable to use by other parts of the code base than just the gpg signing facility. This can be done today without changing disk, possibly with a performance hit until the disk formats match the validatory logic. It will be hard to tell if we have the right routine for that until all the disk changes are complete, so while this is a library only change, its likely one that will be delayed to near the end of the process.
  • Add an explicit API for managing cached annotations. While annotations are considered a cache this is not exposed in such a way that cache operations like 'drop the cache' can be performed. On current disk formats the cache is mandatory, but an API to manage would allow refreshing of the cache (e.g. after ghosts are filled in in baz conversions).
  • Use the _iter_changes API to perform merges. This is a small change that may remove the need to use inventories in merge, making a dramatic difference to merge performance once the tree shape comparison optimisations are implemented.
  • Create a network-efficient revision graph API. This is the logic at the start of push and pull operations, which currently scales O(graph size). Fixing the scaling can be done, but there are tradeoffs to latency and performance to consider, making it a little tricky to get right.
  • Working tree disk operation ordering. We plan to change the order in which some operations are done (specifically TreeTransform ones) to improve performance. There is already a 66% performance boost in that area going through review.
  • Stop requiring full memory copies of files. Currently bzr requires that it can hold 3 copies of any file its versioning in memory. Solving this is tricky, particularly without performance regressions on small files, but without solving it versioning of .iso and other large objects will continue to be extremely painful.
  • Add an API for per-file graph access that alllows incremental access and is suitable for on-demand generation if desired.
  • Repository stacking API. Allowing multiple databases to be stacked to give a single 'repository' will allow implementation of some long desired features like history horizons, and bundle usage where the bundle is not added to the local repository just to examine its contents.
  • Revision data manipulation API. We need a single streaming API for adding data to or getting it from a repository. This will need to allow hints such as 'optimise for size', or 'optimise for fast-addition' to meet the various users planned, but it is a core part of the library today, and its not sufficiently clean to let us simplify/remove a lot of related code today.

1.5.2   Interoperable disk changes

  • New container format to allow single-file description of multiple named objects. This will provide the basis for transmission of revisions over the network, the new bundle format, and possibly a new repository format as well. [Core implemented]
  • Separate the annotation cache from the storage of actual file texts and make the annotation style, and when to do it, configurable. This will reduce data sent over the wire when repositories have had 'needs-annotations' turned off, which very large trees may choose to do - generating just-in-time annotations may be desirable for those trees (even when performing annotation based merges).
  • Repository disk operation ordering. The order that tasks access data within the repository and the layout of the data should be harmonised. This will require disk format changes but does not inherently alter the model, so its straight forward to export from a repository that has been optimised in this way to a 0.16 based repository.
  • Inventory representation. An inventory is a logical description of the shape of a version controlled tree. Currently we operate on the whole inventory as a tree broken down per directory, but we store it as a flat file. This scale very poorly as even a minor change between inventories requires us to scan the entire file, and in large trees this is many megabytes of data to consider. We are investigating the exact form, but the intent is to change the serialisation of inventories so that comparing two inventories can be done in some smaller time - e.g. O(log N) scaling. Whatever form this takes, a repository that can export it directly will be able to perform operations between two historical trees much more efficiently than the current repositories.
  • Delta storage optimisation. We plan to change the delta storage logic to use a binary delta like xdelta rather than using line based deltas from python. These binary deltas could be done along ancestry ordering, or other arbitrary patterns chosen for their intended use. Line based deltas will still be created for cached annotations. This is still under some discussion. http://bazaar-vcs.org/PerformanceRoadmap/Xdelta
  • Greatest distance from origin cache. This is a possible change to introduce, but it may be unnecessary - listed here for completeness till it has been established as [un]needed.

1.5.3   Possibly non-interoperable disk changes

  • Removing of derivable data from the core of bzr. Much of the data that bzr stores is derivable from the users source files. For instance the annotations that record who introduced a line. Given the full history for a repository we can recreate that at any time. We want to remove the dependence of the core of bzr on any data that is derivable, because doing this will give us the freedom to:

    • Improve the derivation algorithm over time.
    • Deal with bugs in the derivation algorithms without having 'corrupt repositories' or such things.

    However, some of the data that is technically derived, like the per-file merge graph, is both considered core, and can be generated differently when certain circumstances arive, by bzr 0.16. Any change to the 'core' status of that data will discard data that cannot be recreated and thus lead to the inability to export from a format where that is derived data to bzr 0.16's formats without errors occuring in those circumstances. Some of the data that may be considered for this includes:

    • Per file merge graphs
    • Annotations

1.5.4   Non-interoperable disk changes

  • Drop the per-file merge graph 'cache' currently held in the FILE-ID.kndx files. A specific case of removing derivable data, this may allow smaller inventory metadata and also make it easier to allow two different trees (in terms of last-change made, e.g. if one is a working tree) to be compared using a hash-tree style approach.
  • Use hash based names for some objects in the bzr database. Because it would force total-knowledge-of-history on the graph revision objects will not be namable via hash's and neither will revisio signatures. Other than that though we can in principle use hash's e.g. SHA1 for everything else. There are many unanswered questions about hash based naming related to locality of reference impacts, which need to be answered before this becomes a definite item.

1.6   Contributing to the performance drive

1.6.1   What needs doing?

There is plenty of code to write. Testers are always welcome for experimental changes. In general, pick a BLUE node from performance.png which has nothing pointing at it that is also BLUE, and start working on that.

Adhoc performance work can also be done. One useful tool is the 'evil' debug flag. For instance running bzr -Devil commit -m "test" will log a backtrace to the bzr log file for every method call which triggers a slow or non-scalable part of the bzr library. So checking that a given command with -Devil has no backtraces logged to the log file is a good way to find problem function calls that might be nested deep in the code base.

1.6.2   Status

The performance drive is well under way. At the moment we are finalising the analysis documents from the sprint at London. If you were not at that sprint, please read the analysis documents - consider them living documents much like code, to be patched and corrected.

If you were at the London performance sprint, please help finish documenting the core command analysis work. The core commands that have not been analysed are listed in performance.dot as the BLUE nodes. For quick reference:

  • status
  • log
  • diff
  • uncommit
  • missing
  • update
  • cbranch

Once a given command has had its analysis created, the bottleneck of 'folk who attended London' is removed. The next task is in general the creation of a targeted API stack for that command. This API stack is done by starting with the cmd object in builtins.py and cleaning up the code so that the API used there is one which allows an implementation matching the analysis document. This then gets repeated, iteratively, on each of the called API's, until all the current slow code is cleanly abstracted behind the Tree, Branch and Repository API's.

1.6.3   Resources

The usual resources for contributing to bzr - the mailing list, wiki, bug tracker and IRC channels are documented in the HACKING document. Additionally there are members of the project focusing on performance at the moment who are willing to mentor contributors on performance issues. Just send a mail to the list asking for mentoring on the step of the performance plan you want to help with.

1.7   Integration of performance changes

To deliver a version of bzr with all our planned changes will require significant integration work. Minimally each change needs to integrate with some aspect of the bzr version it's merged into, but in reality many of these changes while conceptually independent will in fact have to integrate with the other changes we have planned before can have a completed system.

Additionally changes that alter disk formats are inherently more tricky to integrate because we will often need to alter apis throughout the code base to expose the increased or reduced model of the preferred disk format.

The dot file performance.dot graphs out the dependencies to let us make accurate assessments of the changes needed in terms of code and API, hopefully minimising the number of different integration steps we have to take, while giving us a broad surface area for development. Its based on a sumary in the next section of this document of the planned changes with their expected collaborators and dependencies. Where a command is listed, the expectation is that all uses of that command - local, remote, dumb transport and smart transport are being addressed together.

The following provides a summary of the planned changes and their expected collaborators within the code base, along with an estimate of whether they are likely to require changes to their collaborators to be considered 'finished'.

  • Use case target APIs: Each of these is likely to alter the Tree interface. Some few of them focus on Branch and will alter Branch and Repository accordingly. As they are targeted APIs we can deep changes all the way down the stack to the underlying representation to make it all fit well. Presenting a top level API for many things will be possible now as long as the exposed data is audited for things we plan to make optional, or remove: Such things cannot be present in the final API. Writing these APIs now will provide strong feedback to the design process for those things which are considered optional or removable, so these APIs should be implemented before removing or making optional existing data.
  • Deprecating versioned files as a supported API: This collaborates with the Repository API but can probably be done by adding a replacement API for places where the versioned-file api is used. We may well want to keep a concept of 'a file over time' or 'inventories over time', so the existing repository model of exposing versioned file objects may be ok; what we need to ensure we do is remove the places in the code base where you create or remove or otherwise describe manipulation of the storage by knit rather than talking at the level of file ids and revision ids. The current versioned-file API would be a burden for implementors of a blob based repository format, so the removal of callers, and deprecation of those parts of the API should be done before creating a blob based repository format.
  • Creating a revision validator: Revision validators may depend on storage layer changes to inventories so while we can create a revision validator API, we cannot create the final one until we have the inventory structural changes completed.
  • Annotation caching API: This API is a prerequisite for new repository formats. If written after they are introduced we may find that the repository is lacking in functionality, so the API should be implemented first.
  • _iter_changes based merging: If the current _iter_changes_ API is insufficient, we should know about that before designing the disk format for generating fast _iter_changes_ output.
  • Network-efficient revision graph API: This influences what questions we will want to ask a local repository very quickly; as such it's a driver for the new repository format and should be in place first if possible. Its probably not sufficiently different to local operations to make this a hard ordering though.
  • Working tree disk ordering: Knowing the expected order for disk operations may influence the needed use case specific APIs, so having a solid understanding of what is optimal - and why - and whether it is pessimal on non linux platforms is rather important.
  • Be able to version files greater than memory in size: This cannot be achieved until all parts of the library which deal with user files are able to provide access to files larger than memory. Many strategies can be considered for this - such as temporary files on disk, memory mapping etc. We should have enough of a design laid out that developers of repository and tree logic are able to start exposing apis, and considering requirements related to them, to let this happen.
  • Per-file graph access API: This should be implemented on top of or as part of the newer API for accessing data about a file over time. It can be a separate step easily; but as it's in the same area of the library should not be done in parallel.
  • Repository stacking API: The key dependency/change required for this is that repositories must individually be happy with having partial data - e.g. many ghosts. However the way the API needs to be used should be driven from the command layer in, because its unclear at the moment what will work best.
  • Revision stream API: This API will become clear as we streamline commands. On the data insertion side commit will want to generate new data. The commands pull, bundle, merge, push, possibly uncommit will want to copy existing data in a streaming fashion.
  • New container format: Its hard to tell what the right way to structure the layering is. Probably having smooth layering down to the point that code wants to operate on the containers directly will make this more clear. As bundles will become a read-only branch & repository, the smart server wants streaming-containers, and we are planning a pack based repository, it appears that we will have three different direct container users. However, the bundle user may in fact be fake - because it really is a repository.
  • Separation of annotation cache: Making the disk changes to achieve this depends on the new API being created. Bundles probably want to be annotation-free, so they are a form of implementation of this and will need the on-demand annotation facility.
  • Repository operation disk ordering: Dramatically changing the ordering of disk operations requires a new repository format. We have most of the analysis done to be able to specify the desired ordering, so it should be possible to write such a format now based on the container logic, but without any of the inventory representation or delta representation changes. This would for instance involve pack combining ordering the existing diffs in reverse order.
  • Inventory representation: This has a dependency on what data is dropped from the core and what is kept. Without those changes being known we can implement a new representation, but it won't be a final one. One of the services the new inventory representation is expected to deliver is one of validators for subtrees -- a means of comparing just subtrees of two inventories without comparing all the data within that subtree.
  • Delta storage optimisation: This has a strict dependency on a new repository format. Optimisation takes many forms - we probably cannot complete the desired optimisations under knits though we could use xdelta within a knit-variation.
  • Greatest distance from origin cache: The potential users of this exist today, it is likely able to be implemented immediately, but we are not sure that its needed anymore, so it is being shelved.
  • Removing derivable data: Its very hard to do this while the derived data is exposed in API's but not used by commands. Implemented the targeted API's for our core use cases should allow use to remove accidental use of derived data, making only explicit uses of it visible, and isolating the impact of removing it : allowing us to experiment sensibly. This covers both dropping the per-file merge graph and the hash-based-names proposals.

2   Analysis of use cases

2.1   Analysing a specific use case

The analysis of a use case needs to provide as outputs:
  • The functional requirements that the use case has to satisfy.
  • The file level operations and access patterns that will give the best performance.
  • A low friction API which will allow the use case to be implemented.
  • The release of bzr (and thus the supported features) for which the analysis was performed. The feature set of bzr defines the access patterns and data required to implement any use case. So when we add features, their design changes the requirements for the parts of the system they alter, so we need to re-analyse use cases when bzr's feature set changes. If future plans are considered in the analysis with the intention of avoiding rework, these should also be mentioned.

2.2   Performing the analysis

The analysis needs to be able to define the characteristics of the involved disk storage and APIs. That means we need to examine the data required for the operation, in what order it is required, on both the read and write sides, and how that needs to be presented to be consistent with our layering.

As a quick example: 'annotation of a file requires the file id looked up from the tree, the basis revision id from the tree, and then the text of that fileid-revisionid pair along with the creating revision id allocated to each line, and the dotted revision number of each of those revision ids.' All three of our key domain objects are involved here, but we haven't defined any characteristics of the api or disk facilities yet. We could then do that by saying something like 'the file-id lookup should degrade gracefully as trees become huge. The tree basis id should be constant time. Retrieval of the annotated text should be roughly constant for any text of the same size regardless of the number of revisions contributing to its content. Mapping of the revision ids to dotted revnos could be done as the text is retrieved, but its completely fine to post-process the annotated text to obtain dotted-revnos.'

2.3   What factors should be considered?

Obviously, those that will make for an extremely fast system :). There are many possible factors, but the ones I think are most interesting to design with are:

  • baseline overhead:

    • The time to get bzr ready to begin the use case.
  • scaling: how does performance change when any of the follow aspects of the system are ratcheted massively up or down:

    • number of files/dirs/symlinks/subtrees in a tree (both working and revision trees)
    • size of any particular file
    • number of elements within a single directory
    • length of symlinks
    • number of changes to any file over time (subordinately also the number of merges of the file)
    • number of commits in the ancestry of a branch (subordinately also the number of merges)
    • number of revisions in a repository
    • number of fileids in a repository
    • number of ghosts in a given graph (revision or per-file)
    • number of branches in a repository
    • number of concurrent readers for a tree/branch/repository
    • number of concurrent writers for objects that support that.
    • latency to perform file operations (e.g. slow disks, network file systems, our VFS layer and FTP/SFTP/etc)
    • bandwidth to the disk storage
    • latency to perform semantic operations (hpss specific)
    • bandwidth when performing semantic operations.
  • locality of reference: If an operation requires data that is located within a small region at any point, we often get better performance than with an implementation of the same operation that requires the same amount of data but with a lower locality of reference. Its fairly tricky to add locality of reference after the fact, so I think its worth considering up front.

Using these factors, to the annotate example we can add that its reasonable to do two 'semantic' round trips to the local tree, one to the branch object, and two to the repository. In file-operation level measurements, in an ideal world there would be no more than one round trip for each semantic operation. What there must not be is one round trip per revision involved in the revisionid->dotted number mapping, nor per each revision id attributed to a line in the text.

Not all the items mentioned above are created equal. The analysis should include the parameters considered and the common case values for each - the optimisation should be around the common cases not around the exceptions.

For instance, we have a smart server now; file level operations are relatively low latency and we should use that as the common case. At this point we intend to preserve the performance of the dumb protocol networking, but focus on improving network performance via the smart server and thus escape the file-level operation latency considerations.

Many performance problems only become visible when changing the scaling knobs upwards to large trees. On small trees its our baseline performance that drives incremental improvements; on large trees its the amount of processing per item that drives performance. A significant goal therefore is to keep the amount of data to be processed under control. Ideally we can scale in a sublinear fashion for all operations, but we MUST NOT scale even linearly for operations that invoke a latency multiplier. For example, reading a file on disk requires finding the inode for the file, then the block with the data and returning the contents. Due to directory grouping logic we pay a massive price to read files if we do not group the reads of files within the same directory.

3   Use cases

3.1   Initial push / pull

3.1.1   Optimal case

(a motivating example of ultimate performance) Assume there is a file with exactly the right data in compressed form. This may be a tarred branch, a bundle, or a blob format. Performance in this case scales with the size of the file.

3.1.2   Disk case

Assume current repo format. Attempt to achieve parity with cp -r. Read each file only 1 time.

  • read knit graph for revisions
  • write filtered copy of revision knit O(d+a)
  • write filtered copy of knit index O(d)
  • Open knit index for inventory
  • Write a filtered copy of inventory knit and simultaneously not all referenced file-ids O(b+d)
  • Write filtered copy of inventory knit index O(d)
  • For each referenced file-id:
    • Open knit index for each file knit O(e)
    • If acceptable threshold of irrelevant data hard-link O(f)
    • Otherwise write filtered copy of text knit and simultaneously write the fulltext to tree transform O(h)
  • Write format markers O(1)
a:size of aggregate revision metadata
b:size of inventory changes for all revisions
c:size of text changes for all files and all revisions (e * g)
d:number of relevant revisions
e:number of relevant versioned files
f:size of the particular versioned file knit index
g:size of the filtered versioned file knit
h:size of the versioned file fulltext
i:size of the largest file fulltext

3.1.3   Smart Network Case

3.1.3.1   Phase 1

Push: ask if there is a repository, and if not, what formats are okay Pull: Nothing

3.1.3.2   Phase 2

Push: send initial push command, streaming data in acceptable format, following disk case strategy Pull: receive initial pull command, specifying format

Pull client complexity: O(a), memory cost O(1) Push client complexity: procesing and memory cost same as disk case

3.1.4   Dumb Network Case

Pull: same as disk case, but request all file knit indices at once and request al file knits at once. Push: same as disk case, but write all files at once.

3.1.5   Wants

  • Read partial graph
  • Read multiple segments of multiple files on http and sftp
  • Write multiple files over SFTP

3.2   Incremental push/pull

This use case covers pulling in or pushing out some number of revisions which is typically a small fraction of the number already present in the target repository. Pushing and pulling are defined as branch level operations for ease of interaction with VCS systems that have no repository abstraction (such as bzr-svn or GNU Arch) but within bzrlib's core they are currently the responsibility of the Repository object.

3.2.1   Functional Requirements

A push or pull operation must:
  • Copy all the data to reconstruct the selected revisions in the target branch. This is the goal of push and pull after all.
  • Reject corrupt data. As bzr has no innate mechanism for discarding corrupted data, corrupted data should not be incorporated accidentally.

3.2.2   Factors which should add work for push/pull

  • Baseline overhead: The time to connect to both branches.
  • Actual new data in the revisions being pulled (drives the amount of data to move around, includes the commit messages etc)
  • Number of revisions in the two repositories (scaling affects the determination of what revisions to move around).

3.2.3   Push/pull overview

  1. New data is identified in the source repository.
  2. That data is read from the source repository.
  3. The same data is verified and written to the target repository in such a manner that its not visible to readers until its ready for use.

3.2.3.1   New data identification

We have a single top level data object: revisions. Everything else is subordinate to revisions, so determining the revisions to propogate should be all thats needed. This depends on revisions with partial data - such as those with no signature - being flagged in some efficient manner.

We could do this in two manners: determine revisions to sync and signatures to sync in two passes, or change the 'value' of a revision implicitly when the signature is different. E.g. by using merkle hash trees with the signature data a separate component the signatures will naturally be identified to sync.

We want to only exchange data proportional to the number of new revisions and signatures in the system though. One way to achieve this for revisions is to walk the graph out from the desired tips until the surface area intersection is found. For signatures a set difference seems to be needed as there is no DAG of signatures: the presence of one has no implications on the presence of another, so a full pass over the set of signatures would be required to confirm no new signatures are needed (let alone replaced signatures).

IFF we can determine 'new revisions' and 'new signatures' without full graph access then we can scale acceptable for push and pull.

Ghosts are revisions which are not present in a particular repository. Filling ghosts refers to removing ghosts in the target repository when the ghost is present in the source repository. Filling ghosts can be either an explicit or implicit action. The common case is no ghosts.

3.2.3.1.1   Set synchronisation approaches

A set synchronisation approach is one which synchronises two sets without regard for innate structure. This can be very efficient but requires adding a new node to be processed with every commit. Caching of the results of the various set based syncs I've seen is possible but because the data structures look different depending on the tip revision being synced up to the cache needs to be very complex. I recommend not using such an approach for the common case pull because of the failure to scale. We can use such an approach for synchronisation of new signatures and ghosts, which should be an explicit option in both cases.

3.2.3.1.2   DAG synchronisation approaches

A DAG based approach to synchronistion is one that uses the DAG structure to determine the difference in present nodes. It can as a result operate from the tip of the DAG backwards. A dag based approach should allow incremental access to data and not require a full-graph scan for incremental operations.

3.2.3.1.3   File level scaling

We should read roughly as much of the revision level graph as is needed from each repository to determine the node difference. If requested we should perform a detailed scan to pick up ghost revisions and revisions which have had signatures added. This should not be the default as it requires full history access in both cases.

Expected file IO and access pattern:

  • Common case: repo with many branches of one project, to the same.

    1. Source and Target branch tips read.
    2. Find the tip of each branch in their repo (will require reading some of the revision graph but is typically near the end of the graph).
    3. Read and parse increasing amounts of the revision graph until one is found to be a subset of the other, or a complete list of revisions to be transmitted is created.
  • Uncommon cases:

    1. Repositories with many projects or branches which are very old may require reading a lot of unrelated graph data.
    1. Initial push/pull scenarios should not require reading an entire graph.
3.2.3.1.4   API scaling
  1. Get branch tips.
  2. Determine one sided graph difference. To avoid obtaining a full graph over the wire this needs to be done without reference to the full graph, and with some logarthmic scaling algorithm. There are several already available for this.

With ghost and new-signature detection:

  • File IO access pattern will read the entire graph on the 'target' side - if no ghosts are present then stop, otherwise seek the new revisions on the source side with the regular algorithm and also explicitly search for the ghost points from the target; plus a set difference search is needed on signatures.
  • Semantic level can probably be tuned, but as its also complex I suggest deferring analysis for optimal behaviour of this use case.

3.2.3.2   Data reading

When transferring information about a revision the graph of data for the revision is walked: revision -> inventory, revision -> matching signature, inventory -> file ids:revision pairs.

3.2.3.2.1   File level scaling

As we're reading already committed data, as long as nothing is mutating data on disk reading should be race free. We will:

  • read each revision object
  • read the matching inventory delta
  • attempt to read a signature object
  • parse the inventory delta
  • read the fileid:revisionid compressed chunk for each line in the inventory delta

Theres no point validating that the data read is valid, as transmission through to the client writing the data might invalidate it; we need to validate before we write.

3.2.3.2.2   API scaling

Given that we have established the revisions needed, a single API call should suffice to obtain all data; the API should present the data in such an order that it can be validated as it arrives and thus not require large scale buffering on disk. Specifically each item of data should be validatable (e.g. for some file data we want the fileid:revisionid:validationhash + content).

3.2.3.3   Data Verification and writing

New data written to a repository should be completed intact when it is made visible. This suggests that either all the data for a revision must be made atomically visible (e.g. by renaming a single file) or the leaf nodes of the reference graph must become visible first.

Data is referred to via the following graph: revision -> revision revision -> signature revision -> inventory inventory -> fileid:revisionid fileid:revisionid -> fileid:revisionid

Data is verifiable via a different ordering: signature -> revision -> inventory -> fileid:revisionid texts.

We dont gpg verify each revision today; this analysis only speaks to hash verification of contents.

To validate a revision we need to validate the data it refers to. But to validate the contents of a revision we need the new texts in the inventory for the revision - to check a fileid:revisionid we need to know the expected sha1 of the full text and thus also need to read the delta chain to construct the text as we accept it to determine if its valid. Providing separate validators for the chosen representation would address this. e.g: For an inventory entry FILEID:REVISIONID we store the validator of the full text :SHA1:. If we also stored the validator of the chosen disk representation (:DELTASHA1:) we could validate the transmitted representation without expanding the delta in the common case. If that failed we could expand the delta chain and try against the full text validator, and finally fail. As different delta generators might generate different deltas, :DELTASHA1: should not become part of the revision validator, only the inventory disk encoding. In a related manner a transmission format that allowed cheap validation of content without applying locally stored deltas would be advantageous because no local reads would be incurred to validate new content. For instance, always sending a full text for any file, possibly with a delta-chain when transmitting multiple revisionids of the file, would allow this. (git pack-files have this property).

3.2.3.3.1   Overview summary

A single-file local format would allow safe atomic addition of data while allowing optimisal transmission order of data. Failing this the validation of data should be tuned to not require reading local texts during data addition even in the presence of delta chains. We should have transmission-validators separate from content validators that allow validation of the delta-transmitted form of objects.

3.2.3.3.2   File level scaling
  • Every new file text requires transmission and local serialisation.
  • Every commit requires transmission and storage of a revision, signature and inventory.

Thus 4000 commits to a 50000 path tree of 10 files on averages requires (with knits) between 26 writes (2*(3+10)) and 80006 (2*(4000*10 + 3)) writes. In all cases there are 4000 * 13 distinct objects to record.

Grouping data by fileid, content and metadata, gives the figures above. Data grouping:

  • File per full identifier (fileid:revisionid:meta|content): 104000
  • Delta-chain per object: object id count * constant overhead per object id (26 -> 80006)
  • Collation/pack file: 1
Performance for these depends heavily on implementation:
  • Using full ids we could name by validator or by id, giving best performance that depends on either receiving data in validator order or in id order.
  • using delta-chain per object we get least seek overhead and syscall overhead if we recieve in topological order within the object id, and object ids in lexical order.
  • Using a collation/pack file we can stream it into place and validate as we go, giving near ideal performance.
3.2.3.3.3   API scaling

The api for writing new data recieved over the network will need to be geared to the transmission and local storage method. What we need is for the transmission method to reasonably closely match the desired write ordering locally. This suggests that once we decide on the best local storage means we should design the api.

take N commits from A to B, if B is local then merge changes into the tree. copy ebough data to recreate snapshots avoid ending up wth corrupt/bad data

3.2.4   Notes from London

  1. setup

look at graph of revisions for ~N comits to deretmine eligibility for if preserve mainline is on, check LH only

identify objects to send that are not on the client repo
  • revision - may be proportional to the graph
  • inventory - proportional to work
  • texts - proportional to work
  • signatures - ???
  1. data transmission
  • send data proportional to the new information
  • validate the data:
  1. validate the sha1 of the full text of each transmitted text.
  2. validate the sha1:name mapping in each newly referenced inventory item.
  3. validate the sha1 of the XML of each inventory against the revision. this is proportional to tree size and must be fixed
  1. write the data to the local repo. The API should output the file texts needed by the merge as by product of the transmission
  2. tree application

Combine the output from the transmission step with additional 'new work data' for anything already in the local repository that is new in this tree. should write new files and stat existing files proportional to the count of the new work and the size of the full texts.

3.3   Add

Add is used to recursively version some paths supplied by the user. Paths that match ignore rules are not versioned, and paths that become versioned are versioned in the nearest containing bzr tree. Currently we only do this within a single tree, but perhaps with nested trees this should change.

3.3.1   Least work we can hope to perform

  • Read a subset of the full versioned paths data for the tree matching the scope of the paths the user supplied.
  • Seek once to each directory within the scope and readdir its contents.
  • Probe if each directory is a child tree to avoid adding data for paths within a child tree.
  • Calculate the ignored status for paths not previously known to be ignored
  • Write data proportional to the newly versioned file count to record their versioning.
  • Assign a fileid for each path (so that merge --uncommitted can work immediately)

Optionally:

  • Print the ignore rule for each ignored path in the scope.
  • Print the path of each added file.
  • Print the total count of ignored files within the scopes.
  • Record the result of calculating ignored status for ignored files. (proportional to the number we actually calculate).

3.3.2   Per file algorithm

  1. If the path is versioned, and it is a directory, push onto the recurse stack.
  2. If the path is supplied by the user or is not ignored, version it, and if a directory, push onto the recurse stack. Versioning the path may require versioning the paths parents.
  3. Output or otherwise record the ignored rule as per the user interface selected.

3.4   Commit Performance Notes

3.4.1   Changes to commit

We want to improve the commit code in two phases.

Phase one is to have a better separation from the format-specific logic, the user interface, and the general process of committing.

Phase two is to have better interfaces by which a good workingtree format can efficiently pass data to a good storage format. If we get phase one right, it will be relatively easy and non-disruptive to bring this in.

3.4.2   Commit: The Minimum Work Required

Here is a description of the minimum work that commit must do. We want to make sure that our design doesn't cost too much more than this minimum. I am trying to do this without making too many assumptions about the underlying storage, but am assuming that the ui and basic architecture (wt, branch, repo) stays about the same.

The basic purpose of commit is to:

  1. create and store a new revision based on the contents of the working tree
  2. make this the new basis revision for the working tree

We can do a selected commit of only some files or subtrees.

The best performance we could hope for is: - stat each versioned selected working file once - read from the workingtree and write into the repository any new file texts - in general, do work proportional to the size of the shape (eg inventory) of the old and new selected trees, and to the total size of the modified files

In more detail:

1.0 - Store new file texts: if a versioned file contains a new text there is no avoiding storing it. To determine which ones have changed we must go over the workingtree and at least stat each file. If the file is modified since it was last hashed, it must be read in. Ideally we would read it only once, and either notice that it has not changed, or store it at that point.

On the other hand we want new code to be able to handle files that are larger than will fit in memory. We may then need to read each file up to two times: once to determine if there is a new text and calculate its hash, and again to store it.

1.1 - Store a tree-shape description (ie inventory or similar.) This describes the non-file objects, and provides a reference from the Revision to the texts within it.

1.2 - Generate and store a new revision object.

1.3 - Do delta-compression on the stored objects. (git notably does not do this at commit time, deferring this entirely until later.) This requires finding the appropriate basis for each modified file: in the current scheme we get the file id, last-revision from the dirstate, look into the knit for that text, extract that text in total, generate a delta, then store that into the knit. Most delta operations are O(n**2) to O(n**3) in the size of the modified files.

1.4 - Cache annotation information for the changes: at the moment this is done as part of the delta storage. There are some flaws in that approach, such as that it is not updated when ghosts are filled, and the annotation can't be re-run with new diff parameters.

2.1 - Make the new revision the basis for the tree, and clear the list of parents. Strictly this is all that's logically necessary, unless the working tree format requires more work.

The dirstate format does require more work, because it caches the parent tree data for each file within the working tree data. In practice this means that every commit rewrites the entire dirstate file - we could try to avoid rewriting the whole file but this may be difficult because variable-length data (the last-changed revision id) is inserted into many rows.

The current dirstate design then seems to mean that any commit of a single file imposes a cost proportional to the size of the current workingtree. Maybe there are other benefits that outweigh this. Alternatively if it was fast enough for operations to always look at the original storage of the parent trees we could do without the cache.

2.2 - Record the observed file hashes into the workingtree control files. For the files that we just committed, we have the information to store a valid hash cache entry: we know their stat information and the sha1 of the file contents. This is not strictly necessary to the speed of commit, but it will be useful later in avoiding reading those files, and the only cost of doing it now is writing it out.

In fact there are some user interface niceties that complicate this:

3 - Before starting the commit proper, we prompt for a commit message and in that commit message editor we show a list of the files that will be committed: basically the output of bzr status. This is basically the same as the list of changes we detect while storing the commit, but because the user will sometimes change the tree after opening the commit editor and expect the final state to be committed I think we do have to look for changes twice. Since it takes the user a while to enter a message this is not a big problem as long as both the status summary and the commit are individually fast.

4 - As the commit proceeds (or after?) we show another status-like summary. Just printing the names of modified files as they're stored would be easy. Recording deleted and renamed files or directories is more work: this can only be done by reference to the primary parent tree and requires it be read in. Worse, reporting renames requires searching by id across the entire parent tree. Possibly full reporting should be a default-off verbose option because it does require more work beyond the commit itself.

5 - Bazaar currently allows for missing files to be automatically marked as removed at the time of commit. Leaving aside the ui consequences, this means that we have to update the working inventory to mark these files as removed. Since as discussed above we always have to rewrite the dirstate on commit this is not substantial, though we should make sure we do this in one pass, not two. I have previously proposed to make this behaviour a non-default option.

We may need to run hooks or generate signatures during commit, but they don't seem to have substantial performance consequences.

If one wanted to optimize solely for the speed of commit I think hash-addressed file-per-text storage like in git (or bzr 0.1) is very good. Remarkably, it does not need to read the inventory for the previous revision. For each versioned file, we just need to get its hash, either by reading the file or validating its stat data. If that hash is not already in the repository, the file is just copied in and compressed. As directories are traversed, they're turned into texts and stored as well, and then finally the revision is too. This does depend on later doing some delta compression of these texts.

Variations on this are possible. Rather than writing a single file into the repository for each text, we could fold them into a single collation or pack file. That would create a smaller number of files in the repository, but looking up a single text would require looking into their indexes rather than just asking the filesystem.

Rather than using hashes we can use file-id/rev-id pairs as at present, which has several consequences pro and con.

3.4.3   Commit vs Status

At first glance, commit simply stores the changes status reports. In fact, this isn't technically correct: commit considers some files modified that status does not. The notes below were put together by John Arbash Meinel and Aaron Bentley in May 2007 to explain the finer details of commit to Ian Clatworthy. They are recorded here as they are likely to be useful to others new to Bazaar ...

  1. Unknown files have a different effect. With --no-strict (the default) they have no effect and can be completely ignored. With --strict they should cause the commit to abort (so you don't forget to add the two new test files that you just created).
  2. Multiple parents. 'status' always compares 2 trees, typically the last-committed tree and the current working tree. 'commit' will compare more trees if there has been a merge.
  1. The "last modified" property for files. A file may be marked as changed since the last commit, but that change may have come in from the merge, and the change could have happened several commits back. There are several edge cases to be handled here, like if both branches modified the same file, or if just one branch modified it.
  2. The trickier case is when a file appears unmodified since last commit, but it was modified versus one of the merged branches. I believe there are a few ways this can happen, like if a merged branch changes a file and then reverts it back (you still update the 'last modified' field). In general, if both sides disagree on the 'last-modified' flag, then you need to generate a new entry pointing 'last-modified' at this revision (because you are resolving the differences between the 2 parents).
  1. Automatic deletion of 'missing' files. This is a point that we go back and forth on. I think the basic idea is that 'bzr commit' by default should abort if it finds a 'missing' file (in case that file was renamed rather than deleted), but 'bzr commit --auto' can add unknown files and remove missing files automatically.

  2. sha1 for newly added files. status doesn't really need this: it should only care that the file is not present in base, but is present now. In some ways commit doesn't care either, since it needs to read and sha the file itself anyway.

  3. Nested trees. status doesn't recurse into nested trees, but commit does. This is just because not all of the nested-trees work has been merged yet.

    A tree-reference is considered modified if the subtree has been committed since the last containing-tree commit. But commit needs to recurse into every subtree, to ensure that a commit is done if the subtree has changed since its last commit. _iter_changes only reports on tree-references that are modified, so it can't be used for doing subtree commits.

3.4.4   Avoiding Work: Smarter Change Detection

Commit currently walks through every file building an inventory. Here is Aaron's brain dump on a better way ...

_iter_changes won't tell us about tree references that haven't changed, even if those subtrees have changed. (Unless we ask for unchanged files, which we don't want to do, of course.)

There is an iter_references method, but using it looks just as expensive as calling kind().

I did some work on updating commit to use iter_changes, but found for multi-parent trees, I had to fall back to the slow inventory comparison approach.

Really, I think we need a call akin to iter_changes that handles multiple parents, and knows to emit entries when InventoryEntry.revision is all that's changed.

3.4.5   Avoiding Work: Better Layering

For each file, commit is currently doing more work than it should. Here is John's take on a better way ...

Note that "_iter_changes" does have to touch every path on disk, but it just can do it in a more efficient manner. (It doesn't have to create an InventoryEntry for all the ones that haven't changed).

I agree with Aaron that we need something a little different than _iter_changes. Both because of handling multiple parents, as well as we don't want it to actually read the files if we have a stat-cache miss.

Specifically, the commit code has to read the files because it is going to add the text to the repository, and we want it to compute the sha1 at that time, so we are guaranteed to have the valid sha (rather than just whatever the last cached one was). So we want the code to return 'None' if it doesn't have an up-to-date sha1, rather than reading the file and computing it, just before it returns it to the parent.

The commit code (0.16) should really be restructured. It's layering is pretty wrong.

Specifically, calling "kind()" requires a stat of the file. But we have to do a stat to get the size/whether the record is up-to-date, etc. So we really need to have a "create_an_up_to_date_inventory()" function. But because we are accessing every object on disk, we want to be working in tuples rather than Inventory objects. And because DirState already has the parent records next to the current working inventory, it can do all the work to do really fast comparison and throw-away of unimportant records.

The way I made "bzr status" fast is by moving the 'ignore this record' ability as deep into the stack as I could get. Status has the property that you don't care about most of the records, just like commit. So the sooner you can stop evaluating the 99% that you don't care about, the less work you do.

3.4.6   Avoiding work: avoiding reading parent data

We would like to avoid the work of reading any data about the parent revisions. We should at least try to avoid reading anything from the repository; we can also consider whether it is possible or useful to hold less parent information in the working tree.

When a commit of selected files is requested, the committed snapshot is a composite of some directories from the parent revision and some from the working tree. In this case it is logically necessary to have the parent inventory information.

If file last-change information or per-file graph information is stored then it must be available from the parent trees.

If the Branch's storage method does delta compression at commit time it may need to retrieve file or inventory texts from the repository.

It is desirable to avoid roundtrips to the Repository during commit, particularly because it may be remote. If the WorkingTree can determine by itself that a text was in the parent and therefore should be in the Repository that avoids one roundtrip per file.

There is a possibility here that the parent revision is not stored, or not correctly stored, in the repository the tree is being committed into, and so the committed tree would not be reconstructable. We could check that the parent revision is present in the inventory and rely on the invariant that if a revision is present, everything to reconstruct it will be present too.

3.4.7   Code structure

Caller starts a commit

>>> Branch.commit(from_tree, options)

This creates a CommitBuilder object matched to the Branch, Repository and Tree. It can vary depending on model differences or by knowledge of what is efficient with the Repository and Tree. Model differences might include whether no-text-change merges need to be reported, and whether the

The basic CommitBuilder.commit structure can be

  1. Ask the branch if it is ready to commit (up to date with master if any.)
  2. Ask the tree if it is ready to commit to the branch (up to date with branch?), no conflicts, etc
  3. Commit changed files; prototype implementation:
    1. Ask the working tree for all committable files; for each it should return the per-file parents, stat information, kind, etc.
    2. Ask the repository to store the new file text; the repository should return the stored sha1 and new revision id.
  4. Commit changed inventory
  5. Commit revision object

3.4.8   Complications of commit

Bazaar (as of 0.17) does not support selective-file commit of a merge; this could be done if we decide how it should be recorded - is this to be stored as an overall merge revision; as a preliminary non-merge revisions; or will the per-file graph diverge from the revision graph.

There are several checks that may cause the commit to be refused, which may be activated or deactivated by options.

  • presence of conflicts in the tree
  • presence of unknown files
  • the working tree basis is up to date with the branch tip
  • the local branch is up to date with the master branch, if there is one and --local is not specified
  • an empty commit message is given,
  • a hook flags an error
  • a "pointless" commit, with no inventory changes

Most of these require walking the tree and can be easily done while recording the tree shape. This does require that it be possible to abort the commit after the tree changes have been recorded. It could be ok to either leave the unreachable partly-committed records in the repository, or to roll back.

Other complications:

  • when automatically adding new files or deleting missing files during commit, they must be noted during commit and written into the working tree at some point
  • refuse "pointless" commits with no file changes - should be easy by just refusing to do the final step of storing a new overall inventory and revision object
  • heuristic detection of renames between add and delete (out of scope for this change)
  • pushing changes to a master branch if any
  • running hooks, pre and post commit
  • prompting for a commit message if necessary, including a list of the changes that have already been observed
  • if there are tree references and recursing into them is enabled, then do so

Commit needs to protect against duplicated file ids

Updates that need to be made in the working tree, either on conclusion of commit or during the scan, include

  • Changes made to the tree shape, including automatic adds, renames or deletes
  • For trees (eg dirstate) that cache parent inventories, the old parent information must be removed and the new one inserted
  • The tree hashcache information should be updated to reflect the stat value at which the file was the same as the committed version, and the content hash it was observed to have. This needs to be done carefully to prevent inconsistencies if the file is modified during or shortly after the commit. Perhaps it would work to read the mtime of the file before we read its text to commit.

3.4.9   Interface stack

The commit api is invoked by the command interface, and copies information from the tree into the branch and its repository, possibly updating the WorkingTree afterwards.

The command interface passes:

  • a commit message (from an option, if any),
  • or an indication that it should be read interactively from the ui object;
  • a list of files to commit
  • an option for a dry-run commit
  • verbose option, or callback to indicate
  • timestamp, timezone, committer, chosen revision id
  • config (for what?)
  • option for local-only commit on a bound branch
  • option for strict commits (fail if there are unknown or missing files)
  • option to allow "pointless" commits (with no tree changes)

(This is rather a lot of options to pass individually and just for code tidyness maybe some of them should be combine into objects.)

>>> Branch.commit(from_tree, message, files_to_commit, ...)

There will be different implementations of this for different Branch classes, whether for foreign branches or Bazaar repositories using different storage methods.

Most of the commit should occur during a single lockstep iteration across the workingtree and parent trees. The WorkingTree interface needs to provide methods that give commit all it needs. Some of these methods (such as answering the file's last change revision) may be deprecated in newer working trees and there we have a choice of either calculating the value from the data that is present, or refusing to support commit to newer repositories.

For a dirstate tree the iteration of changes from the parent can easily be done within its own iter_changes.

Dirstate inventories may be most easily updated in a single operation at the end; however it may be best to accumulate data as we proceed through the tree rather than revisiting it at the end.

Showing a progress bar for commit may not be necessary if we report files as they are committed. Alternatively we could transiently show a progress bar for each directory that's scanned, even if no changes are observed.

This needs to collect a list of added/changed/removed files, each of which must have its text stored (if any) and containing directory updated. This can be done by calling Tree._iter_changes on the source tree, asking for changes

In the 0.17 model the commit operation needs to know the per-file parents and per-file last-changed revision.

(In this and other operations we must avoid having multiple layers walk over the tree separately. For example, it is no good to have the Command layer walk the tree to generate a list of all file ids to commit, because the tree will also be walked later. The layers that do need to operate per-file should probably be bound together in a per-dirblock iterator, rather than each iterating independently.)

3.4.10   Branch->Tree interface

The Branch commit code needs to ask the Tree what should be committed, in terms of changes from the parent revisions. If the Tree holds all the necessary parent tree information itself it can do it single handed; otherwise it may need to ask the Repository for parent information.

This should be a streaming interface, probably like iter_changes returning information per directory block.

The interface should not return a block for directories that are recursively unchanged.

The tree's idea of what is possibly changed may be more conservative than that of the branch. For example the tree may report on merges of files where the text is identical to the parents: this must be recorded for Bazaar branches that record per-file ancestry but is not necessary for all branches. If the tree is responsible for determining when directories have been recursively modified then it will report on all the parents of such files. There are several implementation options:

1. Return all files and directories the branch might want to commit, even if the branch ends up taking no action on them.

2. When starting the iteration, the branch can specify what type of change is considered interesting.

Since these types of changes are probably (??) rare compared to files that are either completely unmodified or substantially modified, the first may be the best and simplest option.

The branch needs to build an inventory to commit, which must include unchanged files within changed directories. This should be returned from the working tree too. Repositories that store per-directory inventories will want to build and store these from the lowest directories up. For 0.17 format repositories with an all-in-one inventory it may be easiest to accumulate inventory entries in arbitrary order into an in-memory Inventory and then serialize it.

It ought to be possible to commit any Tree into a Branch, without requiring a WorkingTree; the commit code should cope if the tree is not interested in updating hashcache information or does not have a last_revision.

3.4.11   Information from the tree to repository

The main things the tree needs to tell the Branch about are:

  • A file is modified from its parent revision (in text, permissions, other), and so its text may need to be stored.

    Files should also be reported if they have more than one unique parent revision, for repositories that store per-file graphs or last-change revisions. Perhaps this behaviour should be optional.

    XXX: are renames/deletions reported here too?

  • The complete contents of a modified directory, so that its inventory text may be stored. This should be done after all the contained files and directories have been reported. If there are unmodified files, or unselected files carried through from

    XXX: Actually perhaps not grouped by directory, but rather grouped appropriately for the shape of inventory storage in the repository.

    In a zoomed-in checkout the workingtree may not have all the shape data for the entire tree.

  • A file is missing -- could cause either automatic removal or an aborted commit.

  • Any unknown files -- can cause automatic addition, abortion of a strict commit, or just reporting.

3.4.12   Information from the repository to the tree

After the commit the tree needs to be updated to the new revision. Some information which was accumulated during the commit must be made available to the workingtree. It's probably reasonable to hold it all in memory and allow the workingtree to get it in whatever order it wants.

  • A list of modified entries, and for each one:

    • The stat values observed when the file was first read.
    • The hash of the committed file text.
    • The file's last-change revision, if appropriate.

    This should include any entries automatically added or removed.

This might be construed as an enhanced version of set_parent_trees. We can avoid a stat on each file by using the value that was observed when it was first read.

3.4.13   Selective commit

For a partial commit the directory contents may need to contain a mix of entries from the working tree and parent trees. This code probably shouldn't live in a specific tree implementation; maybe there should be a general filter that selects paths from one tree into another?

However, the tree walking code does probably need to know about selected paths to avoid examining unselected files or directories.

We never refuse selective file commits (except of merges).

3.4.14   Common commit code

What is common to all commit implementations, regardless of workingtree or repository format?

  • Prompting for a commit message?
  • Strictness/conflict checks?
  • Auto add/remove?

How should this be separated?

3.4.15   Order of traversal

For current and contemplated Bazaar storage formats, we can only finally commit a directory after its contained files and directories have been committed.

The dirstate workingtree format naturally iterates by directory in order by path, yielding directories before their contents. This may also be the most efficient order in which to stat and read the files.

One option would be to construe the interface as a visitor which reports when files are detected to be changed, and also when directories are finished.

3.4.16   Open question: per-file graphs

XXX: If we want to retain explicitly stored per-file graphs, it would seem that we do need to record per-file parents. We have not yet finally settled that we do want to remove them or treat them as a cache. This api stack is still ok whether we do or not, but the internals of it may change.

3.5   diff Performance Analysis

3.5.1   Minimal Work

3.5.1.1   Reuse of historical comparisons

A significant part of the work done by diff is sequence matching. This scales O(n^2) with the number of lines in the file. Therefore, it is worthwile to avoid content comparisons as much as possible.

Our current knit format contains content comparisons, and this data can be converted into lists of matching blocks. Other future formats such as mpdiff may also support such conversion. So it is possible to reuse past comparisons.

It is also possible to combine sequential comparisons. So given a comparison of "foo" to "bar", and "bar" to "baz", it is possible to derive a comparison of "foo" to "baz".

Reuse of historical comparisons will scale with the number of uncommon build-parents between the two historical revisions. This will typically be proportional to the amount of change that the file has undergone. Therefore, in the common case, reuse of historical comparisons will scale with the amount of change.

The downside of such reuse is that it ties the comparison to the historical data. But given the performance improvement, it seems to be worth consideration. Fresh comparisons can be performed if the user requests them.

It may also be possible to accelerate comparisons by including annotation data, thus increasing the number of unique lines.

3.5.1.2   Historical Tree Against Historical Tree

This operation should be strictly proportional to the amount of change, because a comparison has already been done at commit time. Achieving that performance requires the committed data to be properly structured, so that the comparison can be extracted and combined with other comparisons. This comparision extraction should be possible at the inventory and file-content levels.

Minimum work:

  1. Extract and combine inventory comparisons
  2. Extract and combine text comparisions for modified texts

3.5.1.3   Basis Against Historical Tree

This is another case of Historical Tree Against Historical Tree.

3.5.1.4   Basis Against Basis

This is another case of Historical Tree Against Historical Tree.

3.5.1.5   Working Tree Against Basis

This must scale with the number of versioned files, unless the user indicates that only certain files should be compared.

Performance can be further improved by caching comparisons to avoid repeating them. Caching could potentially be performed by diff and perhaps by merge. Merge is aware of the relationship of a text merge's result to the THIS value, and the THIS value is generally the basis value. So the comparison is latent, but present. The only issue is extracting it.

The cache could be indexed by sha1sum pairs. It could also be indexed by file-id, to facilitate removal of stale data.

Minimum work:

  1. Scan working tree for modified files
  2. Retrieve cached comparisons
  3. Perform comparisons on files with no cached comparisons
  4. Cache comparisons for files with no cached comparisons

3.5.1.6   Working Tree Against Historical Tree

This can be structured as a comparison of working tree against basis tree, followed by basis tree against historical tree. Therefore, it combines the performance characteristics of "Working Tree Against Basis" with "Basis Against Historical Tree".

3.5.1.7   Working Tree Against Working Tree

This can be structured as two comparisons against basis, and one comparison of basis against basis. Its performance is therefore similar to Working Tree Against Historical Tree.

3.5.2   API Changes

Desired API:

  • Tree.get_comparision(file_id, tree)

This probably entails:

  • WorkingTree.store_comparison(file_id, revision_id, sha1, comparison)
  • WorkingTree.get_comparison(file_id, revision_id, sha1)
  • Repository.get_comparision(file_id, revision_id, revision_id)
  • merge_comparisions(comparison, comparision)

3.5.3   Storage considerations

It must be cheap (e.g. scale with number of intermediate revisions) to perform comparison of two historical texts. It must be cheap to perform comparison of the inventories of two historical trees.

3.6   Garbage Collection

Garbage collection is used to remove data from a repository that is no longer referenced.

Generally this involves locking the repository and scanning all its branches then generating a new repository with less data.

3.6.1   Least work we can hope to perform

  • Read all branches to get initial references - tips + tags.
  • Read through the revision graph to find unreferenced revisions. A cheap HEADS list might help here by allowing comparison of the initial references to the HEADS - any unreferenced head is garbage.
  • Walk out via inventory deltas to get the full set of texts and signatures to preserve.
  • Copy to a new repository
  • Bait and switch back to the original
  • Remove the old repository.

A possibility to reduce this would be to have a set of grouped 'known garbage free' data - 'ancient history' which can be preserved in total should its HEADS be fully referenced - and where the HEADS list is deliberate cheap (e.g. at the top of some index).

possibly - null data in place without saving size.

3.7   Revert

Change users selected paths to be the same as those in a given revision making backups of any paths that bzr did not set the last contents itself.

3.7.1   Least work we can hope to perform

We should be able to do work proportional to the scope the user is reverting and the amount of changes between the working tree and the revision being reverted to.

This depends on being able to compare unchanged subtrees without recursing so that the mapping of paths to revert to ids to revert can be done efficiently. Specifically we should be able to avoid getting the transitive closure of directory contents when mapping back to paths from ids at the start of revert.

One way this might work is to: for the selected scopes, for each element in the wt:

1. get hash tree data for that scope. 1. get 'new enough' hash data for the siblings of the scope: it can be out of date as long as its not older than the last move or rename out of that siblings scope. 1. Use the hash tree data to tune the work done in finding matching paths/ids which are different in the two trees.

For each thing that needs to change - group by target directory?

1. Extract new content. 1. Backup old content or replace-in-place (except windows where we move and replace).

3.8   The status command

The status command is used to provide a pithy listing of the changes between two trees. Its common case is between the working tree and the basis tree, but it can be used between any two arbitrary trees.

3.8.1   UI Overview

Status shows several things in parallel (for the paths the user supplied mapped across the from and to tree, and any pending merges in the to tree).

  1. Single line summary of all new revisions - the pending merges and their parents recursively.
  2. Changes to the tree shape - adds/deletes/renames.
  3. Changes to versioned content - kind changes and content changes.
  4. Unknown files in the to tree.
  5. Files with conflicts in the to tree.

3.8.2   Ideal work for working tree to historical status

We need to do the following things at a minimum:

  1. Determine new revisions - the pending merges and history.
  1. Retrieve the first line of the commit message for the new revisions.
  1. Determine the tree differences between the two trees using the users paths to limit the scope, and resolving paths in the trees for any pending merges. We arguably don't care about tracking metadata for this - only the value of the tree the user commited.
  1. The entire contents of directories which are versioned when showing unknowns.
  1. Whether a given unversioned path is unknown or ignored.
  1. The list conflicted paths in the tree (which match the users path selection?)

Expanding on the tree difference case we will need to:

  1. Stat every path in working trees which is included by the users path selection to ascertain kind and execute bit.
  1. For paths which have the same kind in both trees and have content, read that content or otherwise determine whether the content has changed. Using our hash cache from the dirstate allows us to avoid reading the file in the common case. There are alternative ways to achieve this - we could record a pointer to a revision which contained this fileid with the current content rather than storing the content's hash; but this seems to be a pointless double-indirection unless we save enough storage in the working tree. A variation of this is to not record an explicit pointer but instead define an implicit pointer as being to the left-hand-parent tree.

3.8.3   Locality of reference

  • We should stat files in the same directory without reading or statting files in other directories. That is we should do all the statting we intend to do within a given directory without doing any other IO, to minimise pressure on the drive heads to seek.
  • We should read files in the same directory without reading or writing files in other directories - and note this is separate to statting (file data is usually physically disjoint to metadata).

3.8.4   Scaling observations

  • The stat operation clearly involves every versioned path in the common case.
  • Expanding out the users path selection in a naive manner involves reading the entire tree shape information for both trees and for all pending-merge trees. (Dirstate makes this tolerably cheap for now, but we're still scaling extra-linearly.)
  • The amount of effort required to generate tree differences between the working tree and the basis tree is interesting: with a tree-like structure and some generatable name for child nodes we use the working tree data to eliminate accessing or considering subtrees regardless of historival age. However, if we have had to access the historical tree shape to perform path selection this rather reduces the win we can obtain here. If we can cause path expansion to not require historical shape access (perhaps by performing the expansion after calculating the tree difference for the top level of the selected path) then we can gain a larger win. This strongly suggests that path expansion and tree difference generation should be linked in terms of API.

3.9   Annotate

Broadly tries to ascribe parts of the tree state to individual commits.

There appear to be three basic ways of generating annotations:

If the annotation works by asking the storage layer for successive full texts then the scaling of this will be proportional to the time to diff throughout the history of thing being annotated.

If the annotation works by asking the storage layer for successive deltas within the history of the thing being annotated we believe we can make it scale broadly proportional to the depth of the tree of revisions of the annotated object.

If the annotation works by combining cached annotations such that creating a full text recreates annotations for it then it will scale with the cost of obtaining that text.

Generally we want our current annotations but it would be nice to be able to do whitespace annotations and potentially other diff based annotations.

Some things to think about:

  • Perhaps multiparent deltas would allow us to not store the cached annotations in each delta without losing performance or accuracy.

3.10   Scaling analysys of Merge

  1. Fetch revisions O(a)
  2. Common Ancestor [O(b)] O(h)
  3. Calculate tree merge O(c) [+ O(b) + O(d)] + O(i)
  • text merge O(e * e * f) + O(b)
  1. Find filesystem conflicts O(c)
  2. Resolve filesystem conflicts O(g)
  3. Apply changes O(c) + O(log(d))
  4. Set pending merges O(1)
  5. Print conflicts O(g)
  6. Print changes O(c)
a:revisions missing from repo:
b:nodes in the revision graph:
c:files that differ between base and other:
d:number of files in the tree
e:number of lines in the text
f:number number of files requiring text merge
g:number of conflicts (g <= c)
h:humber of uncommon ancestors
i:number of revisions between base and other

3.10.1   Needs

  • Access to revision graph proportional to number of revisions read
  • Access to changed file metadata proportional to number of changes and number of intervening revisions.
  • O(1) access to fulltexts

3.10.2   Notes

Multiparent deltas may offer some nice properties for performance of annotation based merging.

3.11   Bundle Creation

  1. Find common ancestor [O(a)] O(b)
  2. Emit bundle [O(a)] O(b) O(h)

Per revision

  1. emit metadata O(1)
  2. emit changes for files
  1. find changed files [O(c)] O(f)
  2. emit file metadata O(d)
  3. emit diff [O(e * e) * O(f) + O(h)] O(i)
  4. base64 encode O(g)
  1. emit overal diff (or maybe do interdiff) O(e * e) * O(f)
a:nodes in revision graph
b:number of descendants of common ancestor
c:number of files in the tree
d:length of metadata
e:number of lines
f:number of modified files
g:length of diff
h:nodes in knit graph of modified files
i:length of stored diff

3.11.1   Needs

  • Improved common ancestor algorithm
  • Access to partial revision graph proportional to relevant revisions
  • Access to changed files proportional to number of change files and intervening revisions
  • Use knit deltas without recomputing
  • Access to knit deltas in O(1) time
  • Access to snapshots in O(1) amortized time
  • All snapshots must have knit deltas

3.12   Uncommit Performance Notes

3.12.1   Specification of uncommit

uncommit removes revisions from the head of a branch. (By default, only the very latest revision is removed, but optionally more can be taken.) Uncommit does not affect the repository (garbage collection is a separate step and not done by default). The working tree is not logically modified (revert is a different operation), except as described below about merges.

Uncommit can be performed on either a branch or a working tree (and implicitly its branch.)

If the uncommitted revisions includes one or more merges, after the uncommit those revisions are in the working tree's list of pending merges, because their tree changes are still present in the tree.

For a bound branch, uncommit fails unless the local branch is up to date.

3.13   Missing

Missing is used to find out the differences between the current branch and another branch.

The performance analysis itself brings no further points than the incremental-push-pull one.

More importantly, the UI have been considered not optimal: missing finds and displays the differences between two branches, presenting the revisions that are not common to both branches as two sets:

  • the revisions that are present only in the current branch,
  • the revisions that are present only in the other branch.

A quick and dirty survey indicates that most of the users are interested in only one set of revisions at a time.

From a performance point of view, it may be more appropriate to calculate only the set the user is asking for.

It has been proposed that the missing command be deprecated in favor of a --dry-run option for the push, pull, merge commands.

In the mean time, the missing command stays interesting as it provides an easy way to test, measure and optimize graph differences processing.

4   Subsystem designs

4.1   Directory fingerprints

4.1.1   Introduction

The basic idea is that for a directory in a tree (committed or otherwise), we will have a single scalar value. If these values are the same, the contents of the subtree under that directory are necessarily the same.

This is intended to help with these use cases, by allowing them to quickly skip over directories with no relevant changes, and to detect when a directory has changed:

  • diff/status (both local trees and historical trees)
  • merge
  • log -v
  • log on a directory
  • commit

4.1.2   Use-case oriented APIs

Most of this will be hidden behind the Tree interface. This should cover log -v, diff, status, merge (and implicit merge during push, pull, update):

tree.iter_changes(other_tree)
tree.get_file_lines(file_id)   # and get_file, get_file_text

4.1.2.1   commit

Commit is similar to iter_changes, but different because it needs to compare to all the trees. Commit currently needs to compare the working tree to all the parent trees, which is needed to update the last_modified field and would be unnecessary if we removed that field (for both files and directories) and did not store per-file graphs. This would potentially speed up commit after merge.

Verbose commit also displays the merged files, which does require looking at all parents of files that aren't identical to the left-hand parent.

4.1.2.2   log

Log is interested in two operations: finding the revisions that touched anything inside a directory, and getting the differences between consecutive revisions (possibly filtered to a directory):

find_touching_revisions(branch, file_id) # should be on Branch?

Log shows the revisions that merged a change. At the moment that is not included in the per-file graph, and it would also not be visible if the directories were hashed.

4.1.3   Open questions

  • Is this a good idea at all?

    If changing a file changes all its parent directories up to the root it will cause more churn on commit. (We currently update the all-in-one inventory, but only have to update one line of it.)

    Every time a child changes, we'll get a new node in the per-directory graph. This is generally useful: it allows bzr log to do the default mode easily, which is to show all changes under that directory. The less common operation, log --no-recursive is still possible by looking only at when the directory itself was renamed, added or removed. (That is what the directory graph describes in bzr 0.18 and it is rarely useful.)

  • Should these be hashes or revision ids or something else?

    Pros of using hashes: hashes are easy to generate by a foreign branch plugin (e.g. bzr-svn). They don't need to get recursive last-changed from the foreign branch, or to walk back through history. They just need the relevant directory state, which any system we support can answer.

    Hashes converge: if you modify and then modify back, you get the same hash. This is a pro because you can detect that there were ultimately no significant changes. And also a con: you cannot use these hashes to form a graph because they get cycles.

  • Are the values unique across the whole tree, or only when comparing different versions of the same object?

    If we use last-changed revisions, then they will be very not unique across the whole tree. To look up the contents, you must pass a composite key like (file_id, last_changed).

    If we use hashes they will be same only when the two contain the same contents. Since we say that file ids must be unique, this means they will match if and only if they are empty. We might relax that in future when we introduce path tokens.

  • Is it reasonable to assume hashes won't collide?

    The odds of SHA-1 hashes colliding "accidentally" are vanishingly small.

    It is possible that a preimage attack against SHA-1 may be discovered in the future. Since we're not proposing in this document to make revision-ids be SHA-1, if SHA-1 was obsoleted then we could rewrite the contents of revisions but would not need to rename revisions. So the impact of such a migration should just be a format upgrade, and a recommendation (but not requirement) to re-sign revisions.

  • If we use hashes, should it be the hash of the representation stored for a directory?

    In other words, should we pun the representation of the directory with the form used for validation.

    If there's some data stored that's not in the hash it's problematic. The hash in no longer (effectively) uniquely identifies the representation.

    It is desirable that we have a hash that covers all data, to guard against bugs, transmission errors, or users trying to hand-hack files. Since we need one hash of everything in the tree, perhaps we should also use it for the fingerprint.

    Testaments explicitly separate the form used for hashing/signing from the form used for storage. This allows us to change the storage form without breaking existing GPG signatures. The downside is that we need to do work O(tree) to make a testament, and this slows down signing, verifying and generating bundles. It also means that there is some stored data which is not protected by the signature: this data is less important, but corruption of it would still cause problems. We have encountered some specific problems with disagreement between inventories as to the last-change of files, which is currently unsigned. These problems can be introduced by ghosts.

    If we hash the representation, there is still a way to support old signatures, assuming that we never discard irreplaceable information. The signature should say what format it applies to (similar to testaments), and we could transform in memory the tree back to that format.

  • Is hashing substantially slower than other possible approaches?

    We already hash all the plain files. Except in unusual cases, the directory metadata will be substantially smaller: perhaps 200:1 as a rule of thumb.

    When building a bzr tree, we spend on the order of 100ms hashing all the source lines to validate them (about 13MB of source).

  • Can you calculate one from a directory in the working tree? Without a basis?

    This seems possible with either hashes or revision ids.

    Using last_changed means that calculating the fingerprint from a working tree necessarily requires reading the inventory for the basis revision, so that we know when unchanged files were last changed. With hashes we could calculate them using the working tree information alone. It's true that we will often then compare that information to the basis tree (e.g. for simple bzr diff), but we may only have to compare at the top level, and sometimes we're comparing to a different tree. This also touches on whether we should store last_modified for files, rather than directories.

    For revision ids we need to assign a value to use for uncommitted changes, but see below about the problems of this.

    In some ways it would be elegant to say (hypothetical):

    wt.get_root().get_last_modified() == branch.get_last_revision()
    

    to know that nothing was changed; but this may not be much better than

    wt.get_root().get_hash() ==
      branch.get_basis().get_root().get_hash()
    
  • Can you use this to compare (directories from) two working trees?

    If you can generate it from a working tree, you should be able to use it to compare them.

    This does rule out for example using last_modified=None or ='current:' to mean "changed in the working tree." Even if this is not supported there seems some risk that we would get the same fingerprint for trees that are actually different.

    We could assign a hypothetical revision id to the tree for uncommitted files. In that case there is some risk that the not-yet-committed id would become visible or committed.

  • Can we use an "approximate basis"?

    When using radix trees, you may need context beyond the specific directory being compared.

  • Can you get the fingerprint of parents directories with only selected file ids taken from the working tree?

    With hashes, we'd want to carry through the unselected files and directories from the values they had in the parent revision.

  • Are unbalanced trees a significant problem? Trees can be unbalanced by having many directories (deep or wide), or many files per directory.

    For small trees like bzr, 744 of 874 are in the bzrlib subtree. In general, larger trees are more balanced, because humans, editors and other tools have trouble managing very unbalanced trees. But there are exceptions: Aaron has one tree with 20,000 generated but versioned entries in one directory.

  • Should we use a radix tree approach where fingerprints are calculated on a synthetic tree that is by definition balanced, even when the actual tree is unbalanced?

  • What are the specific advantages of using recursive-last-modified rather than hashes?

    It may be a smaller step change.

    It's a bidirectional link: given a directory text identifier (file_id, last_changed) you can look up the revision that last changed it.

    From the preceding, even without the per-file graph you can skip through the history of this file: go to the last-changed revision, look at all its parents and repeat.

  • Is it a smaller change to use recursive-last-modified on directories?

    Probably yes:

    1. We can just put it into the current inventory format without changing anything else.

      By contrast to use a hash we'd have to either split up the inventory as stored, or change the sort order for the inventory, or synthesize per-directory inventories in memory for hashing.

      However, xml is somewhat redundant and slow to parse/generate; and reading the whole thing before comparing some sections is only a partial win. It may be a smaller change but we'd be preserving things we want to change.

    1. At present we rarely hash storage representations, only file texts. This is not a large technical change, but it is a conceptual change. This has some consequences for how we can upgrade it in future: all the changed directories need to be rewritten up to the revision level.
    1. If we address directories by hash we need hash-addressed storage.
    1. If we address directories by hash then for consistency we'd probably (not necessarily) want to address file texts by hash.
    1. The per-file graph can't be indexed by hash because they can converge, so we need to either rework or dispose of the per-file graph.
  • Any possibilities for avoiding hashes recurring?

    1. Hash along with an identification of the parents (as in hg). Then you can't convert a tree without all its basis trees, and there is still convergence when the same merge is done by two people, and you can't create it directly from the working tree.
    1. Include last-modified revision id in the hash.
    1. Index by (revision, hash) or vice versa.
    1. Store a per-file graph and allow it to have repeated keys. The graph would tell you about all the parent texts ever seen; you would need to use revision graph information to resolve ambiguities.
  • What are the specific disadvantages of using recursive-last-modified rather than hashes?

    To calculate the last-changed revision, given the last-changed information of the contained files, you need to look at the revision graph. They're not enough because you need to know the relations between the mentioned revisions. In a merge it's possible the correct directory last-modified will not be the same as that of any of the files within it. This can also happen when a file is removed (deleted or renamed) from a directory.

  • Should we split up storage of the inventories?

    This is not quite the same but connected.

  • How does this relate to per-file/per-directory hashes?

    If the version of a file or directory is identified by a hash, we can't use that to point into a per-file graph. We can have a graph indexed by (file_id, hash, revision_id). The last-modified could be stored as part of this graph.

    The graph would no longer be core data; it could be always present but might be rebuilt. Treating it as non-core data may make some changes like shallow branches easier?

  • How do you ask a tree for a given text?

    Right now we say

    revision_tree.get_file_lines(file_id)
    

    so the choice of storage is hidden behind the revision tree: it could be accessed by (file_id, last_changed) or by hash or otherwise.

    At the moment the Repository exports a friend api to RevisionTree, currently usually talking in VersionedFiles.

    We probably wouldn't want Repository to expose a get_text_for_sha1() interface because that would be very difficult to support on old repositories or on foreign branches.