Computing last_modified values


Bazaar (through at least 0.19) computes a last_modified attribute for all inventory entries and stores it at commit time. This is the revision_id that last changed or merged the file. It is used in knit and weave repositories to look up the file text, and to index into the file graph. It’s also used to determine which revisions of the file text to pull during fetch.

This data is not natively stored by most other systems so we need to synthesize it during conversion.

This is a case of non-core data that we might wish to treat as cached, rather than always stored.


Take the set of all “heads”: all the versions of these files in parent trees.

Reduce the heads by eliminating any whose last_modified is an ancestor of the last_modified of any other head.

If there is still more than one head, a new last_modified is assigned. This points to the merge point in the file graph.

If the file text and properties are the same as the sole remaining head, its last_modified is inherited. Property changes include executable bit, filename, and containing directory.

Otherwise, a new last_modified is used.

(This is meant to be the simplest statement, but it may not be the most efficient algorithm; anything that gives equivalent results can be used.)

Generation in commit

Commit and converters both need this when writing into Bazaar native formats.

This is an O(tree) operation because it needs to check for files with multiple heads. It could be reduced to O(changed_or_merged_files) if that was faster to determine. So it needs to be fast.

For the single-parent commit case, we just need to determine which files have changed compared to the parent. If the file was changed, it gets the revision id of the new revision; otherwise it inherits the value from the parent tree.

In the multi-parent commit case (commit of a merge), it can take the value from any of the parent trees, or of the new revision.

Commit in a dirstate tree should be able to do this more easily by looking at a row of the dirstate to get the per-file parents. It still needs to look at the revision or file graph information to work out whether heads can be eliminated as previously merged. At the moment find_previous_heads works on inventories, so needs to spend considerable effort building whole inventories, including files that are not modified or merged. (Called from record_entry_contents.) It might be better to have the commit builder pass in the per-entry parents so that dirstate can generate just those that are necessary. (See also the spec for iter_changes_multiple_parents.)

If merge used a per-file graph then it would know when one version fully supersedes another, and it could emit only a single parent. Merge could in fact do this even when not using per-file graphs. In the current dirstate format we need to store the full data for all trees because they can be extracted from the dirstate, but it could mark some parents as already merged.

Alternatively, we could change the dirstate to include only the base and current trees, and cache the merged-in parents elsewhere.

(Offtopic other dirstate changes: we could also omit the working-copy hash, and just have a stat-fingerprint of when it was last known equal to the basis revision. That reduces the amount of data stored and possibly makes it simpler to update, and shouldn’t penalize common cases.)

Generation during conversion

Accessing a foreign branch requires synthesizing this information. If last_modified is removed from a future bzr version, we will also need to synthesize it to pull back to earlier formats.

Because last_modified is not natively stored in the foreign branches, we want to take advantage of any conversion we’ve already done, so that we don’t need to recursively generate them on every access. We’d prefer to find a revision that’s already converted to a Bazaar inventory within another related repository, such as the target of a conversion.

Avoiding last_modified

last_modified is potentially expensive to determine and we may not want to store it in inventories in future. Therefore we should use it only when necessary:

  • When writing out an inventory format that includes it.
  • In Bazaar formats that use it as a key for the file text or file ancestry. This should be hidden behind the Repository/RevisionTree interface.
  • When a user operation specifically requires the last_modified (e.g. hypothetical annotate directory).

We already do this in most cases.

Compared to annotate

Use cases

Cases to test

  1. Single parent, unmodified file
  2. Single parent, modified file
  3. Two parents, one descended from the other, modified in one parent only
  4. Two parents, one descended from the other, modified in one parent only, but also modified locally.
  5. Two parents, not descended from each other, modified in one parent only.
  6. Two parents, not descended from each other, modified in one parent only, but also modified locally.
  7. Two parents, modified in both to different values.
  8. Two parents, modified in both to the same value.
  9. Two parents, modified in both, and reverted in both back to the original text.
  10. Three parents, modified in only one
  11. Three parents, modified in only one, also modified locally.
  12. Three parents, modified in 2
  13. Three parents, modified in 2, and locally.
  14. Three parents, modified in 2, but one is a descendant of the other.

Performance considerations

Often we’ll want the last_modified information for multiple files, perhaps everything in a directory or in a whole tree. It may be more efficient for the api to accommodate this. Often the last_modified will be similar for multiple files, and if we process them all at once we can avoid some repeated work in calculating their heads.

Open questions

  • How does caching find_heads interact with cherry-picks?

Possible structure

For a single file, if I am different from all parents, ‘new’. (Do not need to evaluate last modified).