1
Commit Performance Notes
2
========================
9
We want to improve the commit code in two phases.
11
Phase one is to have a better separation from the format-specific logic,
12
the user interface, and the general process of committing.
14
Phase two is to have better interfaces by which a good workingtree format
15
can efficiently pass data to a good storage format. If we get phase one
16
right, it will be relatively easy and non-disruptive to bring this in.
19
Commit: The Minimum Work Required
20
---------------------------------
22
Here is a description of the minimum work that commit must do. We
23
want to make sure that our design doesn't cost too much more than this
24
minimum. I am trying to do this without making too many assumptions
25
about the underlying storage, but am assuming that the ui and basic
26
architecture (wt, branch, repo) stays about the same.
28
The basic purpose of commit is to:
30
1. create and store a new revision based on the contents of the working tree
31
2. make this the new basis revision for the working tree
33
We can do a selected commit of only some files or subtrees.
35
The best performance we could hope for is:
36
- stat each versioned selected working file once
37
- read from the workingtree and write into the repository any new file texts
38
- in general, do work proportional to the size of the shape (eg
39
inventory) of the old and new selected trees, and to the total size of
44
1.0 - Store new file texts: if a versioned file contains a new text
45
there is no avoiding storing it. To determine which ones have changed
46
we must go over the workingtree and at least stat each file. If the
47
file is modified since it was last hashed, it must be read in.
48
Ideally we would read it only once, and either notice that it has not
49
changed, or store it at that point.
51
On the other hand we want new code to be able to handle files that are
52
larger than will fit in memory. We may then need to read each file up
53
to two times: once to determine if there is a new text and calculate
54
its hash, and again to store it.
56
1.1 - Store a tree-shape description (ie inventory or similar.) This
57
describes the non-file objects, and provides a reference from the
58
Revision to the texts within it.
60
1.2 - Generate and store a new revision object.
62
1.3 - Do delta-compression on the stored objects. (git notably does
63
not do this at commit time, deferring this entirely until later.)
64
This requires finding the appropriate basis for each modified file: in
65
the current scheme we get the file id, last-revision from the
66
dirstate, look into the knit for that text, extract that text in
67
total, generate a delta, then store that into the knit. Most delta
68
operations are O(n**2) to O(n**3) in the size of the modified files.
70
1.4 - Cache annotation information for the changes: at the moment this
71
is done as part of the delta storage. There are some flaws in that
72
approach, such as that it is not updated when ghosts are filled, and
73
the annotation can't be re-run with new diff parameters.
75
2.1 - Make the new revision the basis for the tree, and clear the list
76
of parents. Strictly this is all that's logically necessary, unless
77
the working tree format requires more work.
79
The dirstate format does require more work, because it caches the
80
parent tree data for each file within the working tree data. In
81
practice this means that every commit rewrites the entire dirstate
82
file - we could try to avoid rewriting the whole file but this may be
83
difficult because variable-length data (the last-changed revision id)
84
is inserted into many rows.
86
The current dirstate design then seems to mean that any commit of a
87
single file imposes a cost proportional to the size of the current
88
workingtree. Maybe there are other benefits that outweigh this.
89
Alternatively if it was fast enough for operations to always look at
90
the original storage of the parent trees we could do without the
93
2.2 - Record the observed file hashes into the workingtree control
94
files. For the files that we just committed, we have the information
95
to store a valid hash cache entry: we know their stat information and
96
the sha1 of the file contents. This is not strictly necessary to the
97
speed of commit, but it will be useful later in avoiding reading those
98
files, and the only cost of doing it now is writing it out.
100
In fact there are some user interface niceties that complicate this:
102
3 - Before starting the commit proper, we prompt for a commit message
103
and in that commit message editor we show a list of the files that
104
will be committed: basically the output of bzr status. This is
105
basically the same as the list of changes we detect while storing the
106
commit, but because the user will sometimes change the tree after
107
opening the commit editor and expect the final state to be committed I
108
think we do have to look for changes twice. Since it takes the user a
109
while to enter a message this is not a big problem as long as both the
110
status summary and the commit are individually fast.
112
4 - As the commit proceeds (or after?) we show another status-like
113
summary. Just printing the names of modified files as they're stored
114
would be easy. Recording deleted and renamed files or directories is
115
more work: this can only be done by reference to the primary parent
116
tree and requires it be read in. Worse, reporting renames requires
117
searching by id across the entire parent tree. Possibly full
118
reporting should be a default-off verbose option because it does
119
require more work beyond the commit itself.
121
5 - Bazaar currently allows for missing files to be automatically
122
marked as removed at the time of commit. Leaving aside the ui
123
consequences, this means that we have to update the working inventory
124
to mark these files as removed. Since as discussed above we always
125
have to rewrite the dirstate on commit this is not substantial, though
126
we should make sure we do this in one pass, not two. I have
127
previously proposed to make this behaviour a non-default option.
129
We may need to run hooks or generate signatures during commit, but
130
they don't seem to have substantial performance consequences.
132
If one wanted to optimize solely for the speed of commit I think
133
hash-addressed file-per-text storage like in git (or bzr 0.1) is very
134
good. Remarkably, it does not need to read the inventory for the
135
previous revision. For each versioned file, we just need to get its
136
hash, either by reading the file or validating its stat data. If that
137
hash is not already in the repository, the file is just copied in and
138
compressed. As directories are traversed, they're turned into texts
139
and stored as well, and then finally the revision is too. This does
140
depend on later doing some delta compression of these texts.
142
Variations on this are possible. Rather than writing a single file
143
into the repository for each text, we could fold them into a single
144
collation or pack file. That would create a smaller number of files
145
in the repository, but looking up a single text would require looking
146
into their indexes rather than just asking the filesystem.
148
Rather than using hashes we can use file-id/rev-id pairs as at
149
present, which has several consequences pro and con.
155
At first glance, commit simply stores the changes status reports. In fact,
156
this isn't technically correct: commit considers some files modified that
157
status does not. The notes below were put together by John Arbash Meinel
158
and Aaron Bentley in May 2007 to explain the finer details of commit to
159
Ian Clatworthy. They are recorded here as they are likely to be useful to
160
others new to Bazaar ...
162
1) **Unknown files have a different effect.** With --no-strict (the default)
163
they have no effect and can be completely ignored. With --strict they
164
should cause the commit to abort (so you don't forget to add the two new
165
test files that you just created).
167
2) **Multiple parents.** 'status' always compares 2 trees, typically the
168
last-committed tree and the current working tree. 'commit' will compare
169
more trees if there has been a merge.
171
a) The "last modified" property for files.
172
A file may be marked as changed since the last commit, but that
173
change may have come in from the merge, and the change could have
174
happened several commits back. There are several edge cases to be
175
handled here, like if both branches modified the same file, or if
176
just one branch modified it.
178
b) The trickier case is when a file appears unmodified since last
179
commit, but it was modified versus one of the merged branches. I
180
believe there are a few ways this can happen, like if a merged
181
branch changes a file and then reverts it back (you still update
182
the 'last modified' field).
183
In general, if both sides disagree on the 'last-modified' flag,
184
then you need to generate a new entry pointing 'last-modified' at
185
this revision (because you are resolving the differences between
188
3) **Automatic deletion of 'missing' files.** This is a point that we go
189
back and forth on. I think the basic idea is that 'bzr commit' by
190
default should abort if it finds a 'missing' file (in case that file was
191
renamed rather than deleted), but 'bzr commit --auto' can add unknown
192
files and remove missing files automatically.
194
4) **sha1 for newly added files.** status doesn't really need this: it should
195
only care that the file is not present in base, but is present now. In
196
some ways commit doesn't care either, since it needs to read and sha the
199
5) **Nested trees.** status doesn't recurse into nested trees, but commit does.
200
This is just because not all of the nested-trees work has been merged yet.
202
A tree-reference is considered modified if the subtree has been
203
committed since the last containing-tree commit. But commit needs to
204
recurse into every subtree, to ensure that a commit is done if the
205
subtree has changed since its last commit. _iter_changes only reports
206
on tree-references that are modified, so it can't be used for doing
210
Avoiding Work: Smarter Change Detection
211
---------------------------------------
213
Commit currently walks through every file building an inventory. Here is
214
Aaron's brain dump on a better way ...
216
_iter_changes won't tell us about tree references that haven't changed,
217
even if those subtrees have changed. (Unless we ask for unchanged
218
files, which we don't want to do, of course.)
220
There is an iter_references method, but using it looks just as expensive
223
I did some work on updating commit to use iter_changes, but found for
224
multi-parent trees, I had to fall back to the slow inventory comparison
227
Really, I think we need a call akin to iter_changes that handles
228
multiple parents, and knows to emit entries when InventoryEntry.revision
229
is all that's changed.
232
Avoiding Work: Better Layering
233
------------------------------
235
For each file, commit is currently doing more work than it should. Here is
236
John's take on a better way ...
238
Note that "_iter_changes" *does* have to touch every path on disk, but
239
it just can do it in a more efficient manner. (It doesn't have to create
240
an InventoryEntry for all the ones that haven't changed).
242
I agree with Aaron that we need something a little different than
243
_iter_changes. Both because of handling multiple parents, as well as we
244
don't want it to actually read the files if we have a stat-cache miss.
246
Specifically, the commit code *has* to read the files because it is
247
going to add the text to the repository, and we want it to compute the
248
sha1 at *that* time, so we are guaranteed to have the valid sha (rather
249
than just whatever the last cached one was). So we want the code to
250
return 'None' if it doesn't have an up-to-date sha1, rather than reading
251
the file and computing it, just before it returns it to the parent.
253
The commit code (0.16) should really be restructured. It's layering is
256
Specifically, calling "kind()" requires a stat of the file. But we have
257
to do a stat to get the size/whether the record is up-to-date, etc. So
258
we really need to have a "create_an_up_to_date_inventory()" function.
259
But because we are accessing every object on disk, we want to be working
260
in tuples rather than Inventory objects. And because DirState already
261
has the parent records next to the current working inventory, it can do
262
all the work to do really fast comparison and throw-away of unimportant
265
The way I made "bzr status" fast is by moving the 'ignore this record'
266
ability as deep into the stack as I could get. Status has the property
267
that you don't care about most of the records, just like commit. So the
268
sooner you can stop evaluating the 99% that you don't care about, the
272
Avoiding work: avoiding reading parent data
273
-------------------------------------------
275
We would like to avoid the work of reading any data about the parent
276
revisions. We should at least try to avoid reading anything from the
277
repository; we can also consider whether it is possible or useful to hold
278
less parent information in the working tree.
280
When a commit of selected files is requested, the committed snapshot is a
281
composite of some directories from the parent revision and some from the
282
working tree. In this case it is logically necessary to have the parent
283
inventory information.
285
If file last-change information or per-file graph information is stored
286
then it must be available from the parent trees.
288
If the Branch's storage method does delta compression at commit time it
289
may need to retrieve file or inventory texts from the repository.
291
It is desirable to avoid roundtrips to the Repository during commit,
292
particularly because it may be remote. If the WorkingTree can determine
293
by itself that a text was in the parent and therefore should be in the
294
Repository that avoids one roundtrip per file.
296
There is a possibility here that the parent revision is not stored, or not
297
correctly stored, in the repository the tree is being committed into, and
298
so the committed tree would not be reconstructable. We could check that
299
the parent revision is present in the inventory and rely on the invariant
300
that if a revision is present, everything to reconstruct it will be
307
Caller starts a commit
309
>>> Branch.commit(from_tree, options)
311
This creates a CommitBuilder object matched to the Branch, Repository and
312
Tree. It can vary depending on model differences or by knowledge of what
313
is efficient with the Repository and Tree. Model differences might
314
include whether no-text-change merges need to be reported, and whether the
316
The basic CommitBuilder.commit structure can be
318
1. Ask the branch if it is ready to commit (up to date with master if
321
2. Ask the tree if it is ready to commit to the branch (up to date with
322
branch?), no conflicts, etc
324
3. Commit changed files; prototype implementation:
326
a. Ask the working tree for all committable files; for each it should
327
return the per-file parents, stat information, kind, etc.
329
b. Ask the repository to store the new file text; the repository should
330
return the stored sha1 and new revision id.
332
4. Commit changed inventory
334
5. Commit revision object
344
Complications of commit
345
-----------------------
347
Bazaar (as of 0.17) does not support selective-file commit of a merge;
348
this could be done if we decide how it should be recorded - is this to be
349
stored as an overall merge revision; as a preliminary non-merge revisions;
350
or will the per-file graph diverge from the revision graph.
352
There are several checks that may cause the commit to be refused, which
353
may be activated or deactivated by options.
355
* presence of conflicts in the tree
357
* presence of unknown files
359
* the working tree basis is up to date with the branch tip
361
* the local branch is up to date with the master branch, if there
362
is one and --local is not specified
364
* an empty commit message is given,
366
* a hook flags an error
368
* a "pointless" commit, with no inventory changes
370
Most of these require walking the tree and can be easily done while
371
recording the tree shape. This does require that it be possible to abort
372
the commit after the tree changes have been recorded. It could be ok to
373
either leave the unreachable partly-committed records in the repository,
378
* when automatically adding new files or deleting missing files during
379
commit, they must be noted during commit and written into the working
382
* refuse "pointless" commits with no file changes - should be easy by
383
just refusing to do the final step of storing a new overall inventory
386
* heuristic detection of renames between add and delete (out of scope for
389
* pushing changes to a master branch if any
391
* running hooks, pre and post commit
393
* prompting for a commit message if necessary, including a list of the
394
changes that have already been observed
396
* if there are tree references and recursing into them is enabled, then
399
Commit needs to protect against duplicated file ids
402
Updates that need to be made in the working tree, either on conclusion
403
of commit or during the scan, include
405
* Changes made to the tree shape, including automatic adds, renames or
408
* For trees (eg dirstate) that cache parent inventories, the old parent
409
information must be removed and the new one inserted
411
* The tree hashcache information should be updated to reflect the stat
412
value at which the file was the same as the committed version, and the
413
content hash it was observed to have. This needs to be done carefully to
414
prevent inconsistencies if the file is modified during or shortly after
415
the commit. Perhaps it would work to read the mtime of the file before we
416
read its text to commit.
422
The commit api is invoked by the command interface, and copies information
423
from the tree into the branch and its repository, possibly updating the
424
WorkingTree afterwards.
426
The command interface passes:
428
* a commit message (from an option, if any),
429
* or an indication that it should be read interactively from the ui object;
430
* a list of files to commit
431
* an option for a dry-run commit
432
* verbose option, or callback to indicate
433
* timestamp, timezone, committer, chosen revision id
435
* option for local-only commit on a bound branch
436
* option for strict commits (fail if there are unknown or missing files)
437
* option to allow "pointless" commits (with no tree changes)
439
(This is rather a lot of options to pass individually and just for code tidyness maybe some of them should be combine into objects.)
441
>>> Branch.commit(from_tree, message, files_to_commit, ...)
443
There will be different implementations of this for different Branch
444
classes, whether for foreign branches or Bazaar repositories using
445
different storage methods.
447
Most of the commit should occur during a single lockstep iteration across
448
the workingtree and parent trees. The WorkingTree interface needs to
449
provide methods that give commit all it needs. Some of these methods
450
(such as answering the file's last change revision) may be deprecated in
451
newer working trees and there we have a choice of either calculating the
452
value from the data that is present, or refusing to support commit to
455
For a dirstate tree the iteration of changes from the parent can easily be
456
done within its own iter_changes.
458
Dirstate inventories may be most easily updated in a single operation at
459
the end; however it may be best to accumulate data as we proceed through
460
the tree rather than revisiting it at the end.
462
Showing a progress bar for commit may not be necessary if we report files
463
as they are committed. Alternatively we could transiently show a progress
464
bar for each directory that's scanned, even if no changes are observed.
466
This needs to collect a list of added/changed/removed files, each of which
467
must have its text stored (if any) and containing directory updated. This
468
can be done by calling Tree._iter_changes on the source tree, asking for
471
In the 0.17 model the commit operation needs to know the per-file parents
472
and per-file last-changed revision.
474
(In this and other operations we must avoid having multiple layers walk
475
over the tree separately. For example, it is no good to have the Command
476
layer walk the tree to generate a list of all file ids to commit, because
477
the tree will also be walked later. The layers that do need to operate
478
per-file should probably be bound together in a per-dirblock iterator,
479
rather than each iterating independently.)
481
Branch->Tree interface
482
----------------------
484
The Branch commit code needs to ask the Tree what should be committed, in
485
terms of changes from the parent revisions. If the Tree holds all the
486
necessary parent tree information itself it can do it single handed;
487
otherwise it may need to ask the Repository for parent information.
489
This should be a streaming interface, probably like iter_changes returning
490
information per directory block.
492
The interface should not return a block for directories that are
493
recursively unchanged.
495
The tree's idea of what is possibly changed may be more conservative than
496
that of the branch. For example the tree may report on merges of files
497
where the text is identical to the parents: this must be recorded for
498
Bazaar branches that record per-file ancestry but is not necessary for all
499
branches. If the tree is responsible for determining when directories
500
have been recursively modified then it will report on all the parents of
501
such files. There are several implementation options:
503
1. Return all files and directories the branch might want to commit, even
504
if the branch ends up taking no action on them.
506
2. When starting the iteration, the branch can specify what type of change
507
is considered interesting.
509
Since these types of changes are probably (??) rare compared to files that
510
are either completely unmodified or substantially modified, the first may
511
be the best and simplest option.
513
The branch needs to build an inventory to commit, which must include
514
unchanged files within changed directories. This should be returned from
515
the working tree too. Repositories that store per-directory inventories
516
will want to build and store these from the lowest directories up.
517
For 0.17 format repositories with an all-in-one inventory it may be
518
easiest to accumulate inventory entries in arbitrary order into an
519
in-memory Inventory and then serialize it.
521
It ought to be possible to commit any Tree into a Branch, without
522
requiring a WorkingTree; the commit code should cope if the tree is not
523
interested in updating hashcache information or does not have a
527
Information from the tree to repository
528
---------------------------------------
530
The main things the tree needs to tell the Branch about are:
532
* A file is modified from its parent revision (in text, permissions,
533
other), and so its text may need to be stored.
535
Files should also be reported if they have more than one unique parent
536
revision, for repositories that store per-file graphs or last-change
537
revisions. Perhaps this behaviour should be optional.
539
**XXX:** are renames/deletions reported here too?
541
* The complete contents of a modified directory, so that its inventory
542
text may be stored. This should be done after all the contained files
543
and directories have been reported. If there are unmodified files,
544
or unselected files carried through from
546
XXX: Actually perhaps not grouped by directory, but rather grouped
547
appropriately for the shape of inventory storage in the repository.
549
In a zoomed-in checkout the workingtree may not have all the shape data
552
* A file is missing -- could cause either automatic removal or an aborted
555
* Any unknown files -- can cause automatic addition, abortion of a strict
556
commit, or just reporting.
559
Information from the repository to the tree
560
-------------------------------------------
562
After the commit the tree needs to be updated to the new revision. Some
563
information which was accumulated during the commit must be made available
564
to the workingtree. It's probably reasonable to hold it all in memory and
565
allow the workingtree to get it in whatever order it wants.
567
* A list of modified entries, and for each one:
569
* The stat values observed when the file was first read.
571
* The hash of the committed file text.
573
* The file's last-change revision, if appropriate.
575
This should include any entries automatically added or removed.
577
This might be construed as an enhanced version of ``set_parent_trees``.
578
We can avoid a stat on each file by using the value that was observed when
586
For a partial commit the directory contents may need to contain a mix of
587
entries from the working tree and parent trees. This code probably
588
shouldn't live in a specific tree implementation; maybe there should be a
589
general filter that selects paths from one tree into another?
591
However, the tree walking code does probably need to know about selected
592
paths to avoid examining unselected files or directories.
594
We never refuse selective file commits (except of merges).
601
What is common to all commit implementations, regardless of workingtree or
604
* Prompting for a commit message?
605
* Strictness/conflict checks?
608
How should this be separated?
615
For current and contemplated Bazaar storage formats, we can only finally
616
commit a directory after its contained files and directories have been
619
The dirstate workingtree format naturally iterates by directory in order
620
by path, yielding directories before their contents. This may also be the
621
most efficient order in which to stat and read the files.
623
One option would be to construe the interface as a visitor which reports
624
when files are detected to be changed, and also when directories are
628
Open question: per-file graphs
629
------------------------------
631
**XXX:** If we want to retain explicitly stored per-file graphs, it would
632
seem that we do need to record per-file parents. We have not yet finally
633
settled that we do want to remove them or treat them as a cache. This api
634
stack is still ok whether we do or not, but the internals of it may