hgbook
view en/ch04-concepts.tex @ 557:23dc79421e06
Add XSL stylesheets
| author | Bryan O'Sullivan <bos@serpentine.com> |
|---|---|
| date | Mon Feb 09 23:04:31 2009 -0800 (2009-02-09) |
| parents | 5cd47f721686 |
| children |
line source
1 \chapter{Behind the scenes}
2 \label{chap:concepts}
4 Unlike many revision control systems, the concepts upon which
5 Mercurial is built are simple enough that it's easy to understand how
6 the software really works. Knowing this certainly isn't necessary,
7 but I find it useful to have a ``mental model'' of what's going on.
9 This understanding gives me confidence that Mercurial has been
10 carefully designed to be both \emph{safe} and \emph{efficient}. And
11 just as importantly, if it's easy for me to retain a good idea of what
12 the software is doing when I perform a revision control task, I'm less
13 likely to be surprised by its behaviour.
15 In this chapter, we'll initially cover the core concepts behind
16 Mercurial's design, then continue to discuss some of the interesting
17 details of its implementation.
19 \section{Mercurial's historical record}
21 \subsection{Tracking the history of a single file}
23 When Mercurial tracks modifications to a file, it stores the history
24 of that file in a metadata object called a \emph{filelog}. Each entry
25 in the filelog contains enough information to reconstruct one revision
26 of the file that is being tracked. Filelogs are stored as files in
27 the \sdirname{.hg/store/data} directory. A filelog contains two kinds
28 of information: revision data, and an index to help Mercurial to find
29 a revision efficiently.
31 A file that is large, or has a lot of history, has its filelog stored
32 in separate data (``\texttt{.d}'' suffix) and index (``\texttt{.i}''
33 suffix) files. For small files without much history, the revision
34 data and index are combined in a single ``\texttt{.i}'' file. The
35 correspondence between a file in the working directory and the filelog
36 that tracks its history in the repository is illustrated in
37 figure~\ref{fig:concepts:filelog}.
39 \begin{figure}[ht]
40 \centering
41 \grafix{filelog}
42 \caption{Relationships between files in working directory and filelogs in repository}
43 \label{fig:concepts:filelog}
44 \end{figure}
46 \subsection{Managing tracked files}
48 Mercurial uses a structure called a \emph{manifest} to collect
49 together information about the files that it tracks. Each entry in
50 the manifest contains information about the files present in a single
51 changeset. An entry records which files are present in the changeset,
52 the revision of each file, and a few other pieces of file metadata.
54 \subsection{Recording changeset information}
56 The \emph{changelog} contains information about each changeset. Each
57 revision records who committed a change, the changeset comment, other
58 pieces of changeset-related information, and the revision of the
59 manifest to use.
61 \subsection{Relationships between revisions}
63 Within a changelog, a manifest, or a filelog, each revision stores a
64 pointer to its immediate parent (or to its two parents, if it's a
65 merge revision). As I mentioned above, there are also relationships
66 between revisions \emph{across} these structures, and they are
67 hierarchical in nature.
69 For every changeset in a repository, there is exactly one revision
70 stored in the changelog. Each revision of the changelog contains a
71 pointer to a single revision of the manifest. A revision of the
72 manifest stores a pointer to a single revision of each filelog tracked
73 when that changeset was created. These relationships are illustrated
74 in figure~\ref{fig:concepts:metadata}.
76 \begin{figure}[ht]
77 \centering
78 \grafix{metadata}
79 \caption{Metadata relationships}
80 \label{fig:concepts:metadata}
81 \end{figure}
83 As the illustration shows, there is \emph{not} a ``one to one''
84 relationship between revisions in the changelog, manifest, or filelog.
85 If the manifest hasn't changed between two changesets, the changelog
86 entries for those changesets will point to the same revision of the
87 manifest. If a file that Mercurial tracks hasn't changed between two
88 changesets, the entry for that file in the two revisions of the
89 manifest will point to the same revision of its filelog.
91 \section{Safe, efficient storage}
93 The underpinnings of changelogs, manifests, and filelogs are provided
94 by a single structure called the \emph{revlog}.
96 \subsection{Efficient storage}
98 The revlog provides efficient storage of revisions using a
99 \emph{delta} mechanism. Instead of storing a complete copy of a file
100 for each revision, it stores the changes needed to transform an older
101 revision into the new revision. For many kinds of file data, these
102 deltas are typically a fraction of a percent of the size of a full
103 copy of a file.
105 Some obsolete revision control systems can only work with deltas of
106 text files. They must either store binary files as complete snapshots
107 or encoded into a text representation, both of which are wasteful
108 approaches. Mercurial can efficiently handle deltas of files with
109 arbitrary binary contents; it doesn't need to treat text as special.
111 \subsection{Safe operation}
112 \label{sec:concepts:txn}
114 Mercurial only ever \emph{appends} data to the end of a revlog file.
115 It never modifies a section of a file after it has written it. This
116 is both more robust and efficient than schemes that need to modify or
117 rewrite data.
119 In addition, Mercurial treats every write as part of a
120 \emph{transaction} that can span a number of files. A transaction is
121 \emph{atomic}: either the entire transaction succeeds and its effects
122 are all visible to readers in one go, or the whole thing is undone.
123 This guarantee of atomicity means that if you're running two copies of
124 Mercurial, where one is reading data and one is writing it, the reader
125 will never see a partially written result that might confuse it.
127 The fact that Mercurial only appends to files makes it easier to
128 provide this transactional guarantee. The easier it is to do stuff
129 like this, the more confident you should be that it's done correctly.
131 \subsection{Fast retrieval}
133 Mercurial cleverly avoids a pitfall common to all earlier
134 revision control systems: the problem of \emph{inefficient retrieval}.
135 Most revision control systems store the contents of a revision as an
136 incremental series of modifications against a ``snapshot''. To
137 reconstruct a specific revision, you must first read the snapshot, and
138 then every one of the revisions between the snapshot and your target
139 revision. The more history that a file accumulates, the more
140 revisions you must read, hence the longer it takes to reconstruct a
141 particular revision.
143 \begin{figure}[ht]
144 \centering
145 \grafix{snapshot}
146 \caption{Snapshot of a revlog, with incremental deltas}
147 \label{fig:concepts:snapshot}
148 \end{figure}
150 The innovation that Mercurial applies to this problem is simple but
151 effective. Once the cumulative amount of delta information stored
152 since the last snapshot exceeds a fixed threshold, it stores a new
153 snapshot (compressed, of course), instead of another delta. This
154 makes it possible to reconstruct \emph{any} revision of a file
155 quickly. This approach works so well that it has since been copied by
156 several other revision control systems.
158 Figure~\ref{fig:concepts:snapshot} illustrates the idea. In an entry
159 in a revlog's index file, Mercurial stores the range of entries from
160 the data file that it must read to reconstruct a particular revision.
162 \subsubsection{Aside: the influence of video compression}
164 If you're familiar with video compression or have ever watched a TV
165 feed through a digital cable or satellite service, you may know that
166 most video compression schemes store each frame of video as a delta
167 against its predecessor frame. In addition, these schemes use
168 ``lossy'' compression techniques to increase the compression ratio, so
169 visual errors accumulate over the course of a number of inter-frame
170 deltas.
172 Because it's possible for a video stream to ``drop out'' occasionally
173 due to signal glitches, and to limit the accumulation of artefacts
174 introduced by the lossy compression process, video encoders
175 periodically insert a complete frame (called a ``key frame'') into the
176 video stream; the next delta is generated against that frame. This
177 means that if the video signal gets interrupted, it will resume once
178 the next key frame is received. Also, the accumulation of encoding
179 errors restarts anew with each key frame.
181 \subsection{Identification and strong integrity}
183 Along with delta or snapshot information, a revlog entry contains a
184 cryptographic hash of the data that it represents. This makes it
185 difficult to forge the contents of a revision, and easy to detect
186 accidental corruption.
188 Hashes provide more than a mere check against corruption; they are
189 used as the identifiers for revisions. The changeset identification
190 hashes that you see as an end user are from revisions of the
191 changelog. Although filelogs and the manifest also use hashes,
192 Mercurial only uses these behind the scenes.
194 Mercurial verifies that hashes are correct when it retrieves file
195 revisions and when it pulls changes from another repository. If it
196 encounters an integrity problem, it will complain and stop whatever
197 it's doing.
199 In addition to the effect it has on retrieval efficiency, Mercurial's
200 use of periodic snapshots makes it more robust against partial data
201 corruption. If a revlog becomes partly corrupted due to a hardware
202 error or system bug, it's often possible to reconstruct some or most
203 revisions from the uncorrupted sections of the revlog, both before and
204 after the corrupted section. This would not be possible with a
205 delta-only storage model.
207 \section{Revision history, branching, and merging}
209 Every entry in a Mercurial revlog knows the identity of its immediate
210 ancestor revision, usually referred to as its \emph{parent}. In fact,
211 a revision contains room for not one parent, but two. Mercurial uses
212 a special hash, called the ``null ID'', to represent the idea ``there
213 is no parent here''. This hash is simply a string of zeroes.
215 In figure~\ref{fig:concepts:revlog}, you can see an example of the
216 conceptual structure of a revlog. Filelogs, manifests, and changelogs
217 all have this same structure; they differ only in the kind of data
218 stored in each delta or snapshot.
220 The first revision in a revlog (at the bottom of the image) has the
221 null ID in both of its parent slots. For a ``normal'' revision, its
222 first parent slot contains the ID of its parent revision, and its
223 second contains the null ID, indicating that the revision has only one
224 real parent. Any two revisions that have the same parent ID are
225 branches. A revision that represents a merge between branches has two
226 normal revision IDs in its parent slots.
228 \begin{figure}[ht]
229 \centering
230 \grafix{revlog}
231 \caption{}
232 \label{fig:concepts:revlog}
233 \end{figure}
235 \section{The working directory}
237 In the working directory, Mercurial stores a snapshot of the files
238 from the repository as of a particular changeset.
240 The working directory ``knows'' which changeset it contains. When you
241 update the working directory to contain a particular changeset,
242 Mercurial looks up the appropriate revision of the manifest to find
243 out which files it was tracking at the time that changeset was
244 committed, and which revision of each file was then current. It then
245 recreates a copy of each of those files, with the same contents it had
246 when the changeset was committed.
248 The \emph{dirstate} contains Mercurial's knowledge of the working
249 directory. This details which changeset the working directory is
250 updated to, and all of the files that Mercurial is tracking in the
251 working directory.
253 Just as a revision of a revlog has room for two parents, so that it
254 can represent either a normal revision (with one parent) or a merge of
255 two earlier revisions, the dirstate has slots for two parents. When
256 you use the \hgcmd{update} command, the changeset that you update to
257 is stored in the ``first parent'' slot, and the null ID in the second.
258 When you \hgcmd{merge} with another changeset, the first parent
259 remains unchanged, and the second parent is filled in with the
260 changeset you're merging with. The \hgcmd{parents} command tells you
261 what the parents of the dirstate are.
263 \subsection{What happens when you commit}
265 The dirstate stores parent information for more than just book-keeping
266 purposes. Mercurial uses the parents of the dirstate as \emph{the parents of a new changeset} when you perform a commit.
268 \begin{figure}[ht]
269 \centering
270 \grafix{wdir}
271 \caption{The working directory can have two parents}
272 \label{fig:concepts:wdir}
273 \end{figure}
275 Figure~\ref{fig:concepts:wdir} shows the normal state of the working
276 directory, where it has a single changeset as parent. That changeset
277 is the \emph{tip}, the newest changeset in the repository that has no
278 children.
280 \begin{figure}[ht]
281 \centering
282 \grafix{wdir-after-commit}
283 \caption{The working directory gains new parents after a commit}
284 \label{fig:concepts:wdir-after-commit}
285 \end{figure}
287 It's useful to think of the working directory as ``the changeset I'm
288 about to commit''. Any files that you tell Mercurial that you've
289 added, removed, renamed, or copied will be reflected in that
290 changeset, as will modifications to any files that Mercurial is
291 already tracking; the new changeset will have the parents of the
292 working directory as its parents.
294 After a commit, Mercurial will update the parents of the working
295 directory, so that the first parent is the ID of the new changeset,
296 and the second is the null ID. This is shown in
297 figure~\ref{fig:concepts:wdir-after-commit}. Mercurial doesn't touch
298 any of the files in the working directory when you commit; it just
299 modifies the dirstate to note its new parents.
301 \subsection{Creating a new head}
303 It's perfectly normal to update the working directory to a changeset
304 other than the current tip. For example, you might want to know what
305 your project looked like last Tuesday, or you could be looking through
306 changesets to see which one introduced a bug. In cases like this, the
307 natural thing to do is update the working directory to the changeset
308 you're interested in, and then examine the files in the working
309 directory directly to see their contents as they were when you
310 committed that changeset. The effect of this is shown in
311 figure~\ref{fig:concepts:wdir-pre-branch}.
313 \begin{figure}[ht]
314 \centering
315 \grafix{wdir-pre-branch}
316 \caption{The working directory, updated to an older changeset}
317 \label{fig:concepts:wdir-pre-branch}
318 \end{figure}
320 Having updated the working directory to an older changeset, what
321 happens if you make some changes, and then commit? Mercurial behaves
322 in the same way as I outlined above. The parents of the working
323 directory become the parents of the new changeset. This new changeset
324 has no children, so it becomes the new tip. And the repository now
325 contains two changesets that have no children; we call these
326 \emph{heads}. You can see the structure that this creates in
327 figure~\ref{fig:concepts:wdir-branch}.
329 \begin{figure}[ht]
330 \centering
331 \grafix{wdir-branch}
332 \caption{After a commit made while synced to an older changeset}
333 \label{fig:concepts:wdir-branch}
334 \end{figure}
336 \begin{note}
337 If you're new to Mercurial, you should keep in mind a common
338 ``error'', which is to use the \hgcmd{pull} command without any
339 options. By default, the \hgcmd{pull} command \emph{does not}
340 update the working directory, so you'll bring new changesets into
341 your repository, but the working directory will stay synced at the
342 same changeset as before the pull. If you make some changes and
343 commit afterwards, you'll thus create a new head, because your
344 working directory isn't synced to whatever the current tip is.
346 I put the word ``error'' in quotes because all that you need to do
347 to rectify this situation is \hgcmd{merge}, then \hgcmd{commit}. In
348 other words, this almost never has negative consequences; it just
349 surprises people. I'll discuss other ways to avoid this behaviour,
350 and why Mercurial behaves in this initially surprising way, later
351 on.
352 \end{note}
354 \subsection{Merging heads}
356 When you run the \hgcmd{merge} command, Mercurial leaves the first
357 parent of the working directory unchanged, and sets the second parent
358 to the changeset you're merging with, as shown in
359 figure~\ref{fig:concepts:wdir-merge}.
361 \begin{figure}[ht]
362 \centering
363 \grafix{wdir-merge}
364 \caption{Merging two heads}
365 \label{fig:concepts:wdir-merge}
366 \end{figure}
368 Mercurial also has to modify the working directory, to merge the files
369 managed in the two changesets. Simplified a little, the merging
370 process goes like this, for every file in the manifests of both
371 changesets.
372 \begin{itemize}
373 \item If neither changeset has modified a file, do nothing with that
374 file.
375 \item If one changeset has modified a file, and the other hasn't,
376 create the modified copy of the file in the working directory.
377 \item If one changeset has removed a file, and the other hasn't (or
378 has also deleted it), delete the file from the working directory.
379 \item If one changeset has removed a file, but the other has modified
380 the file, ask the user what to do: keep the modified file, or remove
381 it?
382 \item If both changesets have modified a file, invoke an external
383 merge program to choose the new contents for the merged file. This
384 may require input from the user.
385 \item If one changeset has modified a file, and the other has renamed
386 or copied the file, make sure that the changes follow the new name
387 of the file.
388 \end{itemize}
389 There are more details---merging has plenty of corner cases---but
390 these are the most common choices that are involved in a merge. As
391 you can see, most cases are completely automatic, and indeed most
392 merges finish automatically, without requiring your input to resolve
393 any conflicts.
395 When you're thinking about what happens when you commit after a merge,
396 once again the working directory is ``the changeset I'm about to
397 commit''. After the \hgcmd{merge} command completes, the working
398 directory has two parents; these will become the parents of the new
399 changeset.
401 Mercurial lets you perform multiple merges, but you must commit the
402 results of each individual merge as you go. This is necessary because
403 Mercurial only tracks two parents for both revisions and the working
404 directory. While it would be technically possible to merge multiple
405 changesets at once, the prospect of user confusion and making a
406 terrible mess of a merge immediately becomes overwhelming.
408 \section{Other interesting design features}
410 In the sections above, I've tried to highlight some of the most
411 important aspects of Mercurial's design, to illustrate that it pays
412 careful attention to reliability and performance. However, the
413 attention to detail doesn't stop there. There are a number of other
414 aspects of Mercurial's construction that I personally find
415 interesting. I'll detail a few of them here, separate from the ``big
416 ticket'' items above, so that if you're interested, you can gain a
417 better idea of the amount of thinking that goes into a well-designed
418 system.
420 \subsection{Clever compression}
422 When appropriate, Mercurial will store both snapshots and deltas in
423 compressed form. It does this by always \emph{trying to} compress a
424 snapshot or delta, but only storing the compressed version if it's
425 smaller than the uncompressed version.
427 This means that Mercurial does ``the right thing'' when storing a file
428 whose native form is compressed, such as a \texttt{zip} archive or a
429 JPEG image. When these types of files are compressed a second time,
430 the resulting file is usually bigger than the once-compressed form,
431 and so Mercurial will store the plain \texttt{zip} or JPEG.
433 Deltas between revisions of a compressed file are usually larger than
434 snapshots of the file, and Mercurial again does ``the right thing'' in
435 these cases. It finds that such a delta exceeds the threshold at
436 which it should store a complete snapshot of the file, so it stores
437 the snapshot, again saving space compared to a naive delta-only
438 approach.
440 \subsubsection{Network recompression}
442 When storing revisions on disk, Mercurial uses the ``deflate''
443 compression algorithm (the same one used by the popular \texttt{zip}
444 archive format), which balances good speed with a respectable
445 compression ratio. However, when transmitting revision data over a
446 network connection, Mercurial uncompresses the compressed revision
447 data.
449 If the connection is over HTTP, Mercurial recompresses the entire
450 stream of data using a compression algorithm that gives a better
451 compression ratio (the Burrows-Wheeler algorithm from the widely used
452 \texttt{bzip2} compression package). This combination of algorithm
453 and compression of the entire stream (instead of a revision at a time)
454 substantially reduces the number of bytes to be transferred, yielding
455 better network performance over almost all kinds of network.
457 (If the connection is over \command{ssh}, Mercurial \emph{doesn't}
458 recompress the stream, because \command{ssh} can already do this
459 itself.)
461 \subsection{Read/write ordering and atomicity}
463 Appending to files isn't the whole story when it comes to guaranteeing
464 that a reader won't see a partial write. If you recall
465 figure~\ref{fig:concepts:metadata}, revisions in the changelog point to
466 revisions in the manifest, and revisions in the manifest point to
467 revisions in filelogs. This hierarchy is deliberate.
469 A writer starts a transaction by writing filelog and manifest data,
470 and doesn't write any changelog data until those are finished. A
471 reader starts by reading changelog data, then manifest data, followed
472 by filelog data.
474 Since the writer has always finished writing filelog and manifest data
475 before it writes to the changelog, a reader will never read a pointer
476 to a partially written manifest revision from the changelog, and it will
477 never read a pointer to a partially written filelog revision from the
478 manifest.
480 \subsection{Concurrent access}
482 The read/write ordering and atomicity guarantees mean that Mercurial
483 never needs to \emph{lock} a repository when it's reading data, even
484 if the repository is being written to while the read is occurring.
485 This has a big effect on scalability; you can have an arbitrary number
486 of Mercurial processes safely reading data from a repository safely
487 all at once, no matter whether it's being written to or not.
489 The lockless nature of reading means that if you're sharing a
490 repository on a multi-user system, you don't need to grant other local
491 users permission to \emph{write} to your repository in order for them
492 to be able to clone it or pull changes from it; they only need
493 \emph{read} permission. (This is \emph{not} a common feature among
494 revision control systems, so don't take it for granted! Most require
495 readers to be able to lock a repository to access it safely, and this
496 requires write permission on at least one directory, which of course
497 makes for all kinds of nasty and annoying security and administrative
498 problems.)
500 Mercurial uses locks to ensure that only one process can write to a
501 repository at a time (the locking mechanism is safe even over
502 filesystems that are notoriously hostile to locking, such as NFS). If
503 a repository is locked, a writer will wait for a while to retry if the
504 repository becomes unlocked, but if the repository remains locked for
505 too long, the process attempting to write will time out after a while.
506 This means that your daily automated scripts won't get stuck forever
507 and pile up if a system crashes unnoticed, for example. (Yes, the
508 timeout is configurable, from zero to infinity.)
510 \subsubsection{Safe dirstate access}
512 As with revision data, Mercurial doesn't take a lock to read the
513 dirstate file; it does acquire a lock to write it. To avoid the
514 possibility of reading a partially written copy of the dirstate file,
515 Mercurial writes to a file with a unique name in the same directory as
516 the dirstate file, then renames the temporary file atomically to
517 \filename{dirstate}. The file named \filename{dirstate} is thus
518 guaranteed to be complete, not partially written.
520 \subsection{Avoiding seeks}
522 Critical to Mercurial's performance is the avoidance of seeks of the
523 disk head, since any seek is far more expensive than even a
524 comparatively large read operation.
526 This is why, for example, the dirstate is stored in a single file. If
527 there were a dirstate file per directory that Mercurial tracked, the
528 disk would seek once per directory. Instead, Mercurial reads the
529 entire single dirstate file in one step.
531 Mercurial also uses a ``copy on write'' scheme when cloning a
532 repository on local storage. Instead of copying every revlog file
533 from the old repository into the new repository, it makes a ``hard
534 link'', which is a shorthand way to say ``these two names point to the
535 same file''. When Mercurial is about to write to one of a revlog's
536 files, it checks to see if the number of names pointing at the file is
537 greater than one. If it is, more than one repository is using the
538 file, so Mercurial makes a new copy of the file that is private to
539 this repository.
541 A few revision control developers have pointed out that this idea of
542 making a complete private copy of a file is not very efficient in its
543 use of storage. While this is true, storage is cheap, and this method
544 gives the highest performance while deferring most book-keeping to the
545 operating system. An alternative scheme would most likely reduce
546 performance and increase the complexity of the software, each of which
547 is much more important to the ``feel'' of day-to-day use.
549 \subsection{Other contents of the dirstate}
551 Because Mercurial doesn't force you to tell it when you're modifying a
552 file, it uses the dirstate to store some extra information so it can
553 determine efficiently whether you have modified a file. For each file
554 in the working directory, it stores the time that it last modified the
555 file itself, and the size of the file at that time.
557 When you explicitly \hgcmd{add}, \hgcmd{remove}, \hgcmd{rename} or
558 \hgcmd{copy} files, Mercurial updates the dirstate so that it knows
559 what to do with those files when you commit.
561 When Mercurial is checking the states of files in the working
562 directory, it first checks a file's modification time. If that has
563 not changed, the file must not have been modified. If the file's size
564 has changed, the file must have been modified. If the modification
565 time has changed, but the size has not, only then does Mercurial need
566 to read the actual contents of the file to see if they've changed.
567 Storing these few extra pieces of information dramatically reduces the
568 amount of data that Mercurial needs to read, which yields large
569 performance improvements compared to other revision control systems.
571 %%% Local Variables:
572 %%% mode: latex
573 %%% TeX-master: "00book"
574 %%% End: