The "TODO" file! -*-Indented-Text-*- 22. Catch signals for cleanup when "add"ing files. 24. Insist on a log message. (If done, this should be configurable via commitinfo or some new config file -kingdon, Jun 1995). 30. Add "rdiff" program option to the modules database (and "diff" too?). (perhaps should think a little harder about what this is trying to accomplish and what the best way is -kingdon, Jul 1997). 31. Think hard about ^C recovery. 38. Think hard about using RCS state information to allow one to checkin a new vendor release without having it be accessed until it has been integrated into the local changes. 39. Think about a version of "cvs update -j" which remembers what from that other branch is already merged. This has pitfalls--it could easily lead to invisible state which could confuse users very rapidly--but having to create a tag or some such mechanism to keep track of what has been merged is a pain. Take a look at PRCS 1.2. PRCS 1.0 was particularly bad the way it handled the "invisible state", but 1.2 is significantly better. 45. Consider enhancing the "rdiff" and "tag" (rtag??) command support in the module database -- they seem hard to use since these commands deal directly with the RCS ,v files. 49. cvs xxx commands should be able to deal with files in other directories. I want to do a cvs add foo/bar.c. [[ most commands now use the generic recursion processor, but not all; this note is left here to remind me to fix the others ]] 52. SCCS has a feature that I would *love* to see in CVS, as it is very useful. One may make a private copy of SCCS suid to a particular user, so other users in the authentication list may check files in and out of a project directory without mucking about with groups. Is there any plan to provide a similar functionality to CVS? Our site (and, I'd imagine, many other sites with large user bases) has decided against having the user-groups feature of unix available to the users, due to perceived administrative, technical and performance headaches. A tool such as CVS with features that provide group-like functionality would be a huge help. 62. Consider using revision controlled files and directories to handle the new module format -- consider a cvs command front-end to add/delete/modify module contents, maybe. 63. The "import" and vendor support commands (co -j) need to be documented better. 66. Length of the CVS temporary files must be limited to 14 characters for System-V stupid support. As well as the length on the CVS.adm files. 72. Consider re-design of the module -o, -i, -t options to use the file system more intuitively. 73. Consider an option (in .cvsrc?) to automatically add files that are new and specified to commit. 79. Might be nice to have some sort of interface to Sun's Translucent (?) File System and tagged revisions. 82. Maybe the import stuff should allow an arbitrary revision to be specified. 84. Improve the documentation about administration of the repository and how to add/remove files and the use of symbolic links. 85. Make symbolic links a valid thing to put under version control. Perhaps use one of the tag fields in the RCS file? Note that we can only support symlinks that are relative and within the scope of the sources being controlled. 92. Look into this: After a bit of soul searching via dbx, I realized my sin was that I'd specified "echo" as the program to call from loginfo. The commit procedure worked fine till it hit my echo, then silently aborted leaving the lockfiles intact. Since I needn't use the loginfo facility, I simply removed those commands and it all works. 93. Need to think hard about release and development environments. Think about execsets as well. 98. If diff3 bombs out (too many differences) cvs then thinks that the file has been updated and is OK to be commited even though the file has not yet been merged. 100. Checked out files should have revision control support. Maybe. 102. Perhaps directory modes should be propagated on all import check-ins. Not necessarily uid/gid changes. 103. setuid/setgid on files is suspect. 104. cvs should recover nicely on unreadable files/directories. 105. cvs should have administrative tools to allow for changing permissions and modes and what not. In particular, this would make cvs a more attractive alternative to rdist. 107. It should be possible to specify a list of symbolic revisions to checkout such that the list is processed in reverse order looking for matches within the RCS file for the symbolic revision. If there is not a match, the next symbolic rev on the list is checked, and so on, until all symbolic revs are exhausted. This would allow one to, say, checkout "4.0" + "4.0.3" + "4.0.3Patch1" + "4.0.3Patch2" to get the most recent 4.x stuff. This is usually handled by just specifying the right release_tag, but most people forget to do this. 108. If someone creates a whole new directory (i.e. adds it to the cvs repository) and you happen to have a directory in your source farm by the same name, when you do your cvs update -d it SILENTLY does *nothing* to that directory. At least, I think it was silent; certainly, it did *not* abort my cvs update, as it would have if the same thing had happened with a file instead of a directory. 109. I had gotten pieces of the sys directory in the past but not a complete tree. I just did something like: cvs get * Where sys was in * and got the message cvs get: Executing 'sys/tools/make_links sys' sh: sys/tools/make_links: not found I suspect this is because I didn't have the file in question, but I do not understand how I could fool it into getting an error. I think a later cvs get sys seemed to work so perhaps something is amiss in handling multiple arguments to cvs get? 113. The "cvs update" command should tee its output to a log file in ".". (why? What is wrong with piping stdout to "tee"? -kingdon, Jun 1995) 119. When importing a directory tree that is under SCCS/RCS control, consider an option to have import checkout the SCCS/RCS files if necessary. (This is if someone wants to import something which is in RCS or SCCS without preserving the history, but makes sure they do get the latest versions. It isn't clear to me how useful that is -kingdon, June 1996). 122. If Name_Repository fails, it currently causes CVS to die completely. It should instead return NULL and have the caller do something reasonable (??? -what is reasonable? I'm not sure there is a real problem here. -kingdon, June 1996). 123. Add a flag to import to not build vendor branches for local code. (See `importb' tests in src/sanity.sh for more details). 124. Anyway, I thought you might want to add something like the following to the cvs man pages: BUGS The sum of the sizes of a module key and its contents are limited. See ndbm(3). 126. Do an analysis to see if CVS is forgetting to close file descriptors. Especially when committing many files (more than the open file limit for the particular UNIX). 127. Look at *info files; they should all be quiet if the files are not there. Should be able to point at a RCS directory and go. 130. cvs diff with no -r arguments does not need to look up the current RCS version number since it only cares about what's in the Entries file. This should make it much faster. It should ParseEntries itself and access the entries list much like Version_TS does (sticky tags and sticky options may need to be supported here as well). Then it should only diff the things that have the wrong time stamp (the ones that look modified). 134. Make a statement about using hard NFS mounts to your source repository. Look into checking NULL fgets() returns with ferror() to see if an error had occurred. (we should be checking for errors, quite aside from NFS issues -kingdon, June 1996). 137. Some sites might want CVS to fsync() the RCS ,v file to protect against nasty hardware errors. There is a slight performance hit with doing so, though, so it should be configurable in the .cvsrc file. Also, along with this, we should look at the places where CVS itself could be a little more synchronous so as not to lose data. [[ I've done some of this, but it could use much more ]] 138. Some people have suggested that CVS use a VPATH-like environment variable to limit the amount of sources that need to be duplicated for sites with giant source trees and no disk space. 141. Import should accept modules as its directory argument. If we're going to implement this, we should think hard about how modules might be expanded and how to handle those cases. 143. Update the documentation to show that the source repository is something far away from the files that you work on. (People who come from an RCS background are used to their `repository' being _very_ close to their working directory.) 144. Have cvs checkout look for the environment variable CVSPREFIX (or CVSMODPREFIX or some such). If it's set, then when looking up an alias in the modules database, first look it up with the value of CVSPREFIX attached, and then look for the alias itself. This would be useful when you have several projects in a single repository. You could have aliases abc_src and xyz_src and tell people working on project abc to put "setenv CVSPREFIX abc_" in their .cshrc file (or equivalent for other shells). Then they could do "cvs co src" to get a copy of their src directory, not xyz's. (This should create a directory called src, not abc_src.) 145. After you create revision 1.1.1.1 in the previous scenario, if you do "cvs update -r1 filename" you get revision 1.1, not 1.1.1.1. It would be nice to get the later revision. Again, this restriction comes from RCS and is probably hard to change in CVS. Sigh. |"cvs update -r1 filename" does not tell RCS to follow any branches. CVS |tries to be consistent with RCS in this fashion, so I would not change |this. Within CVS we do have the flexibility of extending things, like |making a revision of the form "-r1HEAD" find the most recent revision |(branch or not) with a "1." prefix in the RCS file. This would get what |you want maybe. This would be very useful. Though I would prefer an option such as "-v1" rather than "-r1HEAD". This option might be used quite often. 146. The merging of files should be controlled via a hook so that programs other than "rcsmerge" can be used, like Sun's filemerge or emacs's emerge.el. (but be careful in making this work client/server--it means doing the interactive merging at the end after the server is done). (probably best is to have CVS do the non-interactive part and tell the user about where the files are (.#foo.c.working and .#foo.c.1.5 or whatever), so they can do the interactive part at that point -kingdon, June 1996). 149. Maybe there should be an option to cvs admin that allows a user to change the Repository/Root file with some degree of error checking? Something like "cvs admin reposmv /old/path /new/pretty/path". Before it does the replace it check to see that the files /new/pretty/path// exist. The obvious cases are where one moves the repository to another machine or directory. But there are other cases, like where the user might want to change from :pserver: to :ext:, use a different server (if there are two server machines which share the repository using a networked file system), etc. The status quo is a bit of a mess (as of, say, CVS 1.9). It is that the -d global option has two moderately different uses. One is to use a totally different repository (in which case we'd probably want to give an error if it disagreed with CVS/Root, as CVS 1.8 and earlier did). The other is the "reposmv" functionality above (in which the two repositories really are the same, and we want to update the CVS/Root files). A related issue is that the fact that CVS only sets and looks at the CVS/Root file in the directory where CVS is run; it doesn't do anything about CVS/Root files in subdirectories. Note also RELATIVE_REPOS in options.h; it needs to be set for changing CVS/Root (not CVS/Repository) to be sufficient in the case where the directory has changed. This whole area is a rather bad pile of individual decisions which accumulated over time, some of them probably bad decisions with hindsight. But we didn't get into this mess overnight, and we're not going to get out of it overnight (that is, we need to come up with a replacement behavior, document what parts of the status quo are deprecated, probably circulate some unofficial patches, &c). (this item originally added 2 Feb 1992 but revised since). 150. I have a customer request for a way to specify log message per file, non-interactively before the commit, such that a single, fully recursive commit prompts for one commit message, and concatenates the per file messages for each file. In short, one commit, one editor session, log messages allowed to vary across files within the commit. Also, the per file messages should be allowed to be written when the files are changed, which may predate the commit considerably. A new command seems appropriate for this. The state can be saved in the CVS directory. I.e., % cvs message foo.c Enter log message for foo.c >> fixed an uninitialized variable >> ^D The text is saved as CVS/foo.c,m (or some such name) and commit is modified to append (prepend?) the text (if found) to the log message specified at commit time. Easy enough. (having cvs commit be non-interactive takes care of various issues like whether to connect to the server before or after prompting for a message (see comment in commit.c at call to start_server). Also would clean up the kludge for what to do with the message from do_editor if the up-to-date check fails (see commit.c client code). I'm not sure about the part above about having commit prompt for an overall message--part of the point is having commit non-interactive and somehow combining messages seems like (excess?) hair. Would be nice to do this so it allows users more flexibility in specifying messages per-directory ("cvs message -l") or per-tree ("cvs message") or per-file ("cvs message foo.c"), and fixes the incompatibility between client/server (per-tree) and non-client/server (per-directory). A few interesting issues with this: (1) if you do a cvs update or some other operation which changes the working directory, do you need to run "cvs message" again (it would, of course, bring up the old message which you could accept)? Probably yes, after all merging in some conflicts might change the situation. (2) How do you change the stored messages if you change your mind before the commit (probably run "cvs message" again, as hinted in (1))? 151. Also, is there a flag I am missing that allows replacing Ulrtx_Build by Ultrix_build? I.E. I would like a tag replacement to be a one step operation rather than a two step "cvs rtag -r Ulrtx_Build Ultrix_Build" followed by "cvs rtag -d Ulrtx_Build" 152. The "cvs -n" option does not work as one would expect for all the commands. In particular, for "commit" and "import", where one would also like to see what it would do, without actually doing anything. 153. There should be some command (maybe I just haven't figured out which one...) to import a source directory which is already RCS-administered without losing all prior RCS gathered data. Thus, it would have to examine the RCS files and choose a starting version and branch higher than previous ones used. (Check out rcs-to-cvs and see if it addresses this issue.) 154. When committing the modules file, a pre-commit check should be done to verify the validity of the new modules file before allowing it to be committed. 155. The options for "cvs history" are mutually exclusive, even though useful queries can be done if they are not, as in specifying both a module and a tag. A workaround is to specify the module, then run the output through grep to only display lines that begin with T, which are tag lines. (Better perhaps if we redesign the whole "history" business -- check out doc/cvs.texinfo for the entire rant.) 156. Also, how hard would it be to allow continuation lines in the {commit,rcs,log}info files? It would probably be useful with all of the various flags that are now available, or if somebody has a lot of files to put into a module. 158. If I do a recursive commit and find that the same RCS file is checked out (and modified!) in two different places within my checked-out files (but within the realm of a single "commit"), CVS will commit the first change, then overwrite that change with the second change. We should catch this (typically unusual) case and issue an appropriate diagnostic and die. 160. The checks that the commit command does should be extended to make sure that the revision that we will lock is not already locked by someone else. Maybe it should also lock the new revision if the old revision was already locked by the user as well, thus moving the lock forward after the commit. 163. The rtag/tag commands should have an option that removes the specified tag from any file that is in the attic. This allows one to re-use a tag (like "Mon", "Tue", ...) all the time and still have it tag the real main-line code. 165. The "import" command will create RCS files automatically, but will screw-up when trying to create long file names on short file name file systems. Perhaps import should be a bit more cautious. 166. There really needs to be a "Getting Started" document which describes some of the new CVS philosophies. Folks coming straight from SCCS or RCS might be confused by "cvs import". Also need to explain: - How one might setup their $CVSROOT - What all the tags mean in an "import" command - Tags are important; revision numbers are not 170. Is there an "info" file that can be invoked when a file is checked out, or updated ? What I want to do is to advise users, particularly novices, of the state of their working source whenever they check something out, as a sanity check. For example, I've written a perl script which tells you what branch you're on, if any. Hopefully this will help guard against mistaken checkins to the trunk, or to the wrong branch. I suppose I can do this in "commitinfo", but it'd be nice to advise people before they edit their files. It would also be nice if there was some sort of "verboseness" switch to the checkout and update commands that could turn this invocation of the script off, for mature users. 173. We have a tagged branch in CVS. How do we get the version of that branch (for an entire directory) that corresponds to the files on that branch on a certain day? I'd like to specify BOTH -r and -D to 'cvs checkout', but I can't. It looks like I can only specify the date for the main line (as opposed to any branches). True? Any workarounds to get what I need? 174. I would like to see "cvs release" modified so that it only removes files which are known to CVS - all the files in the repository, plus those which are listed in .cvsignore. This way, if you do leave something valuable in a source tree you can "cvs release -d" the tree and your non-CVS goodies are still there. If a user is going to leave non-CVS files in their source trees, they really should have to clean them up by hand. 175. And, in the feature request department, I'd dearly love a command-line interface to adding a new module to the CVSROOT/modules file. 176. If you use the -i flag in the modules file, you can control access to source code; this is a Good Thing under certain circumstances. I just had a nasty thought, and on experiment discovered that the filter specified by -i is _not_ run before a cvs admin command; as this allows a user to go behind cvs's back and delete information (cvs admin -o1.4 file) this seems like a serious problem. 177. We've got some external vendor source that sits under a source code hierarchy, and when we do a cvs update, it gets wiped out because its tag is different from the "main" distribution. I've tried to use "-I" to ignore the directory, as well as .cvsignore, but this doesn't work. 179. "cvs admin" does not log its actions with loginfo, nor does it check whether the action is allowed with commitinfo. It should. 180. "cvs edit" should show you who is already editing the files, probably (that is, do "cvs editors" before executing, or some similar result). (But watch out for what happens if the network is down!). 182. There should be a way to show log entries corresponding to changes from tag "foo" to tag "bar". "cvs log -rfoo:bar" doesn't cut it, because it erroneously shows the changes associated with the change from the revision before foo to foo. I'm not sure that is ever a useful or logical behavior ("cvs diff -r foo -r bar" gets this right), but is compatibility an issue? See http://www.cyclic.com/cvs/unoff-log.txt for an unofficial patch. 183. "cvs status" should report on Entries.Static flag and CVS/Tag (how? maybe a "cvs status -d" to give directory status?). There should also be more documentation of how these get set and how/when to re-set them. 184. Would be nice to implement the FreeBSD MD5-based password hash algorithm in pserver. For more info see "6.1. DES, MD5, and Crypt" in the FreeBSD Handbook, and src/lib/libcrypt/crypt.c in the FreeBSD sources. Certainly in the context of non-unix servers this algorithm makes more sense than the traditional unix crypt() algorithm, which suffers from export control problems. 185. A frequent complaint is that keyword expansion causes conflicts when merging from one branch to another. The first step is documenting CVS's existing features in this area--what happens with various -k options in various places? The second step is thinking about whether there should be some new feature and if so how it should be designed. For example, here is one thought: rcs' co command needs a new -k option. The new option should expand $Log entries without expanding $Revision entries. This would allow cvs to use rcsmerge in such a way that joining branches into main lines would neither generate extra collisions on revisions nor drop log lines. The details of this are out of date (CVS no longer invokes "co", and any changes in this area would be done by bypassing RCS rather than modifying it), but even as to the general idea, I don't have a clear idea about whether it would be good (see what I mean about the need for better documentation? I work on CVS full-time, and even I don't understand the state of the art on this subject). 186. There is a frequent discussion of multisite features. * There may be some overlap with the client/server CVS, which is good especially when there is a single developer at each location. But by "multisite" I mean something in which each site is more autonomous, to one extent or another. * Vendor branches are the closest thing that CVS currently has for multisite features. They have fixable drawbacks (such as poor handling of added and removed files), and more fundamental drawbacks (when you import a vendor branch, you are importing a set of files, not importing any knowledge of their version history outside the current repository). * One approach would be to require checkins (or other modifications to the repository) to succeed at a write quorum of sites (51%) before they are allowed to complete. To work well, the network should be reliable enough that one can typically get to that many sites. When a server which has been out of touch reconnects, it would want to update its data before doing anything else. Any of the servers can service all requests locally, except perhaps for a check that they are up-to-date. The way this differs from a run-of-the-mill distributed database is that if one only allows reversible operations via this mechanism (exclude "cvs admin -o", "cvs tag -d", &c), then each site can back up the others, such that failures at one site, including something like deleting all the sources, can be recovered from. Thus the sites need not trust each other as much as for many shared databases, and the system may be resilient to many types of organizational failures. Sometimes I call this design the "CVScluster" design. * Another approach is a master/slave one. Checkins happen at the master site, and slave sites need to check whether their local repository is up to date before relying on its information. * Another approach is to have each site own a particular branch. This one is the most tolerant of flaky networks; if checkins happen at each site independently there is no particular problem. The big question is whether merges happen only manually, as with existing CVS branches, or whether there is a feature whereby there are circumstances in which merges from one branch to the other happen automatically (for example, the case in which the branches have not diverged). This might be a legitimate question to ask even quite aside from multisite features. One additional random tidbit is to note that Eric Raymond has some interest in this sort of thing. The item "Cooperative distributed freeware development" on http://www.ccil.org/~esr/ has a very brief introduction to what he is thinking about. 187. Might want to separate out usage error messages and help messages. The problem now is that if you specify an invalid option, for example, the error message is lost among all the help text. In the new regime, the error message would be followed by a one-line message directing people to the appropriate help option ("cvs -H " or "cvs --help-commands" or whatever, according to the situation). I'm not sure whether this change would be controversial (as defined in HACKING), so there might be a need for further discussion or other actions other than just coding. 188. Option parsing and .cvsrc has at least one notable limitation. If you want to set a global option only for some CVS commands, there is no way to do it (for example, if one wants to set -q only for "rdiff"). I am told that the "popt" package from RPM (http://www.rpm.org) could solve this and other problems (for example, if the syntax of option stuff in .cvsrc is similar to RPM, that would be great from a user point of view). It would at least be worth a look (it also provides a cleaner API than getopt_long). Another issue which may or may not be related is the issue of overriding .cvsrc from the command line. The cleanest solution might be to have options in mutually exclusive sets (-l/-R being a current example, but --foo/--no-foo is a better way to name such options). Or perhaps there is some better solution. 189. Renaming files and directories is a frequently discussed topic. Some of the problems with the status quo: a. "cvs annotate" cannot operate on both the old and new files in a single run. You need to run it twice, once for the new name and once for the old name. b. "cvs diff" (or "cvs diff -N") shows a rename as a removal of the old file and an addition of the new one. Some people would like to see the differences between the file contents (but then how would we indicate the fact that the file has been renamed? Certainly the notion that "patch(1)" has of renames is as a removal and addition). c. "cvs log" should be able to show the changes between two tags/dates, even in the presence of adds/removes/renames (I'm not sure what the status quo is on this; see also item #182). d. Renaming directories is way too hard. Implementations: It is perhaps premature to try to design implementation details without answering some of the above questions about desired behaviors but several general implementations get mentioned. i. No fundamental changes (for example, a "cvs rename" command which operated on directories could still implement the current recommended practice for renaming directories, which is to rename each of the files contained therein via an add and a remove). One thing to note that the status quo gets right is proper merges, even with adds and removals (Well, mostly right at least. There are a *LOT* of different cases; see the testsuite for some of them). ii. Rename database. In this scheme the files in the repository would have some arbitrary name, and then a separate rename database would indicate the current correspondence between the filename in the working directory and the actual storage. As far as I know this has never been designed in detail for CVS. iii. A modest change in which the RCS files would contain some information such as "renamed from X" or "renamed to Y". That is, this would be generally similar to the log messages which are suggested when one renames via an add and a removal, but would be computer-parseable. I don't think anyone has tried to flesh out any details here either. It is interesting to note that in solution ii. version numbers in the "new file" start where the "old file" left off, while in solutions i. and iii., version numbers restart from 1.1 each time a file is renamed. Except perhaps in the case where we rename a file from foo to bar and then back to foo. I'll shut up now. Regardless of the method we choose, we need to address how renames affect existing CVS behaviors. For example, what happens when you rename a file on a branch but not the trunk and then try to merge the two? What happens when you rename a file on one branch and delete it on another and try to merge the two? Ideally, we'd come up with a way to parameterize the problem and simply write up a lookup table to determine the correct behavior. 190. The meaning of the -q and -Q global options is very ad hoc; there is no clear definition of which messages are suppressed by them and which are not. Here is a classification of the current meanings of -q; I don't know whether anyone has done a similar investigation of -Q: a. The "warm fuzzies" printed upon entering each directory (for example, "cvs update: Updating sdir"). The need for these messages may be decreased now that most of CVS uses ->fullname instead of ->file in messages (a project which is *still* not 100% complete, alas). However, the issue of whether CVS can offer status as it runs is an important one. Of course from the command line it is hard to do this well and one ends up with options like -q. But think about emacs, jCVS, or other environments which could flash you the latest status line so you can see whether the system is working or stuck. b. Other cases where the message just offers information (rather than an error) and might be considered unnecessarily verbose. These have a certain point to them, although it isn't really clear whether it should be the same option as the warm fuzzies or whether it is worth the conceptual hair: add.c: scheduling %s `%s' for addition (may be an issue) modules.c: %s %s: Executing '%s' (I can see how that might be noise, but...) remove.c: scheduling `%s' for removal (analogous to the add.c one) update.c: Checking out %s (hmm, that message is a bit on the noisy side...) (but the similar message in annotate is not affected by -q). c. Suppressing various error messages. This is almost surely bogus. commit.c: failed to remove tag `%s' from `%s' (Questionable. Rationale might be that we already printed another message elsewhere but why would it be necessary to avoid the extra message in such an uncommon case?) commit.c: failed to commit dead revision for `%s' (likewise) remove.c: file `%s' still in working directory (see below about rm -f analogy) remove.c: nothing known about `%s' (looks dubious to me, especially in the case where the user specified it explicitly). remove.c: removed `%s' (seems like an obscure enough case that I fail to see the appeal of being cryptically concise here). remove.c: file `%s' already scheduled for removal (now it is starting to look analogous to the infamous rm -f option). rtag.c: cannot find tag `%s' in `%s' (more rm -f like behavior) rtag.c: failed to remove tag `%s' from `%s' (ditto) tag.c: failed to remove tag %s from %s (see above about whether RCS_* has already printed an error message). tag.c: couldn't tag added but un-commited file `%s' (more rm -f like behavior) tag.c: skipping removed but un-commited file `%s' (ditto) tag.c: cannot find revision control file for `%s' (ditto, but at first glance seems even worse, as this would seem to be a "can't happen" condition) 191. Storing RCS files, especially binary files, takes rather more space than it could, typically. - The virtue of the status quo is that it is simple to implement. Of course it is also simplest in terms of dealing with compatibility. - Just storing the revisions as separate gzipped files is a common technique. It also is pretty simple (no new algorithms, CVS already has zlib around). Of course for some files (such as files which are already compressed) the gzip step won't help, but something which can at least sometimes avoid rewriting the entire RCS file for each new revision would, I would think, be a big speedup for large files. - Josh MacDonald has written a tool called xdelta which produces differences (that is, sufficient information to transform the old to the new) which looks for common sequences of bytes, like RCS currently does, but which is not based on lines. This seems to do quite well for some kinds of files (e.g. FrameMaker documents, text files), and not as well for others (anything which is already compressed, executables). xdelta 1.10 also is faster than GNU diff. - Karl Fogel has thought some about using a difference technique analogous to fractal compression (see the comp.compression FAQ for more on fractal compression, including at least one patent to watch for; I don't know how analogous Karl's ideas are to the techniques described there). - Quite possibly want some documented interface by which a site can plug in their choice of external difference programs (with the ability to choose the program based on filename, magic numbers, or some such). 192. "cvs update" using an absolute pathname does not work if the working directory is not a CVS-controlled directory with the correct CVSROOT. For example, the following will fail: cd /tmp cvs -d /repos co foo cd / cvs update /tmp/foo It is possible to read the CVSROOT from the administrative files in the directory specified by the absolute pathname argument to update. In that case, the last command above would be equivalent to: cd /tmp/foo cvs update . This can be problematic, however, if we ask CVS to update two directories with different CVSROOTs. Currently, CVS has no way of changing CVSROOT mid-stream. Consider the following: cd /tmp cvs -d /repos1 co foo cvs -d /repos2 co bar cd / cvs update /tmp/foo /tmp/bar To make that example work, we need to think hard about: - where and when CVSROOT-related variables get set - who caches said variables for later use - how the remote protocol should be extended to handle sending a new repository mid-stream - how the client should maintain connections to a variety of servers in a single invocation. Because those issues are hairy, I suspect that having a change in CVSROOT be an error would be a better move. 193. The client relies on timestamps to figure out whether a file is (maybe) modified. If something goes awry, then it ends up sending entire files to the server to be checked, and this can be quite slow especially over a slow network. A couple of things that can happen: (a) other programs, like make, use timestamps, so one ends up needing to do "touch foo" and otherwise messing with timestamps, (b) changing the timezone offset (e.g. summer vs. winter or moving a machine) should work on unix, but there may be problems with non-unix. Possible solutions: a. Store a checksum for each file in CVS/Entries or some such place. What to do about hash collisions is interesting: using a checksum, like MD5, large enough to "never" have collisions probably works in practice (of course, if there is a collision then all hell breaks loose because that code path was not tested, but given the tiny, tiny probability of that I suppose this is only an aesthetic issue). b. I'm not thinking of others, except storing the whole file in CVS/Base, and I'm sure using twice the disk space would be unpopular. 194. CVS does not separate the "metadata" from the actual revision history; it stores them both in the RCS files. Metadata means tags and header information such as the number of the head revision. Storing the metadata separately could speed up "cvs tag" enormously, which is a big problem for large repositories. It could also probably make CVS's locking much less in the way (see comment in do_recursion about "two-pass design"). 195. Many people using CVS over a slow link are interested in whether the remote protocol could be any more efficient with network bandwidth. This item is about one aspect of that--how the server sends a new version of a file the client has a different version of, or vice versa. a. Cases in which the status quo already sends a diff. For most text files, this is probably already close to optimal. For binary files, and anomalous text files, it might be worth looking into other difference algorithms (see item #191). b. Cases in which the status quo does not send a diff (e.g. "cvs commit"). b1. With some frequency, people suggest rsync or a similar algorithm (see ftp://samba.anu.edu.au/pub/rsync/). This could speed things up, and in some ways involves the most minimal changes to the default CVS paradigm. There are some downsides though: (1) there is an extra network turnaround, (2) the algorithm is not as efficient with network bandwidth as difference type programs (it transmits a fair bit of data to discover what a difference program discovers locally). b2. If one is willing to require that users use "cvs edit" before editing a file on the client side (in some cases, a development environment like emacs can make this fairly easy), then the Modified request in the protocol could be extended to allow the client to just send differences instead of entire files. In the degenerate case (e.g. "cvs diff" without arguments) the required network traffic is reduced to zero, and the client need not even contact the server.