Converting from CVS to Subversion?

Bob Bobbinson asks: "I'm currently looking to convert my workplace's CVS repository over to Subversion. The main issue I'm having at the moment before I can commence a roll-out is how we are currently using tags, and sticky tags. The project is a website and we have two tags which are used to put changes from the main trunk live (internally, and externally). All development is committed to the main branch, and when we want a change to go live we move the two 'live' tags to that current version. The live servers are both sticky tagged to these tags, so when we run an update on these servers they will only update to the version that the 'live' tags refer to. Currently I haven't found a satisfactory way to replicate this in Subversion, as moving tags, and updating on separate servers seems to be quite kludgy i.e. need to remove current tagged version from tag, then copy from the main branch over to the tag, then update this on the live server. So I'm trying to look for an alternative way to implement this staggered releasing of code live, and maintaining the ability to see what versions of files constitute what is live. Any ideas?"

Comments

KDE

[JabberWokky]

KDE just did a massive update including all tools and reporting software. It is the largest software repository running on Subversion. You might look into what they did.

--
Evan

use the check-in numbers

[maraist]

First, each checkin should ideally be a legitimate build.. You shouldn't check in to the main branch things that are knowingly broken.

I assume you have a staging environment which would pull from the latest main-branch build.. It's easy to see what checkin version staging is using..

Once you have staging in an acceptible state ready for deployment, simply take the associated checkin-number (this SHOULD be a manual process) and post that to production.

In other words, ditch the remapping of tags as you used in CVS. This is an error prone process which destroys history.

In our development environment, we have a wiki that we (manually) update whenever we post things, and each deployment (to test, stage, production instanceX) has an associated svn checkin number.

If there is some memorable tag that we need, we svn cp a tag.. And obviously long-term modifications get their own branches.

While I'm not as happy with SVN's lack of relocatable tags, I do really like the idea that when using svn + fsfs as a backend, you have read-only back-versions. You can see everything associated with the checkin-number. CVS, on the other hand only lets you see the current state of the rcs files.. Sure there's a lot of history in there, but there is no associated "undo".. Yout have to use the cvs tools, and then you have to keep your fingers crossed that you're doing it right.

Re:use the check-in numbers

[EnronHaliburton2004]

First, each checkin should ideally be a legitimate build.. You shouldn't check in to the main branch things that are knowingly broken.

Whoa there cowboy. There will be many cases where this isn't always practical.

For example, lets pretend you have a programmer working on the application code, and a designer working on the layout of a program. In this case, there is going to be a point where one person checks in their changes, which leaves the main branch in a broken state until the second person checks in their code.

You could solve this problem by creating an interim branch, and then checking that branch into main; but that adds alot of work to what is otherwise a simple, fast process.

Do you have any other suggestions for this situation. I'd be willing to bet this is the case for close to 50% of code checkins.

Re:use the check-in numbers

[dubl-u]

Do you have any other suggestions for this situation. I'd be willing to bet this is the case for close to 50% of code checkins.

On your projects, perhaps. On mine, never. I use continuous integration tools like Anthill and Cruise Control, and anytime a checkin breaks the build, the developers hear about it pronto. This is a little annoying to start, but once you get broken in it's fabulous: what's in CVS is always trustworthy, and if you see a problem you know it's your problem.

If you're just using CVS to pass around files, then you could just email the files around. If you keep the units of work small, clear, and discreet, this works fine.

Another option is to have areas in CVS for things that are half finished; the person to do the last bit of work moves things to their proper place.

With the designer/programmer split, one helpful approach is to make the last step connecting it to the rest of the app. So if it's a web app, you put up the new pages in a way where you can get to them only if you know the URLs. Then only when they're perfect do you add links from other pages.

You can also make features configurable, so that a feature is only visible if certain configuration options are set. This is especially valuable in environments where you have to roll out multiple related servers but need to bring everything live at the same time.

But my favorite approach is to get the necessary people together in one place. E.g., the programmer pairs with the designer, and you get things done in one go. Not only is it faster and more likely to work, but both sides learn a lot more about how the other guy works.

Re:CVS was designed for text files...

[EnronHaliburton2004]

I mean, if it's not broken... why fix it?

Because there are a ton of features in CVS which are broken? Moving directories, versioning metadata & permissions; etc.

I'm not really into SVN's versioning system either. I suppose I could get used to it, but I still like the flexibility of CVS tags. But there are many compelling reasons to move from CVS to SVN.

If I understand correctly, you want...

[danpat]

svn co repos/trunk

svn commit

When /trunk is ready to go live:
svn rm repos/livetag
svn cp repos/trunk repos/livetag

Then, on the live servers:
svn update (where they have working copies of repos/livetag)

OR, you could do:
svn cp repos/trunk repos/tag-20050403
(i.e. create a unique tag for each release)

then, on the live machines do:
svn switch repos/tag-20050403

Either way should do the trick. And as other posters have pointed out, have you asked this question on the Subversion mailing list?

Re:If I understand correctly, you want...

[bigbadbob0]

I have a dev server that is always on trunk. When trunk is ready to push up I tag it. I switch to stage and do a svn switch to the new tag. I check for bugs on stage, address them on the tag and commit, merging up to trunk. When all is ready I do a final svn switch on the live site.

Re:If I understand correctly, you want...

[jbellis]

It's not exactly hard to prevent apache etc to not serve .svn directories.

It's very worth it in terms of how much easier it is to sync with the latest bugfixes in your live branch.

Just stop thinking in terms of "sticky" tags...

[photon317]

In the subversion philosophy of doing things, ideally a tag shouldn't ever get changed. A tag should be pretty much a snapshot in time that stays as it is forever.

For instance, here's an example of branching/tagging the way subversion thinks of it (although you are free to twist it to other methods):

I have my main trunk, which lines in "trunk/..." in my repository. The trunk is at version 31666 right now. I decide that rev 31453 (the last good stable build from 2 days ago before I started mucking with a new feature) will be the branching point for my 1.0 stable branch. So I issue an "svn copy" command to copy rev 31453 of the turnk directory to a new place in the repository called "branches/1.x-stable". I tell the team that that branch is in feature freeze - debug it. They find a few bugs, which they commit to either the branches/1.x-stable tree, or the main trunk, and in either case they also copy their change to the other if it's applicable.

After a few weeks of debugging like this, I decide that at repository revision 31942, the "branches/1.x-stable" branch is stable enough to release the real 1.0. So I issue an "svn copy" command to copy rev 31942 of "branches/1.x-stable" to "tags/1.0". I never update this tree in tags/1.0 again. It essentially serves as a permanent record of what we released as 1.0, and a convenient spot to pull the 1.0 release from for packaging (or for comparison down the road).

After giving 1.0 to users, based on user feedback 3 new bugs were found in the 1.0 release. The coders fix these in the "branches/1.x-stable" tree, and reflect those changes back to the trunk if they're applicable (trunk may no longer have the affected code for all we know at this point). We decide that the 3 bugs were serious enough to warrant a 1.1 release to customers, so we make yet another "svn copy" off of "branches/1.x-stable", after the 3 fixes are in, as "tags/1.1", and release that copy as 1.1.

etc...etc...

To apply things to your situation (where you seem to essentially be using tagging but no branching, which is something you may wish to rethink down the line), essentially what you would do is every time you want to publish a "stable" copy internally or externally, you would just make a brand new tag. Instead of "updating" a tag named "internal" or "external" with your new code, you would create a new tag named "internal-1.1.2" or whatever you choose to version it as - but never change old tags, just make new ones.

Once you've done "svn copy" to make your new internal or external release tag, then you go to the actual webserver where you want to check it out, and in place of your previous "cvs update" on a checkout of a sticky "external" tag, you instead do an "svn switch" command from "tags/external-1.1.1" to "tags/external-1.1.2", which will alter the working copy as minimally as neccesary to switch you over to the new tag.

Re:Just stop thinking in terms of "sticky" tags...

[Parity]

I don't think that does what the poster wants to
do though.

The poster wants, I believe, to have a 'live'
release, so that some 'get the current "live"'
will always get the code considered to be
released. For example, by a cron-job doing
an update at 3:47 AM each night.

This kind of thing is not really amenable to doing an 'svn switch', which requires knowing whether or not things have changed and what the new version is called.

Perhaps the correct answer is 'write a more complex cron job that detects whether there is
a tag 'live-.*' where .* is a number greater
than the current version.'

Or perhaps the correct answer is to use a branch.

Or perhaps something else that a subversion user
would know.

(I'm not, and don't know, the original poster -
but having worked with CVS as web-site tracker
and CVS as compiled-code tracker... the poster
sounds like he's doing something much more like
the former... )

Re:Just stop thinking in terms of "sticky" tags...

[photon317]

If it has to be a cron job type thing that updates the internal and external websites to the "sticky" tags, then yeah I would recommend the cronjob pull the version from something else. One idea would be something like this:

myproj/trunk/[...]
myproj/tags/external-1.1/[.. .]
myproj/tags/external-1.2/[...]
myproj/externa l-release

Where the tags work as I outlined, and "myproj/external-release" is a text file which is manually updated to read "1.1" or "1.2" or whatever.

Then the cronjob on the production server essentially does

svn checkout svn://mysvnsvr/myproj/external-release /tmp/relnum
svn switch svn://mysvnsvr/myproj/tags/external-`cat /tmp/relnum` /web/root

Or what-have-you for the specifics.

use branches, OR revision numbers

[Anonymous Coward]

First of all, Subversion has no tags. They might call them "tags" but they are really just branches. I consider this a bit of a flaw but whatever.

What you need to do is either

1) remember the revision number of the repository which has the "good" code. Write it on a stick note or something. This is as close to a "tag" as you'll get. Then just tell your build process "revision #123124 is the one to use".

or

2) use a separate stable branch, and ONLY merge known working code to this branch.

I used to think #2 was almost impossible in Subversion (no tags!) but then I discovered svnmerge. See if you can find it, I can't find it anywhere in my FreeBSD or Gentoo installs and I had to download it separately. But it's a godsend! It's even *easier* than using floating tags in CVS to keep track of merge points.

Basically svnmerge will remember which revisions (specifically, which diffs between pairs of revisions) have been applied to a branch, by storing the information in properties. I have a big project with multiple long-lived branches (one for each client), and I could not figure out how to do this in Subversion until I discovered svnmerge.

In CVS it's easy ("easy" meaning, I can write down a set of instructions that can be mindlessly followed to get the right results), just create a floating tag that represents the last point you merged from. One on each branch for two-way merges.

svnmerge makes this easy in subversion too.

(It would be even *easier* if Subversion tracked changesets instead of tree revisions but that's a rant for another day.. Darcs is looking promising though).

Re:use branches, OR revision numbers

[Bassman59]

First of all, Subversion has no tags. They might call them "tags" but they are really just branches. I consider this a bit of a flaw but whatever.

1) remember the revision number of the repository which has the "good" code. Write it on a stick note or something. This is as close to a "tag" as you'll get. Then just tell your build process "revision #123124 is the one to use".

Ah, you don't get it -- to "tag" a release, you simply copy the head to your "tag." You can even mark the tag as read-only. Then you tell the production people to check out the tag. It's really simple.

svn:externals?

[malachid69]

Pardon me if this is completely off-base, but I only switched from CVS to Subversion about a week ago.

Couldn't you use svn:externals to do what you want? I mean, couldn't you have a 'live' project or something that simply listed whatever tag you wanted in your externals? Then, whenever you get the 'live' site, it would get whichever tag was currently referenced?

Instructions are at http://svnbook.red-bean.com/en/1.0/ch07s03.html

Re:Clarification

[brunson]

You prefer CVS over SVN because you can do a deployment with two (non-atomic) commandos (tag,update) instead with three (atomic) ones (remove,copy,update)?

I prefer atomic commandos. GI Joe, the great american hero!

Subversion software -vs philosphy

[MobyDisk]

Subversion doesn't support "tagging" like you call it. There are constant of posts the subversion mailing lists about this. 90% of them are ignored, and a few get responses. They prefer to use their branching concept.

This is a case where the authors of Subversion are trying to force their philosphy of how to use version control on the users. I understand that: It is nice when your users are using the software the way that yoy think is best. But that isn't always possible, and they've removed a feature that is common to 90% of all source control systems.

That makes it very hard to migrate to Subversion. There are many build tools, developer's brains, and business procedures that depend on the concept of tagging. Their arguments against it only take into account one side of things. And unfortunately, it has become more of a holy war with the developers ignoring legitimate reasons to support it rather than address them. I really hope this changes, or I fear Subversion will never gain the "critical mass" that CVS has.

With that complaint out of the way, you can somewhat do what you want with two ways. One is to use a branch instead of a tag, and have some way to identify which branch is the "live" branch. A text file. A custom field on one of the files. A rule like "live-###" where you take the largest number. These are your best bet.

If you use a system where you have a large hierarchy, and you cannot deal with all those branches showing up on the tree (it can get really messy) than you can delete old branches, or move them elsewhere. If that's not possible, you are SOL, and Subversion won't work for you.