AlliedModders Planning

The complexity of SourceMod's branching (even as trivial as it is) is too much for Subversion and our build process. There are a bunch of things at play here, and not all are related. This is sort of a major push for AlliedModders to be a bit more streamlined as we're having trouble accommodating multiple projects, documentation sets, build systems, mirrors etc.

This document discusses some changes to the build process and a possible move to Mercurial.

Some of the benefits we'd gain from the ideas in this document:

• Much easier branching
• Merging changesets would be FAR easier
• Merges between branches would be tracked
• All the benefits of a distributed system -- pulling changes from other parties is much easier
• Getting rid of revision numbers
• Making fetchdlls work seamlessly
• Getting rid of package/symbol svn and doing 30day rotation to save disk space

Overview

Subversion doesn't let us branch nicely or create user-specific branches without forever mucking up the central repository. It's difficult to do merges. Mercurial seems to solve these problems, but our build process will be majorly affected.

The Revision Problem

Mercurial tracks revisions with 160-bit IDs (40 digits of hex), called "changeset IDs." Mercurial has aesthetic revision numbers only; they're not guarantees like they are in Subversion. Revision numbers are properties of a repository clone, not the overall repository. This means that if two people make changes locally, those revisions will be different, and will change again once pushed upstream.

This problem is somewhat mitigated by the fact that the central repository will have reliable revision numbers, but that's still not good enough since branching will generate multiple revisions. Branching also makes the revision numbers awkward since users assign meaning to them where none exists (for example, "1.0.1.1234 versus 1.1.0.1231 doesn't make sense because 1.1 is later than 1.0.1").

This means we must change SourceMod to not use revision numbers, a change we should make regardless of a Mercurial merge. The major problems to keep in mind are: What do we display for the version number? and How do we keep fetchdlls working? (Fetchdlls needs to be able to get a changeset ID from the MINIDUMP_MODULE struct, which has a checksum and version info.)

Solutions

There are a few ways to go about this. These are the two I came up with while writing this.

• Revision numbers that are per major release. We'd have a "build number server" that would dish out new numbers on request. These numbers would be used by the build scripts.
  • A decent bit of work but would preserve the most compatibility with fetchdlls.
  • [svnrev] (or [hgrev] or whatever) would have mismatching revision/build # ids which could get confusing. Thus we'd have to try and not display the revision numbers publicly (they'd only be used for linking).
• Simply assign 0 to the build number and ignore it entirely.
  • Seems easy but we'd need to have a mapping between VS's "checksum" values and a changeset ID if we wanted fetchdlls to work. Not too difficult really as this could be done with a simple HTTP push.
  • For visual display, we'd totally eliminate the build number. It'd only be visible on Windows from the file information.
  • We'd add a secret date+time field to plugins that would be filled in by the compiler.

For plugins, displaying a timestamp is not a problem. We could put some sort of "public String:__date[] = __DATE__" into core.inc.

Right now, I'm preferring the second solution. There's no significance to a build number if we can replace it with a timestamp, and a checksum -> changeset mapping is trivial.

Automation

Right now we use subversion for all of the packages. This is getting to be troublesome. It makes fetchdlls extraordinarily complicated when it really does not have to be, and the hourly builds just accumulate with no way of being deleted. The package repository is 1.7GB and growing.

It would be a lot simpler if we could just throw a single package on the webserver and forget about it. We could even keep a rotation of, say, 30 days' worth of packages in a rotation. This would let us remove cruft and keep enough binaries around for debugging. The problem is that people rely on the repositories to perform diffs.

I don't have any good ideas on solving that problem, because there's no way to selectively rotate things out of a versioning system. The only solution I can think of is providing people with, say, a perl or bash script that does the job for them. This could go through two routes: using local repositories, or simply doing a HTTP fetch from the web server and then doing a diff.

License Changes

I don't want to keep hacking up our private->public sync script. DS has faithfully maintained it until now, but I believe he will be as glad as I to do away with it. A move to Mercurial would have either of the following benefits:

• We could keep the JIT closed source and shove it in a private repository, OR
• Just open source the thing finally, it's not very special, and have everything be together.

Phases

Phase 0

• Apply new versioning scheme, whatever it may be

Phase 1

• Redo the automation so it simply sends the packages upstream to FTP
• Use an SSH tunnel to notify of new packages, so we don't have to do something idiotic like mirrors.pl for the mirror system
• Redo the package system so only 30 days worth of builds are available
• Write a script to help users diff between packages?
• Transition C# build tool to vs2k5 during this downtime
• Remove the Packages repository entirely (transitioning AMX Mod X to this will be trivial).

Phase 2

• Move the pdbstore to be centralized client-side. This can be accomplished by making a more advanced symstore wrapper that will remove anything older than 30 days, and then (in a similar SSH tunnel method as above), rsync all changes upstream to the webserver. The exact mechanism is still up in the air, but rsync would probably suffice (as long as we can stomach cygwin).
• Remove the Symbols repository.

Phase 3

(Maybe)

• Add hg.alliedmods.net pointing to our "repository roots," which will be per-project. For example, hg.alliedmods.net/sourcemod.
• From there, import trunk as "sourcemod."
• Branch every major release off as separate repositories, from the appropriate revisions.
• Tag minor releases.
• Update all links we can find on the site and in the docs.

Phase 4

(Assuming Phase 3)

• Add /metamodsource, import trunk and branches.
• Branch every major release off as separate repositories, from the appropriate revisions.
• Tag minor releases.
• Update all links we can find on the site and in the docs.
• Create new repositories for the hl2sdk stuff. This way, the AlliedModders root organization will look like:
  • hl2sdk
    • hl2sdk
    • hl2sdk-ob
  • metamodsource
    • metamodsource
    • metamodsource-1.0
    • metamodsource-1.1
    • metamodsource-1.2
    • metamodsource-1.3
    • metamodsource-1.4
    • metamodsource-1.6
  • sourcemod
    • sourcemod
    • sourcemod-1.0

Phase 5

We really, really still need a CMS to make this packaging, mirroring, etc business much more streamlined. But this phase is so far in the future I expect other phases to creep in before it. This is more a mind dump than anything, but here goes --

• Moving from flyspray to bugzilla. After using bugzilla, I'm convinced that its massive feature set far outweights the few permissions issues we ran into. flyspray is severely lacking, and even though bugzilla isn't the nicest looking tool, it's templated and professionally done. It has little things that really count, like..
  • Not e-mailing the person who made the change to a report
  • A much, MUCH better attachment system that allows for attachment reviews
  • More focused on triaging and patch management than flyspray, which is simply a list of tasks with some basic statistics
• Moving from MediaWiki, which (I view) as a very, very outdated piece of software. From version 1.5 to 1.10 I have noticed no real usability or feature set changes, and the installation/upgrade process is always very buggy for us. Its administration and permissions model is laughably limited. A possible target for this is DekiWiki which Mozilla is supposedly converting to. It's open source and it's actually a CMS, which (perhaps) could kill two birds with one stone.