Hacker News new | past | comments | ask | show | jobs | submit login
Still All on One Server: Perforce at Scale (2011) [pdf] (perforce.com)
60 points by guessmyname 9 months ago | hide | past | web | favorite | 14 comments

This is correct, but perforce got them to the level they could create their own tooling around git to make it work.

Google replaced perforce with Piper, a perforce compatible system written in-house. This system was not based on git.

There were systems that allowed you to mirror portions of piper into local git repo, giving you a git-like experience on top of piper, but the underlaying system was piper and not git. (At this point my knowledge is 4 years out of date.)

Git? I thought Google uses a custom system?

FWIW, the metadata file locking was improved with lockless reads as of the 2013.3 release, which made a huge difference to ease of scalability.

Various customers with repos in the PB range these days (metadata TB).

Disclaimer: current Perforce employee

Disclaimer: current Perforce employee

That's a claimer, rather than a disclaimer!

Ha! Remember when source control used locks! Waiting 30-60secs for p4 commands...

"we went from seeing the server blocked routinely for up to ten minutes at a time in 2005 to gaps of thirty seconds to a minute in 2010..."

Now instead, I run git pull; git push in a loop and hope I can get in before a coworker. Progress?

Do you really have that many coworkers that people are pushing every few seconds on the same branch?

Yes, during the work day, many people would like to commit their changes, often during the same timeframe. Every step takes a second or two, so there's plenty of room to get conflicts (I moved closer to the git server, so now I have an advantage!)

CVS was nice(ish) since changes were independent, any number of people could check in code to limit of io parallelism. Svn was nice, the server would wait for a lock and if my changes didn't conflict, they would be applied after the rest of the pending commits. Git is a PITA, large companies build elaborate systems to queue commits, but we're not big enough to use that yet.

Does everyone in your company push to master all the time? Or is it different PRs being merged all the time?

Pretty much yes. We like to push to master, and then push to production. Some teams cherry pick changes to a deployment branch and deploy from there instead of master.

Edit to add: We may be "doing it wrong", but this is the least effort way to do centralized development.

No seriously, consider git branches. It's gotten 100x easier to create, manage, and merge branch pull requests. Most projects I join these days in fact "lock" master and you are forced to develop in branches.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact