This is exactly how updates should be done. Usually, I'm not that big of a fan of things that Google does, but this is a great achievement. Hats off to those who contributed to this. I hope all update systems eventually work this way.

Well, most binary updates do go through bsdiff, which was designed for exactly this. I'm surprised they were able to beat it by such a margin. Shit, the guy who made bsdiff did his doctoral thesis on the subject (actually, the algorithm used in bsdiff generates patches about 25% larger than the one in the thesis, but at huge performance increase (or, the thesis generates patches 20% smaller)).

Really, I wasn't aware that that many projects used bsdiff.

Yeah, me neither. What projects are we talking about here? I'm in Linux so I'm getting the full deb each time. I'm okay with that considering it's managed for me. I'll take that any day.

Adobe? Apple? Those are the two vendors that I can imagine people having installed on most, if not all Windows computers. And I know for a fact they both distribute updates as full copies of their software waying in at a couple hundred to several hundred megabytes, requiring manual updates, pop up windows and restarts.

I've yet to see anyone do updates as fast or seamlessly as Google.

Call me stupid, but why do you have to disassemble the program to get the symbol pointers when you just finished compiling the thing in the first place?

Couldn't this intermediate step of pointer collection be part of the prelink process and skip all this guesswork?

If the source is in a single language, compiled by a single tool, with one single version, and entirely linked by a single linker strictly at the end of the process, then that might be feasible. But that's probably not the case. If there is much heterogeneity at all, getting proper symbol information may be a lot more work than analysing the binary for control and data flow. For example, intermediate linking steps may resolve symbol fixups internally and just leave a list of relocations behind, or perhaps use strictly relative addresses within what looks to the final linker like a monolithic blob.

If we're talking about a way to generate an upgrade image for a single architecture, taking a known version # of an executable to another known version #, built on a production build machine, I would think that could cover the case easily.

That kind of led to my second question: this makes really small images, but only from one known version to another, right? What happens if the target to be upgraded is 80 revisions behind?

We want smaller updates because it narrows the window of vulnerability. If the update is a tenth of the size, we can push ten times as many per unit of bandwidth. We have enough users that this means more users will be protected earlier.

They're trying to get the vulnerability window down to minutes! If they succeed, this is going to have an impact on the economic viability of running malware.

We're also giving google implicit permission to push code to our machines anytime they want.

We should have conflicting feelings about this.

The last I heard, Google were being sued due to the use of Courgette allegedly infringing a patent and, as a result, they had stopped using Courgette to distribute Chrome updates. Does anyone know the result of that case? Are they now using Courgette again?

http://news.softpedia.com/news/Google-Sued-over-the-Courgett...

Looks like it is still ongoing: http://dockets.justia.com/docket/massachusetts/madce/1:2009c... reports news from Nov 2010. Can't decode the legal language.