Use container based builds by default?

[hadley]

• sudo: false
• Figure out how to take advantage of caching

[gaborcsardi]

Yeah, I was thinking about that, too. Not easy, though, because you cannot do apt-get install without sudo.....

I guess once R is a first class language on Travis-CI, this problem largely goes away, so it might not be worth "fixing" the current r-travis.

[hadley]

@gaborcsardi we should be working on getting R to be a first class language then!

[gaborcsardi]

@hadley, Yep, @craigcitro will do that soon: https://twitter.com/GaborCsardi/status/542712472155983873

FWIW, I have binary builds for many R versions and R-devel (hourly) at https://github.com/metacran/r-builder/releases

Using these it is easier to get rid of sudo, the bigger pieces still installed by apt-get are TeX, qpdf and gfortran: https://github.com/metacran/r-builder/blob/master/pkg-build.sh

TeX is a really a beast, but I guess we can just pull this from a binary as well, and then cache it. @krlmlr worked on this if I remember well.

I am kind of ambivalent about caching, btw., it makes the build much faster, but it also introduces a lot of potential errors if you do not start from a clean state. Anyway, caching tex, gfortran, etc. is probably fine. Maybe the R binaries are also fine. Caching the R library directory is more risky imho.

[krlmlr]

I have prepared both R and TeXlive binaries in two GitHub repos for running on Snap CI, but this is CentOS. It should be easy to port these to Travis/Ubuntu. gfortran and qpdf can be provided in a similar fashion.

To take advantage of the cache, there needs to be some mechanism to detect that the cached binaries are out of date. Git offers this -- if R or TeXlive are already cached, a simple git pull suffices to get the latest updates with minimal overhead. That's the reason why the stuff is in a GitHub repo and not in an archive. Perhaps there are better ways to achieve this, but this now works for me well enough for quite some time.

[gaborcsardi]

@krlmlr I think the git repo is elegant, but it might be overkill. Unless you prune older files, it will be huge soon, because you are storing binaries. (I guess, I might be wrong.) I am also not sure about Github's policy about storing essentially binaries in the repos....

An alternative solution is simply using a hash. The hash of the tarball can be included in the file name, and then if it is newer than the cached version, then we can download it.

[hadley]

I think the cache will be most important for r packages, as other big dependencies should be included in top level language support.

@gaborcsardi is right that it's easy to cause a lot of problems with caching, so it probably shouldn't be the default. But it will be useful for packages with many deps like dplyr and devtools.

[gaborcsardi]

I think the cache will be most important for r packages, as other big dependencies should be included in top level language support.

Ideally yes. But if Travis keeps relying on Ubuntu 12.x, then we need to install a newer TexLive, which might not be possible at the top level. Anyway, this is too speculative. We'll see.