Chris Marusich <cmmarus...@gmail.com> skribis: > No, I'm not using that at the moment. In the future, if I set up any > nginx servers to accomplish the same task, I will definitely use it. I > would prefer to run my own servers, but for now this is something I can > do immediately to help the project, so I decided to do it. > > CloudFront is a service, so you use its API (or the AWS Management > Console, which is a web UI for the API) to create a "distribution" and > configure it to use hydra.gnu.org as its "origin". A little extra work > is required to glue everything together. For example, I had to create a > CNAME pointing from hydra-mirror.marusich.info to > d2xj50ygrk34qq.cloudfront.net, which is the canonical name of my > distribution. Once it's configured, all requests sent to > hydra-mirror.marusich.info are serviced by a nearby point of presence in > the CloudFront content distribution network, and the results are cached.
OK. Do you know exactly how much is cached? Also, when does caching happen? When using nginx as a proxy as on mirror.guixsd.org, it fetches things lazily, so on a cache miss it goes connect to hydra.gnu.org. > I've noticed that Hydra does not include cache-related headers (e.g., > Cache-Control). Perhaps for this reason, the nginx config you linked > seems to pick arbitrary caching settings. When using CloudFront, a > distribution can be configured to respect the Cache-Control headers sent > by the origin server. Does nginx provide similar functionality? Would it > make sense to have hydra.gnu.org return such headers? Dunno, maybe! Maybe we could tell nginx to add such headers? What would be the right thing? > For now, I've configured hydra-mirror.marusich.info to cache all > successful requests for 1 week, and to respect cache-related headers > from the hydra.gnu.org if it ever decides to send them. This seemed like > a reasonable configuration for data which is not expected to change. Yeah, mirror.guixsd.org also caches for one week now. I wasn’t sure how much disk space that would represent, but so far we’re around 10G, so increasing to 1 week seemed reasonable. Thanks, Ludo’.
