https://pantsbuild.org/ logo
Powered by
#general
  • g

    gentle-potato-32845

    02/26/2016, 12:48 AM
    ok varnish is what i remember imaging it was
  • g

    gentle-potato-32845

    02/26/2016, 12:48 AM
    just making sure im not crazy
  • w

    witty-crayon-22786

    02/26/2016, 12:48 AM
    but yea, varnish+nginx should go a very long way.
  • w

    witty-crayon-22786

    02/26/2016, 12:49 AM
    @bored-art-40741: would it be easier to just scale that cache out?
  • b

    bored-art-40741

    02/26/2016, 12:49 AM
    Yeah if you're trying to setup an artifact cache quickly and on the cheap, I can provide our current config @gentle-potato-32845
  • w

    witty-crayon-22786

    02/26/2016, 12:49 AM
    N instances or whatever
  • b

    bored-art-40741

    02/26/2016, 12:49 AM
    I should really check it in as a demo
  • g

    gentle-potato-32845

    02/26/2016, 12:49 AM
    @bored-art-40741: that would be ❤️
  • r

    rough-minister-58256

    02/26/2016, 12:49 AM
    couldn’t you just put a layer of nginx instances in front of it acting as a load balancer?
  • w

    witty-crayon-22786

    02/26/2016, 12:50 AM
    or that. unless you're write bound.
  • b

    bored-art-40741

    02/26/2016, 12:50 AM
    It depends a bit
  • b

    bored-art-40741

    02/26/2016, 12:50 AM
    We're definitely read bound
  • r

    rough-minister-58256

    02/26/2016, 12:50 AM
    if the same workload saturating the nic is write traffic, then uh
  • r

    rough-minister-58256

    02/26/2016, 12:51 AM
    id be real surprised
  • w

    witty-crayon-22786

    02/26/2016, 12:51 AM
    heh
  • b

    bored-art-40741

    02/26/2016, 12:51 AM
    If we were write bound on scalac output then we wouldn't need a remote cache
  • b

    bored-art-40741

    02/26/2016, 12:51 AM
    And also I'd eat my hat
  • r

    rough-minister-58256

    02/26/2016, 12:51 AM
    heh yup
  • b

    bored-art-40741

    02/26/2016, 12:52 AM
    But there's also the question of whether you're using a virtual IP, round robin DNS, or a HAProxy instance to switch between these load balancers (and whether or not you want to do sharding as well)
  • r

    rough-minister-58256

    02/26/2016, 12:52 AM
    this looks interesting: http://ceph.com/
  • b

    bored-art-40741

    02/26/2016, 12:52 AM
    RR DNS is a little iffy because honestly I'm not sure to what extent a single process respects it
  • r

    rough-minister-58256

    02/26/2016, 12:52 AM
    kinda like the utilitarian approach of varnish/nginx tbh
  • g

    gentle-potato-32845

    02/26/2016, 12:52 AM
    heh ceph
  • r

    rough-minister-58256

    02/26/2016, 12:53 AM
    doesnt seem like a good best of breed open source object store exists afaict
  • b

    bored-art-40741

    02/26/2016, 12:53 AM
    yeah
  • w

    witty-crayon-22786

    02/26/2016, 12:53 AM
    Screen Shot 2016-02-25 at 4.53.08 PM.png
  • b

    bored-art-40741

    02/26/2016, 12:54 AM
    But yeah, we came up with what Aaron is working now because the backing store is already fault tolerant, and we want the frontend to it to be as simple as possible and trivial to scale up
  • b

    bored-art-40741

    02/26/2016, 12:55 AM
    We're going to do another round of tweaking once we see how this works out
  • b

    bored-art-40741

    02/26/2016, 12:55 AM
    Ditching HDFS is definitely a possibility
  • b

    bored-art-40741

    02/26/2016, 12:55 AM
    In practice I don't think something on the order of 10k small files is going to burn the cluster down--I'm on the hadoop team, and I'm familiar enough with what normal cluster usages look like
1...101410151016...2420Latest