Thanks for the context. So, if I get what you're writing, the goal is too redirect to a node that has the object in cache?
Question is, what's the time needed to: - send a request to the server on the LAN and receive the object - send the redirect across the web, and wait for the client to send a new request, again across the web. If the former is not at least an order of magnitude larger than the latter, I wouldn't bother. The issues I have with your redirection scheme are that: - IIUC, you are basically explaining to people where the backend is, instead of shielding it with Varnish - it doesn't lower the backend traffic - as said, I'm not even sure the user-experience is better/faster -- Guillaume Quintard On Sun, Dec 18, 2016 at 9:21 PM, Anton Berezhkov < [email protected]> wrote: > We have 6 servers with 6TB on each one (same video files). Currently > they're hitting iowait limit of SATA disk (240ops total). At same time, > each server can provide 500Mbit of guaranteed bandwidth. > > With HDD restriction, each server provides about 320Mbit. There is also > problem with fragmentation, which is caused by nature of HTML5 video > players & HTTP (allowing to request partial data with Range header). > > Before this moment, we were scaling horizontally, by duplicating this > servers. > > There is also option to get same server with 2x480GB SSDs. As i > reasearched from nginx logs, 98% of daily traffic lays in ≈800GB of files. > > What i want to achieve: To build a varnish server with 2x480GB SSDs(no > raid), with storage for varnish about 800GBs. Which will guarantedly fill > all available bandwidth for a server. > > Also, I built simple load-balancer, which monitors each server for current > eth0 load (in Mbps) and decide to which one redirect (using HTTP Location > header). > > Request for Video -> LBServer: Find lowest loaded(1 of 6) & Redirect to > LBNode -> LBN: serve request > > To add new HDD-LBN, i need to setup server, sync videos, setup some > additional software. > > My wish: add new SSD-LBN, setup & sync varnish config, and it will build > cached pool itself. > > Why i need redirect? > 1. It will offload bandwidth of SSD-LBN, pass-through will take bandwidth > of both servers + still cause iowait problems on HDD-LBN. > 2. It will "prove" that uncached video will be take from HDD-LBN which > always have all videos. > > Currently all LBN servers are hosted on OVH and we're good with them, > especially because of low price :) > > If you have any suggestions, i'll be glad to hear them :) > > > On Dec 18, 2016, at 10:59 PM, Guillaume Quintard < > [email protected]> wrote: > > > > I think Jason is right in asking "why?". What do you want to achieve > specifically with this behavior? > > > > Varnish has streaming and request coalescing, meaning a request can be > served as soon as data starts being available AND the backend doesn't > suffer from simultaneous misses on the same object. I feel that should > cover almost all your needs, so I'm curious about the use-case. > > > > On Dec 18, 2016 20:27, "Jason Price" <[email protected]> wrote: > > It would be possible to do this with varnish... but I have to ask... why > bother? > > > > If the purpose is to offload the IO load, then varnish is good, but you > need to prime the cache... TBH, what I'd do first is put one or a pair of > varnish boxes really close to the overloaded box, and force all traffic to > that server through the close varnish boxes... using the do_stream feature, > you'll get stuff out there fairly quickly. > > > > After that is working nicely, I'd layer in the further out varnish boxes > which interact with the near-varnish boxes to get their data. > > > > This works well at scale since the local caches offer whatever's useful > local to them, and the 'near-varnish' boxes handle the 'global caching' > world. > > > > This was how I arranged it at $PreviousGig and the outer CDN was getting > a 85-90% cache hit ratio, and the inner tier was seeing 60% cache hit > ratio's. (The inner tier's ratio will depend heavily on how many outer > tier's there are...) > > > > On Sat, Dec 17, 2016 at 8:09 PM, Anton Berezhkov < > [email protected]> wrote: > > This is how I semi-implemented: http://pastebin.com/drDP8JxP > > Now i need to use script which will run "curi -I -X PUT > <url-to-put-into-cache>". > > > > > > > On Dec 18, 2016, at 3:58 AM, Mark Staudinger <[email protected]> > wrote: > > > > > > Hi Anton, > > > > > > Have you looked into the "do_stream" feature of Varnish? This will > begin serving the content to the visitor without waiting for the entire > object to be downloaded and stored in cache. Set in vcl_backend_response. > > > > > > https://github.com/mattiasgeniar/varnish-4.0- > configuration-templates/blob/master/default.vcl > > > > > > Cheers, > > > Mark > > > > > > On Sat, 17 Dec 2016 19:05:48 -0500, Anton Berezhkov < > [email protected]> wrote: > > > > > >> Hello. > > >> > > >> Switched to Varnish from Nginx for additional functionality and > better control of handling requests. > > >> But still can't implement what i want. And I want simple behaviour > "Redirect on MISS/PASS". > > >> I want to use VC for deploying quick "cdn" servers for our > mp4-video-servers (used for HTML5 players), without need to store all files > on this quick (ssd, upto 2x480GB space, full database about 6TB). > > >> > > >> Currently we have 6 servers with SATA HDDs and hitting iowait like a > trucks :) > > >> > > >> Examples: > > >> - Request -> Varnish -> HIT: serve it using Varnish. > > >> - Request -> Varnish -> MISS: start caching data from backend, and > instantly reply to client: `Location: http://backend/$req.url"; > > >> - Request -> Varnish -> UPDATE: see `-> MISS` behaviour. > > >> > > >> From my perspective, i should do this "detach & reply redirect" > somewhere in `vcl_miss` OR `vcl_backend_fetch`, because if i understood > correctly https://www.varnish-cache.org/docs/4.1/reference/states.html, i > need vcl_backend_response to keep run in background (as additional thread) > while doing return(synth(...)) to redirect user. > > >> > > >> Similiar thing is "hitting stale content while object is updating". > > >> But in my case "replying redirect while object is updating". > > >> > > >> Also, i pray to implement this without writing additional scripts, > why? I could do external php/ruby checker/cache-pusher with nginx & etc. > But scared by performance downgrade :( > > >> _______________________________________________ > > >> varnish-misc mailing list > > >> [email protected] > > >> https://www.varnish-cache.org/lists/mailman/listinfo/varnish-misc > > > > > > _______________________________________________ > > varnish-misc mailing list > > [email protected] > > https://www.varnish-cache.org/lists/mailman/listinfo/varnish-misc > > > > > > _______________________________________________ > > varnish-misc mailing list > > [email protected] > > https://www.varnish-cache.org/lists/mailman/listinfo/varnish-misc > >
_______________________________________________ varnish-misc mailing list [email protected] https://www.varnish-cache.org/lists/mailman/listinfo/varnish-misc
