We really need to get positive confirmation that this plan is both required and sufficient. Also we need confirmation of usage numbers and dates.
On Wed, Aug 13, 2014 at 5:43 PM, Ben Bangert <[email protected]> wrote: > Hello Dev-Media! > > I'm Ben Bangert, engineer in Cloud Services, leading technical development > on Push. > > This is a general update on the status of simplepush as it pertains to > usage by loop for call push notifications. > > Currently we are proceeding with the assumption that the only viable way > to alert the client of a call is via simplepush, which loop currently uses > with our FxOS simplepush cluster. I've talked with Adam Roach about some > ways to alter the loop client's use slightly so that we can stand up a > separate set of 'simplepush' clusters specifically for use by loop. These > separate clusters will use a lighter-weight version of simplepush (in > development now) that doesn't do longer-term state maintenance which is > unnecessary for loop's use. > > Simplepush, the version currently running, has been benchmarked to about 1 > million clients using a cluster of 5 machines. This cluster required > memcached as it stored some state that loop's version (loopPush) doesn't > require. > > Adam indicated that it seems reasonable that 20% of Firefox users might > click the Loop icon, which would be around 100 million connections. > Confirmation or refinement of that number would be helpful. For this load > we would deploy multiple loopPush clusters and provide a URL that the loop > client would query before initially connecting to determine which clusters > have capacity to handle more connections. > > Remaining work to be done: > - simplepush codebase > - refactoring already under way for not using memcached > - additional clean-ups, performance optimizations > - cluster setup automation > > - loads (or some other testing tool) > - needs dynamic test client handling (we will need over 1,000 instances > to be spun up for larger scale test runs) > - ideally some automation integration (so that we can wire it into > jenkins for automatic runs) > > - integrated complete testing (Tokbox, Simplepush, Loop) > Tokbox has contacted us regarding how we plan on handling load testing of > the push system and ensuring their TURN/STUN servers can meet requirements > at scale. If we're looking at 100 million people possibly using this, how > many are behind firewalls that will require complete TURN server proxying? > If we have some ideas about what capacity Tokbox can/should handle, they'll > want to sync up with us on that. > > We will also need to do some failure-case testing of the loop client to > determine behavior under various failures in the SimplePush service. > > Regarding timelines, we believe it is possible to have a basic service up > around Firefox 33 hitting the public but as we haven't gotten to a point > where we've done any load-testing its hard to determine what kind of > capacity can be handled. I'm also not sure what timeline Loop or Tokbox is > operating with when it comes to having production deployments ready that > can handle 100 million users. > > Does anyone have roadmaps and timelines for these components? > > For those interested, we have the Push Meeting on Weds at 11am PST in the > Vidyo Services channel. > > Cheers, > Ben > _______________________________________________ > dev-media mailing list > [email protected] > https://lists.mozilla.org/listinfo/dev-media > -- Ian Bicking | Cloud Services Engineering Manager _______________________________________________ dev-media mailing list [email protected] https://lists.mozilla.org/listinfo/dev-media
