Lessons from Building and Scaling Reddit's Ad Serving Platform with Go

Post by GFischer » Fri, Sep 14, 2018, 9:04 PM UTC

diminoten wrote: post #96

kokey wrote: post #11

Earlier quoted context omitted.

I remember when Digg was bigger than Reddit and then after a certain Digg site redesign they lost a significant number of their users to Reddit and never really recovered.

If memory serves, the primary driver of reduced traffic to Digg 4.0 was the fundamental shift in how their submission mechanism worked, not in a change to their layout. I recall some kind of promotion mechanism where superusers gained some amount of control over their "front page" equivalent. Also, unlike Digg, Reddit hasn't been bleeding members to another site for some time now, so there's nowhere for Reddit users…

Network effect is one for sure. I'm not a fan of the redesign (and submitted my thoughts several times), but the content is there, and it's a reasonable platform overall.

Post by katzgrau » Sat, Sep 15, 2018, 1:52 PM UTC

wink wrote: post #129

katzgrau wrote: post #99

Earlier quoted context omitted.

> if they control everything and can drop a standard 90ms down to 10ms 10ms, after request processing, ad selection logic, and faster-than-light packet travel? I think that's far too optimistic of an improvement. As the post said, 30ms was the goal. My adserver delivers in 20-40ms. But since ads are routinely loaded asynchronously, is speed concern truly valid? The adserver can usually make a selection and deliver be…

30ms really sounded a bit much for me, if you take network-locality as a precondition (i.e. adserver on same physical sub-1ms-network than webserver). But yeah, it should be loaded async. Been a while since I worked with ad stuff, but I thought 10ms-20ms sounded more likely.

If it was straight ad delivery I'd agree that being closer to 10ms is achievable. But in the process of pulling in ad data, user profile data, running it through a selection algorithm that can be pretty complex for even a simple adserver, calculating and plugging in macros, gzipping, and doing that thousands of times per second per server, the latency tends to creep upward.

More/faster hardware can obviously help, but since everything's async from the browser's perspective you tend to run into higher costs and diminishing returns by pushing for the lowest latency possible.

Re: Lessons from Building and Scaling Reddit's Ad Serving Platform with Go

Re: Lessons from Building and Scaling Reddit's Ad Serving Platform with Go