Hacker News Reader

Comment by smaili | Hacker News Reader

One interesting thing I've noticed from this (not sure if this is just my experience or if others noticed as well) but none of the ads seem to be impacted by this. Certainly makes me wonder if those are on a more prioritized and entirely separate infra and SLA that is designed to be more resilient and highly available.

View on HN

It's a lot easier to serve ads than your custom photos. The ads are more generic, easier to cache globally, and show to a lot of people. They can also fall back to super generic ads if the database is unavailable to customize them.

But your photos don't have a fallback.

Fallback is the main reason. Some webservices even fallback to non-moneymaking ads for charities in the case of a technical fault, because shareholders react badly to an ads outage, but they don't notice an hour or two of charity ads...

Our local transit system installed video boards showing arrival times for the next trains. It was paid for and operated by an advertising company who in return were able to run ads on their video boards. The functionality for arrival times rarely worked but the displaying of ads never failed. It was obvious where the company's priorities were at.

Intersection right?

> Certainly makes me wonder if those are on a more prioritized and entirely separate infra and SLA that is designed to be more resilient and highly available.

There's absolutely no doubt that the ads are served differently. Think about it this way:

- Ads: Small number, each served to many people, funded with real cash, but you can't get the cash if you don't serve the ads.

- Photos: Large number, each served to few people, funded by money left over from ads, and if you don't serve them your (non-paying) customers get frustrated.

My guess is that ads is funded well enough that they can run at lower resource utilization, and much more effort is made to run photos at high resource utilization. That's how I'd run it.

Images:. Serve requested image.

Ads:. Conduct an auction in milliseconds between hundreds of parties between hundreds of millions of ad creatives with lots of very large in-memory machine learning systems, all to decide which ad to serve to maximize revenue.

Different problem entirely.

In systems like this, you tend to want to maximise some other outcome than revenue (like clicks or conversions). In general, both GOOG and FB make money out of DR advertisers who care about this, so you tend to make more money if you optimise for the thing they care about (revenue goes up as a second-order effect).

Reply on news.ycombinator.com