Hi all! My name is Mark, I’m the engineering manager on Sharethrough’s team that builds out features with a data driven feedback loop. In this blog post, we’ll spotlight a feature that makes it so exciting to work at Sharethrough because it involves machine learning, finding practical applications, and working on projects that align with our values, like sustainability.
Some Context
At Sharethrough, one of the products we manage is an ad exchange. For those uninitiated into the wild and wacky world of adtech, this can be viewed as a web service that facilitates communication between two parties:
- Web pages, apps, or streaming services that basically say, “hey, we’ve got a person who’s looking at our stuff, and we’d like to show them an ad!” (we call these entities “supply” or “publishers”)
- Advertising platforms that programmatically respond with, “nice, we’ve got an ad we’d like to pay you to run, and here’s how much we’ll pay to do it!” (we call these entities “demand” or “DSPs”)
Sharethrough then figures out which demand partner will result in the best profit for the supply partner, takes our cut, and sends down the ad (plus our secret sauce, enhancements). Everybody wins.
There’s just one catch.
The Problem: Missed Opportunities
A large proportion of requests from publishers don’t get responses from DSPs. This might be due to lack of interest, for example, a mall in Singapore probably isn't serving ads to people living in the American Midwest (not a real world example… that I know of). Or an advertiser might have already used up their monthly budget. In any case, a lot of requests to Sharethrough are simply wasted.
As of December 2022, Sharethrough was receiving 125 billion requests a day (that’s about 1.5 million requests per second), which is really exciting and fun from an engineering perspective. However, with great request count comes great AWS bill responsibility. What’s more, consuming those computational resources has a very real impact on the environment, especially when you consider that sending a request to a demand side partner often sets off a chain of requests from them to their partners.
More often than not, it’s just servers playing telephone with other servers, chewing up hardware and guzzling electricity, without any real value added to anyone.
Or at least, it would be, if it weren’t for a Sharethrough feature called Smart Throttling.
The Solution: Optimized Requests
With Smart Throttling, we segment out data about the traffic we receive each day and use it to predict if a request is likely to result in a bid. We then only send the request to a particular demand partner if those predictions suggest they will have interest. It’s a massively successful project, eliminating about 75% of our normal bandwidth costs and drastically cutting the number of server instances we need to run to handle our request load while having minimal effects to our partners.
Smart Throttling is obviously exciting to the finance team here at Sharethrough, but it’s also an exciting realm of learning and growth on the engineering side. Not only are there interesting scaling issues to be solved (daily request count tripled last year), but we’re also in the process of transitioning from a statistical model (which predicted whether to throttle traffic based on simple heuristics) to a more fully featured machine learning model that’s trained daily. We’re constantly learning and iterating on our process and framework and that’s a lot of fun.
Feeling Good
At the end of the day, Smart Throttling is the type of project I love working on at Sharethrough. It scratches the common engineering itch to make things more efficient. It gives us daily opportunities to learn new technologies. It lets us brag to our friends that we work on ML.
Most importantly, it shows that it’s possible to align financial priorities with doing what’s right. Smart Throttling saves Sharethrough and its partners a ton of money. But it also helps us cut out the unnecessary resource use that often underpins our favorite apps and damages the planet we live on. More green for our partners, and more green for Planet Earth: that’s about as good as it gets.
Contact us to start spending sustainably or to learn more about Smart Throttling and other features of the Sharethrough Exchange
Hi all! My name is Mark, I’m the engineering manager on Sharethrough’s team that builds out features with a data driven feedback loop. In this blog post, we’ll spotlight a feature that makes it so exciting to work at Sharethrough because it involves machine learning, finding practical applications, and working on projects that align with our values, like sustainability.
Some Context
At Sharethrough, one of the products we manage is an ad exchange. For those uninitiated into the wild and wacky world of adtech, this can be viewed as a web service that facilitates communication between two parties:
- Web pages, apps, or streaming services that basically say, “hey, we’ve got a person who’s looking at our stuff, and we’d like to show them an ad!” (we call these entities “supply” or “publishers”)
- Advertising platforms that programmatically respond with, “nice, we’ve got an ad we’d like to pay you to run, and here’s how much we’ll pay to do it!” (we call these entities “demand” or “DSPs”)
Sharethrough then figures out which demand partner will result in the best profit for the supply partner, takes our cut, and sends down the ad (plus our secret sauce, enhancements). Everybody wins.
There’s just one catch.
The Problem: Missed Opportunities
A large proportion of requests from publishers don’t get responses from DSPs. This might be due to lack of interest, for example, a mall in Singapore probably isn't serving ads to people living in the American Midwest (not a real world example… that I know of). Or an advertiser might have already used up their monthly budget. In any case, a lot of requests to Sharethrough are simply wasted.
As of December 2022, Sharethrough was receiving 125 billion requests a day (that’s about 1.5 million requests per second), which is really exciting and fun from an engineering perspective. However, with great request count comes great AWS bill responsibility. What’s more, consuming those computational resources has a very real impact on the environment, especially when you consider that sending a request to a demand side partner often sets off a chain of requests from them to their partners.
More often than not, it’s just servers playing telephone with other servers, chewing up hardware and guzzling electricity, without any real value added to anyone.
Or at least, it would be, if it weren’t for a Sharethrough feature called Smart Throttling.
The Solution: Optimized Requests
With Smart Throttling, we segment out data about the traffic we receive each day and use it to predict if a request is likely to result in a bid. We then only send the request to a particular demand partner if those predictions suggest they will have interest. It’s a massively successful project, eliminating about 75% of our normal bandwidth costs and drastically cutting the number of server instances we need to run to handle our request load while having minimal effects to our partners.
Smart Throttling is obviously exciting to the finance team here at Sharethrough, but it’s also an exciting realm of learning and growth on the engineering side. Not only are there interesting scaling issues to be solved (daily request count tripled last year), but we’re also in the process of transitioning from a statistical model (which predicted whether to throttle traffic based on simple heuristics) to a more fully featured machine learning model that’s trained daily. We’re constantly learning and iterating on our process and framework and that’s a lot of fun.
Feeling Good
At the end of the day, Smart Throttling is the type of project I love working on at Sharethrough. It scratches the common engineering itch to make things more efficient. It gives us daily opportunities to learn new technologies. It lets us brag to our friends that we work on ML.
Most importantly, it shows that it’s possible to align financial priorities with doing what’s right. Smart Throttling saves Sharethrough and its partners a ton of money. But it also helps us cut out the unnecessary resource use that often underpins our favorite apps and damages the planet we live on. More green for our partners, and more green for Planet Earth: that’s about as good as it gets.
Contact us to start spending sustainably or to learn more about Smart Throttling and other features of the Sharethrough Exchange
Behind Headlines: 180 Seconds in Ad Tech is a short 3-minute podcast exploring the news in the digital advertising industry. Ad tech is a fast-growing industry with many updates happening daily. As it can be hard for most to keep up with the latest news, the Sharethrough team wanted to create an audio series compiling notable mentions each week.
Hi all! My name is Mark, I’m the engineering manager on Sharethrough’s team that builds out features with a data driven feedback loop. In this blog post, we’ll spotlight a feature that makes it so exciting to work at Sharethrough because it involves machine learning, finding practical applications, and working on projects that align with our values, like sustainability.
Some Context
At Sharethrough, one of the products we manage is an ad exchange. For those uninitiated into the wild and wacky world of adtech, this can be viewed as a web service that facilitates communication between two parties:
- Web pages, apps, or streaming services that basically say, “hey, we’ve got a person who’s looking at our stuff, and we’d like to show them an ad!” (we call these entities “supply” or “publishers”)
- Advertising platforms that programmatically respond with, “nice, we’ve got an ad we’d like to pay you to run, and here’s how much we’ll pay to do it!” (we call these entities “demand” or “DSPs”)
Sharethrough then figures out which demand partner will result in the best profit for the supply partner, takes our cut, and sends down the ad (plus our secret sauce, enhancements). Everybody wins.
There’s just one catch.
The Problem: Missed Opportunities
A large proportion of requests from publishers don’t get responses from DSPs. This might be due to lack of interest, for example, a mall in Singapore probably isn't serving ads to people living in the American Midwest (not a real world example… that I know of). Or an advertiser might have already used up their monthly budget. In any case, a lot of requests to Sharethrough are simply wasted.
As of December 2022, Sharethrough was receiving 125 billion requests a day (that’s about 1.5 million requests per second), which is really exciting and fun from an engineering perspective. However, with great request count comes great AWS bill responsibility. What’s more, consuming those computational resources has a very real impact on the environment, especially when you consider that sending a request to a demand side partner often sets off a chain of requests from them to their partners.
More often than not, it’s just servers playing telephone with other servers, chewing up hardware and guzzling electricity, without any real value added to anyone.
Or at least, it would be, if it weren’t for a Sharethrough feature called Smart Throttling.
The Solution: Optimized Requests
With Smart Throttling, we segment out data about the traffic we receive each day and use it to predict if a request is likely to result in a bid. We then only send the request to a particular demand partner if those predictions suggest they will have interest. It’s a massively successful project, eliminating about 75% of our normal bandwidth costs and drastically cutting the number of server instances we need to run to handle our request load while having minimal effects to our partners.
Smart Throttling is obviously exciting to the finance team here at Sharethrough, but it’s also an exciting realm of learning and growth on the engineering side. Not only are there interesting scaling issues to be solved (daily request count tripled last year), but we’re also in the process of transitioning from a statistical model (which predicted whether to throttle traffic based on simple heuristics) to a more fully featured machine learning model that’s trained daily. We’re constantly learning and iterating on our process and framework and that’s a lot of fun.
Feeling Good
At the end of the day, Smart Throttling is the type of project I love working on at Sharethrough. It scratches the common engineering itch to make things more efficient. It gives us daily opportunities to learn new technologies. It lets us brag to our friends that we work on ML.
Most importantly, it shows that it’s possible to align financial priorities with doing what’s right. Smart Throttling saves Sharethrough and its partners a ton of money. But it also helps us cut out the unnecessary resource use that often underpins our favorite apps and damages the planet we live on. More green for our partners, and more green for Planet Earth: that’s about as good as it gets.
Contact us to start spending sustainably or to learn more about Smart Throttling and other features of the Sharethrough Exchange
Founded in 2015, Calibrate is a yearly conference for new engineering managers hosted by seasoned engineering managers. The experience level of the speakers ranges from newcomers all the way through senior engineering leaders with over twenty years of experience in the field. Each speaker is greatly concerned about the craft of engineering management. Organized and hosted by Sharethrough, it was conducted yearly in September, from 2015-2019 in San Francisco, California.
Stay Up-to-Date—
Subscribe to our newsletter and receive cutting-edge digital advertising insights, including our weekly Behind Headlines episodes, delivered right to your inbox.
.png)
