How the NFL Scheduled 272 Football Games Using 4,000 Virtual AWS Servers (amazon.com) 34
Long-time Slashdot reader theodp writes: AWS offered A Look Inside the Making of an NFL Football Schedule in conjunction with Thursday's release of the 2023 NFL Schedule Powered by AWS. AWS notes that producing the schedule required the use of 4,000+ AWS EC2 Spot Instances. An AWS promotional video claims they "saved the NFL an estimated $2 million each season" by leveraging AWS Spot Instances for a discount of up to 90% off compared to AWS On-Demand pricing..
"In just three months," AWS explains, "National Football League (NFL) schedule makers methodically build an exciting 18 week 272-game schedule spanning 576 possible game windows." Up until 10 years ago, AWS notes in an accompanying infographic, the NFL used a white-boarding process to manually craft its schedule.
Not to diminish the NFL's and AWS's 2023 scheduling achievement, but the 2013 documentary The Schedule Makers told the remarkable tale of the husband-and-wife duo of Henry and Holly Stephenson, who for almost a quarter of a century in the pre-Cloud era managed the scheduling for 30 Major League Baseball (MLB) teams who each played 162 regular season games a year. According to the May 1985 Atari Compendium (pg. 38), the Stephensons were using a self-written program running on a 64K IMS-8000 to help schedule games for the MLB (2,106 games over a 6-month season), NBA, and NASL/MISL (defunct soccer leagues). So perhaps the NFL's claim that "There's no way the NFL could deliver the quality of schedule that we put out every year for our fans and television partners without the contributions of our friends at AWS" should be taken with a grain of salt.
"In just three months," AWS explains, "National Football League (NFL) schedule makers methodically build an exciting 18 week 272-game schedule spanning 576 possible game windows." Up until 10 years ago, AWS notes in an accompanying infographic, the NFL used a white-boarding process to manually craft its schedule.
Not to diminish the NFL's and AWS's 2023 scheduling achievement, but the 2013 documentary The Schedule Makers told the remarkable tale of the husband-and-wife duo of Henry and Holly Stephenson, who for almost a quarter of a century in the pre-Cloud era managed the scheduling for 30 Major League Baseball (MLB) teams who each played 162 regular season games a year. According to the May 1985 Atari Compendium (pg. 38), the Stephensons were using a self-written program running on a 64K IMS-8000 to help schedule games for the MLB (2,106 games over a 6-month season), NBA, and NASL/MISL (defunct soccer leagues). So perhaps the NFL's claim that "There's no way the NFL could deliver the quality of schedule that we put out every year for our fans and television partners without the contributions of our friends at AWS" should be taken with a grain of salt.
Re: (Score:1)
Re: (Score:2)
Does the article say they use python? I didn't notice that. It was pretty thin on technical details.
Re: (Score:2)
Does the article say they use python? I didn't notice that. It was pretty thin on technical details.
If they used Python I would think they would need 3 times the number of compute instances to run that code to solve this problem. Python is SLOW.
Algorithms? (Score:3)
Just wondering if anybody has leads on the algorithms they use. The description sounds like they explore the search tree in parallel, but doesn't give any hints on algorithms. The Operations Research community often uses things like Constraint Propagation (CP), or Mixed-Integer Linear Programming (MILP) to solve scheduling problems, and there's a ton of literature on using these techniques.
A.
Re:Algorithms? (Score:5, Insightful)
The description sounds like they explore the search tree in parallel, but doesn't give any hints on algorithms.
Well, of course not. This story is ultimately intended as a marketing and sales piece for AWS. They don't make any money if they tell you how to do it.
Re: Algorithms? (Score:2, Insightful)
They 'saved' $2M how, exactly? In transportation cost? On the AWS charges? Where?
Why does this require 4,000 AWS 'spot' instances?
How is this measurably better than the husband and wife team that did it for (presumably) well under the cost savings AWS brought?
This is a non-sensical story - as presented - there's too many marketing over-simplifications to actually make sense of what they are describing...
I'd like to know the actual cost of this exercise, either before or after they 'saved' $2B by not using t
Re:Algorithms? (Score:4, Interesting)
They talk about "art and science". They also talk about a rules engine and 20,000 rules. They mention trees. They also say there isn't a perfect solution.
My guess is they build a loss function based on their 20,000 rules then entirely traverse the problem space multiple times with different weights for the rules. It would explain the huge number of instances they use.
It's probably cheaper to use brute force and ignorance with 4000 spot instances for a few days than get a team who can implement optimization with 20,000 constants. A very rough price calculation based on guesswork: 4000 X spot m7g.16xlarge = $8506.40 an hour.
This is total guesswork of course. There were few details in the article,
Re: Algorithms? (Score:2)
Re: Algorithms? (Score:1)
Yet, somehow, the husband and wife team did it for a quarter-century... This is a solvable problem (obviously), but the struggle is the conflicting goals amongst the various owners I suspect, they likely change their priorities frequently...
Re: (Score:2)
Probably running a script that's fed a seed for its random number generator. Then it runs all night spitting out randoms schedules and checking with a bunch of rules just written in that script.
Probably some perl script they wrote back on their "100" machines they used to run on. Just taking AWS practically free compute time they're giving away for the free advertising so they can just run it more.
I don't believe their is any super advanced algorithm involved.
Re: (Score:2)
Re:Algorithms? RTFA (Score:2)
RTFA
RTFARTFARTFA.
Heh. It says fart fart.
nothing was scheduled before AWS (Score:1)
Re: (Score:2, Funny)
Sports leagues didn't exist before AWS the same way our immune systems didn't exist before COVID.
WE NEED THESE MEGA CORPORATIONS TO LIVE. What don't you understand! Your ignorance is literally killing people.
Waaaahhhh
Re: (Score:2, Insightful)
Anyone who ever died had an immune system.
Re:nothing was scheduled before AWS (Score:4, Interesting)
They make it sound like it was impossible to make schedules before.
The NFL has, over time, made scheduling more difficult due to their various rules about local team/market carriage, coverage, blackouts, etc. with each content provider(s) contract having various assurances of games of likely interest (and viewers).
Re: nothing was scheduled before AWS (Score:3, Insightful)
So it's complex because they choose to make it complicated, got it.
Re: (Score:2)
Re: (Score:2)
Scope Creep [wikipedia.org]
Re: (Score:2)
Scope Creep [wikipedia.org]
Eeww!! Sounds like a colonoscopy treatment.
Slashvertisement (Score:2)
They make it sound like it was impossible to make schedules before
That's because quite simply, this story is pure slashvertisement.
The Titans asked people on Broadway for help ... (Score:2)
https://twitter.com/i/status/1... [twitter.com]
In recent years, NFL teams have been getting more creative with how they announce their upcoming schedules, and the winner this year goes to the Tennessee Titans, who recently went viral with their unforgettable release of their 2023 matchups.
In a clip posted Thursday, the Tennessee team tweeted simply, “We asked people on Broadway to help us with our 2023 schedule release.” It included a laughing face emoji as well.
The video showed just how many people who visit
Re: (Score:2)
Yeah, it's amazing the entire country doesn't have absolute knowledge of team logos. It's as if millions of people's lives don't revolve around the sport and the teams therein.
Re: (Score:1)
Why would you expect anyone other than NFL fans to recognize NFL team logos?
4,000 AWS instances for THAT?? (Score:1)
Is this scheduling NP-Hard? (Score:3)
It would make sense for a big bunch of computers to run all the possible schedules if this is NP-Hard, right?
Now I'd be interested in seeing how they defined all the likely constraints and preferences to feed all those "computers".
Re: (Score:2)
Re: (Score:2)
Likely if you want the PERFECT solution.
One interesting thing is, just because it's NP-Hard, doesn't mean "good enough" solutions aren't easier to find. You can approach the ideal solution and in many cases even know how close you are to the final solution using much easier algori
Re: (Score:2)
"Good enough" is probably the best you can get because the assumptions behind many of the constraints are uncertain and fuzzy. They are probably treating the fuzzy assumptions as non-fuzzy in their model. If there were a perfect schedule it would just be the same every year and that would be pretty boring. "Good enough" also leaves room for exciting surprises to happen that you don't anticipate ahead of time, but I'm pretty sure that kind of thinking is anathema to the NFL.
That really seems like something... (Score:2)
...a fairly short Prolog program could find out in minutes. Of course an exhaustive search would be way to long, but there are likely so many constraints that the number of possible arrangements probably falls to something you can just try out.
Hello? (Score:1)
Excel anyone?