Abstract:

Most products we release into the world are destined for abject failure. A very small number will become runaway successes. What separates one from the other, as N.N. Taleb never tires of reminding us, is not sage-like foresight or iron-clad work ethic but luck.

Taleb¹s advice to casino dwellers everywhere holds important lessons for those of us who plan and build the virtual infrastructure on which future products will thrive or, more frequently, sink into oblivion.

To maximize our chances of a lucky break, we have to settle into a rhythm of placing small bets, none of them so costly failure would send in the bailiffs. We have to minimize the cost of each investment made without sacrificing our ability to scale up massively in the never unexpected but highly unlikely event of a smash hit.

To borrow one of Taleb¹s key terms, the goal is to thrive on optionality¹ redundant resources until traffic ramps up. We are left with pairs of autoscaling groups (one empty, the other nearly so) and geographically distributed copies of our infrastructure that are instantiated occasionally for testing and immediately torn down.

Scaling out using many inexpensive commodity components that often fail¹ [2] has allowed the web giants to grow at an astonishing rate, but it also holds out the promise of an architecture that can be collapsed back to a single node running on bargain bin hardware.

This talk is an invitation to start asking questions we are often too embarrassed to ask: what is the least performant model you have in stock? Can you order one that is cheaper still? What is the lowest achievable throughput for that database-as-a-service your sales rep has praised to the skies? Once that is out of the way, you will find that there is plenty to like in an occasionally sluggish service used by not very many people. Crucially, though, we can always scale up in the unlikely event that we unwrap the golden ticket.

Sadly the talk has no golden ticket to offer, but it does come with a budget-friendly, open source CloudFormation template that puts this approach into practice.

1 N.N. Taleb, Antifragile: Things That Gain from Disorder, esp. Book IV.

[2] gfs-sosp2003.pdf http://static.googleusercontent.com/media/research.google.com/en//archive/gfs-sosp2003.pdf

Speaker: Gerald Schmidt

blog comments powered by Disqus