[zbobet2012]

I think "industrial computing just isn't that hard relative to math and the sciences" is only true if you are working on a limited set of applications.

In my job as a software engineer, I've had to: Invent new algorithms and prove their correctness. Formalize datatypes as CRDTs. Design (and model check) a specialized distributed consensus algorithm. Read and implement _many_ algorithms from academic papers, including motion vector estimation. And build complex statistical models.

All of the above require either formal use or the actual practice of mathematics. Sometimes very advanced mathematics including multivariate calculus, statistics, graph theory, number theory, category theory, and so on.

I guess what I'm saying is, software engineering can be as easily as challenging as math and science. Because at its "peak", it is math and science. Not all mathematicians and scientist work on hard problems. Not all software engineers do either. But don't be to quick to judge the field based on a limited sample size.

[protomikron]

Can you elaborate on what you worked on? Was it part of academics or industry? I guess it was embedded systems, as these domain seems pretty open to formalization (as they are pretty low-level), as they often demand high reliability.

[zbobet2012]

Most of my formal work was for distributed systems. Some of it was for video transcode and packaging work. All of it was for industry.

For example, I implemented a distributed adaptive bitrate video packager whose core synchronization algorithm I needed to both develop, and then for my own comfort, formally prove. I did a proof of correctness via exhaustion and a proof that it achieved consensus in a minimal number of steps by induction.

This is pretty typical for any "distributed" algorithm because intuiting correctness for concurrent and distributed stuff is ... really hard. That's why people prove garbage collection algorithms.

When designing large distributed systems a thorough understanding of statistics is required to understand workloads. How many transactions per second should this micro service support? Better pull out that Poisson distribution. What cache size do I need? Better grab a zipf distribution, some sample data, and R. Want to understand the interplay of several factors on workload? I hope you are comfortable with multivariable calculus.

I don't face problems like these daily, but when I do I'm fucking glad for every inch of math I know. Which to be honest is still probably not enough.

[protomikron]

Thanks for this thoughtful answer.