Hacker News Reader

Comment by dirtydroog | Hacker News Reader

> We don't write programs for machines to understand. We write programs for humans to understand.

True, but that program ends up executing on a machine. You normally want that program to execute as efficiently as possible. 'Efficiency' can be measured several ways... time to complete, resources required, financial cost.

A case in point is our company switched from using Cassandra (JVM based) to ScyllaDB (a C++ clone of Cassandra). Query latencies were the same if not faster, and we needed fewer nodes to handle the same load. All because the language used has some mechanical sympathy for the CPU it's going to run on, and no GC.

I see Scala is somewhat the defacto language used for 'big data' stuff. I have no idea why. I reckon if some bean counter translated that choice into $$$ things would change.

View on HN

> True, but that program ends up executing on a machine. You normally want that program to execute as efficiently as possible. 'Efficiency' can be measured several ways... time to complete, resources required, financial cost.

No, you normally want that program to execute efficiently enough to do its required job. Abstractions are often necessarily costly, but that's ok because their value is greater than their cost. Otherwise, you'd just throw all the abstractions away and use assembler.

> Query latencies were the same if not faster, and we needed fewer nodes to handle the same load. All because the language used has some mechanical sympathy for the CPU it's going to run on.

It seems like there's a whole load of reasons why it might have been more efficient on ScyllaDB that are unrelated to what language it was written in. Maybe the software design fit your requirements better? Maybe it was just better implemented? Maybe it was easier to configure for your load? Etc. etc. Programs aren't magically "more efficient" because they're written in C.

> No, you normally want that program to execute efficiently enough to do its required job

but... the "required job" is to go as fast as possible - in two ways :

1/ You must go faster than your competitors. Else people will switch, your software will get bad reviews, etc.

2/ you must go fast enough for anything that people will throw at you. And, in many cases, people can throw arbitrarily large problems to your program, and telling them "sorry but it will take 1 hour to process your 2TB dataset" just won't cut it at all. And even when you manage to cut it, then people come at you like "now can this run on a raspberry pi?"

Take for instance image processing software, audio or multimedia software... they will never be fast enough, and there is always an use case where getting a 1% perf improvement will allow a project to make instead of break (especially when working with uncompromising artists).

> but... the "required job" is to go as fast as possible - in two ways :

This is sometimes true, but hardly ever. Most software is not image/audio processing software. Most software is not even interacted with directly by humans. Most software exists as a component as part of a larger system, the performance of which is typically determined by a small fraction of that system. In other words, most software is not on the critical path.

Even if you're on the critical path, the performance requirements could be of the form "Here's some data, I need an answer in an hour". If your current system spends 2 minutes computing and delivering the result, why spend time and money making it faster? Just because you're on the critical path for you, doesn't mean you're on the client's critical path.

This is the whole point of performance engineering - identifying which bits are important and then figuring out how to optimise those. Optimising everything to run as fast as possible is a waste of time and effort - worse, it often degrades the overall system by making previously easy-to-understand components more complex and more difficult to reason about. See the whole "Worse is better" thing.

If you're on the critical performance path, go nuts. Throw away the abstractions and take it down to the metal. If you're not? Why bother?

Perhaps the 'I need answer in a hour' is currently acceptable because that's how long people are used to waiting for the answer. If they knew that you could get that answer in 2 minute do you not think they'd ask for that instead?

Like I mentioned previously, Scala's use in Big Data is perplexing. I reckon Cloud providers are making an absolute mint from design / language inefficiencies.

> the performance requirements could be of the form "Here's some data, I need an answer in an hour"

I have never been in the case where the answer to the question "when" wasn't "as fast as physically possible".

Reply on news.ycombinator.com