Hacker News Reader

Comment by wiseowise | Hacker News Reader

Awaiting “something, something premature optimization” person to show up in the comments.

People always seem to think "premature optimization" means "premature optimization for speed". But there are lots of otherwise-good things that you can prematurely optimize for:

    - modularity
    - readability
    - abstraction
    - maintainability
    - scalability
    - reusability
    - testability
    - cost

We've all seen codebases that are over-abstracted in the name of code reuse, or to adhere to this or that software design philosophy, and are nightmares as a result. "You Ain't Gonna Need It" is a restatement of the premature-optimization adage, just for a different metric.

And the "premature optimization" excuse is used almost every time someone argues that over-abstraction can cause performance issues.

Or other forms of laziness.

"Premature optimisation" refers to speed or size. Using the phrase to talk about those other things is a stretch. (Except maybe cost.)

You can design for modularity, readability etc., but optimising for them doesn't sound right. Something to do with them not having well-defined, agreed-upon definitions of where the optimum is.

I work with a Principle of Maximum Power guy. He’s insufferable. Too smart to browbeat into behaving, and too old to chalk it up to youthful exuberance or lack of experience. He’s made big chunks of the project comfortable for him and miserable for everyone else, so of course he’ll never leave, and people who know they deserve better do.

I think it's fair to talk about optimisation (premature or otherwise) for anything fitting Blum's axioms. The classic ones are number of instructions executed, wall-clock time, memory usage, network requests, user interactions, calls to some slow function (e.g. the comparison function, for sorting algorithms), etc. but I think it's also applicable to fuzzier concepts.

For example, modularity could be modelled as the time/edit-distance/etc. it would take to swap-out N out of C components, where C >> N. If we're defining things centrally and passing them around via dependency-injection then it's O(N) (each component is defined once, so change N of them); if we're hard-coding references everywhere then it's O(N*C) since we'll have to check for references in all of the components, and each one needs to be checked for the N possible swaps.

On the other hand, dependency injection is less optimal than hard-coding for a metric like API size: for N components, the number of references grows as O(N^2) (since new components need to reference existing ones, and existing ones will need to reference them); so APIs using DI will have O(N^2) arguments.

Fair enough. You can optimise for lots of things.

But when people warn against "premature optimisation", specifically, they are advising you to give higher priority to things like modularity, readability, abstraction and maintainability, and not sacrifice them all for a singular goal, typically speed.

People deal with the cognitive dissonance by calling these optimizations other things, like “architecture”. But at the end of the day, if you’re changing your architecture for any reason other than correctness, then you’re optimizing. And if you’re having a hard time doing it, then you probably waited too long.

“The Last Responsible Moment” is subtle. You have to include user benefit (they may not care yet) versus snowballing consequences of delay, versus having more and better information later on.

One of the things that drew me to CI/CD was the notion that you’re not going to get better at handling your problems by avoiding them. You can’t learn to pick your optimization battles if you never fight them. You won’t learn something I’ve been saying for almost 25 years now and I still haven’t heard come out of anyone else’s mouth: there are code changes that improve readability and performance. Making those changes as refactors is hardly ever premature.

Lol, yeah I get this quoted sometimes merely for suggesting the right data structure to use.

"Use a set instead of a list, it's faster". "that's premature optimization!"

Usually this criticism doesn't even hold water if you put performance aside, since a set will often have nicer ergonomics if you find yourself reaching for it. I see far too much code that is like items.remove((item) => item == value) and if you're using the proper datastructure you don't have to handroll your implementations each time, because they are provided by default.

Premature pessimization.

The real premature optimization is using List instead of Set for guaranteed small sizes, since List is going to outperform Set in those cases.

That depends on how list and set are implemented.

Except in your example I’ve had numerous cases where bugs have been introduced because the developer’s primary thinking to pick a set was “fast” and “no duplicates”, only to later realize that order of items is important. And looking back at their choice, dealing with 5-20 items, using a list and checking contains() not only had no performance impact but better reflects the intent of “no duplicates”. Instead the set is left in place and the dev fixing it usually copies the results in a list to be sorted the same order prior to being in the set...

I've never seen this example. There are set implementations which maintain order, but in my experience it's fairly rare to both need "no duplicates" and "items inserted maintain order".

It's fairly easy to remove the set and replace it with something else if this does come up.

Maybe I’m misunderstanding, but isn’t it pretty common to e.g. have a list of non-duplicate items on screen which have a specific order they’re supposed to be displayed in which can be changed as a result of user actions (like adding new items or editing existing ones)?

Maybe? Perhaps that's just a part of my daily dev that is lacking. I generally work on the backend and not the frontend. I could envision how keeping stuff in the same order on a UX would be important.

Doesn't seem like it'd be the general case though. And it seems like it'd be fairly obvious when something like this applies.

Hash tables contain a list internally and sets are hash tables with just the keys in them so they are already lists. With a level of indirection, it's easy to make them insertion ordered just like normal lists. Just make the hash table a map of keys to list indexes. That way the table can be rehashed freely while the internal list remains ordered.

"In established engineering disciplines a 12% improvement, easily obtained, is never considered marginal; and I believe the same viewpoint should prevail in software engineering." - Also Knuth

A 12% overall improvement might be worthwhile. But you have to look at the bigger picture: if part A of your project takes 99.9% of the time, and part B takes 0.1% of the time, even a 100% improvement in part B (ie making it take 0 time) might not be worthwhile.

a.k.a. [Amdahl's Law](https://en.wikipedia.org/wiki/Amdahl%27s_law)

I agree. But it is critical to ensure that what one is improving is something worthwhile. 12% improvement in compute costs for a project with lots of compute spend = massive win. 12% improvement in compute costs for a hobby project running on a single VM = probably negligible. 12% improvement in function runtime some arbitrary place in the stack = possibly net negative, taking into account developer opportunity costs. It is almost always useful to talk in terms of outcomes, ie benefits to customer/users/team/company.

The original quote is not by Knuth, though he made it well known, it is by Tony Hoare.

Any improvement is significant!! I spent a week implementing an optimization for my language (pypy's storage strategies) and ended up significantly slowing it down instead. It's not guaranteed.

I have massive respect for anyone who can pull off steady improvements, no matter how small.

You’d think they would catch on. But then a foolish consistency is the hobgoblin of little minds.

Reply on news.ycombinator.com