Comment by GuB-42
5 hours ago
> Premature optimization is the root of all evil.
There are few principle of software engineering that I hate more than this one, though SOLID is close.
It is important to understand that it is from a 1974 paper, computing was very different back then, and so was the idea of optimization. Back then, optimizing meant writing assembly code and counting cycles. It is still done today in very specific applications, but today, performance is mostly about architectural choices, and it has to be given consideration right from the start. In 1974, these architectural choices weren't choices, the hardware didn't let you do it differently.
Focusing on the "critical 3%" (which imply profiling) is still good advice, but it will mostly help you fix "performance bugs", like an accidentally quadratic algorithms, stuff that is done in loop but doesn't need to be, etc... But once you have dealt with this problem, that's when you notice that you spend 90% of the time in abstractions and it is too late to change it now, so you add caching, parallelism, etc... making your code more complicated and still slower than if you thought about performance at the start.
Today, late optimization is just as bad as premature optimization, if not more so.
The most misunderstood statement in all of programming by a wide margin.
I really encourage people to read the Donald Knuth essay that features this sentiment. Pro tip: You can skip to the very end of the article to get to this sentiment without losing context.
Here ya go: https://dl.acm.org/doi/10.1145/356635.356640
Basically, don't spend unnecessary effort increasing performance in an unmeasured way before its necessary, except for those 10% of situations where you know in advance that crucial performance is absolutely necessary. That is the sentiment. I have seen people take this to some bizarre alternate insanity of their own creation as a law to never measure anything, typically because the given developer cannot measure things.
> I have seen people take this to some bizarre alternate insanity of their own creation as a law to never measure anything, typically because the given developer cannot measure things.
Similar to the "code should be self documenting - ergo: We don't write any comments, ever"
It is to me incredible, how many „developers“, even “10 years senior developers” have no idea how to use a dubugger and or profiler. I’ve even met some that asked “what is a profiler?” I hope I’m not insulting anybody, but to me is like going to an “experienced mechanic” and they don’t know what a screwdriver is.
13 replies →
> "code should be self documenting
It should be to the greatest extent possible. Strive to write literate code before writing a comment. Comments should be how and why, not what.
> - ergo: We don't write any comments, ever"
Indeed this does not logically follow. Writing fluent, idiomatic code with real names for symbols and obvious control flow beats writing brain teasers riddled with comments that are necessary because of the difficulty in parsing a 15-line statement with triply-nested closures and single-letter variable names. There's a wide middle ground where comments are leveraged, not made out of necessity.
3 replies →
Laziness in moral clothing.
> Similar to the "code should be self documenting - ergo: We don't write any comments, ever"
My counterpoint: Code can be self-documenting, reality isn't. You can have a perfectly clear method that does something nobody will ever understand unless you have plenty of documentation about why that specific thing needs to be done, and why it can't be simpler. Like having special-casing for DST in Arizona, which no other state seems to need:
https://en.wikipedia.org/wiki/Time_in_the_United_States
3 replies →
(this is the correct answer, parent needs to understand this better)
This is crucial detail that almost everyone misses when they are skimming the topic on surface. The implication is that this statement/law is referenced more often to shut down the architecture designs/discussions
Even moreso . I like the Rob Pike restatement of this principle, it really makes it crystal clear:
"You can't tell where a program is going to spend its time. Bottlenecks occur in surprising places, so don't try to second guess and put in a speed hack until you've proven that's where the bottleneck is."
Moreso, in my personal experience, I've seen a few speed hacks cause incorrect behavior on more than one occasion.
In particular I've seen way too many people use this term as an excuse to write obviously poor performing code. That's not what Knuth said. He never said it's ok to write obviously bad code.
I'm still salty about that time a colleague suggested adding a 500 kb general purpose js library to a webapp that was already taking 12 seconds on initial load, in order to fix a tiny corner case, when we could have written our own micro utility in 20 lines. I had to spend so much time advocating to management for my choice to spend time writing that utility myself, because of that kind of garbage opinion that is way too acceptable in our industry today. The insufferable bastard kept saying I had to do measurements in order to make sure I wasn't prematurely optimizing. Guy adding 500 kb of js when you need 1 kb of it is obviously a horrible idea, especially when you're already way over the performance budget. Asshat. I'm still salty he got so much airtime for that shitty opinion of his and that I had to spend so much energy defending myself.
Reminds me of a codebase that was littered with SQL injection opportunities because doing it right would have been "premature optimization" since it was "just" a research spike and not customer facing. Guess what happened when it got promoted to a customer facing product?
4 replies →
That is the giat of the leftpad history, isn’t it?
> except for those 10% of situations where you know in advance that crucial performance is absolutely necessary
Yeah like, NOT indexing any fields in a database, that'll become a problem very quickly. ;)
The biggest issue I have with premature optimization is stuff that really doesn't matter.
For example, in Java I usually use ConcurrentHashMap, even in contexts that a regular HashMap might be ok. My reasoning for this is simple: I might want to use it in a multithreaded context eventually and the performance differences really aren't that much for most things; uncontested locks in Java are nearly free.
I've gotten pull requests rejected because regular HashMaps are "faster", and then the comments on the PR ends up with people bickering about when to use it.
In that case, does it actually matter? Even if HashMap is technically "faster", it's not much faster, and maybe instead we should focus on the thing that's likely to actually make a noticeable difference like the forty extra separate blocking calls to PostgreSQL or web requests?
So that's the premature optimization that I think is evil. I think it's perfectly fine at the algorithm level to optimize early.
With modern tools it should be pretty easy to build scalable solutions. I take premature optimization as going out of your way to optimize something that's already reasonable. Not that you should write really bad code as a starting point.
The problem is that that this term gets misused to say the opposite of what it was intended for.
It's particularly the kind of people who like to say "hur hur don't prematurely optimize" that don't bother writing decent software to begin with and use the term as an excuse to write poor performing code.
Instead of optimizing their code, these people end up making excuses so they can pessimize it instead.
To me that's the people who write desktop software in Electron. Hate that.
In my career Ive seen about 1000 instances of somebody trying to optimize something prematurely.
Usually those people also have a good old whinge about the premature optimization quote being wrong or misinterpreted and general attitudes to software efficiency.
Not once have I ever seen somebody try to derail a process of "ascertain speed is an issue that should be tackled" -> "profile" -> fix the hot path.
Yeah, I interpret "premature optimization" as taking a request that takes 500ms and focusing on saving a couple ms by refactoring logic to avoid a SQL JOIN or something.
Your users are not going to notice. Sure, it's faster but it's not focused on the problem.
Spent 6 months last year ripping out an abstraction layer that made every request 40ms slower. We profiled, found the hot path, couldn't fix it without a rewrite. The "optimize later" school never tells you later sometimes means never
I'd say it usually means "never".
I also find it a bit annoying is that most people just make shit up about stuff that is "faster". Instead of measuring and/or looking at the compiled bytecode/assembly, people just repeat tribal knowledge about stuff that is "faster" with no justification. I find that this is common amongst senior-level people at BigCos especially.
When I was working in .NET land, someone kept telling me that "switch statements are faster" than their equivalent "if" statements, so I wrote a very straightforward test comparing both, and used dotpeek to show that they compile to the exact same thing. The person still insisted that switch is "faster", I guess because he had a professor tell him this one time (probably with more appropriate context) and took whatever the professor said as gospel.
I've seen a lot of requests to obtimize code where we can measure the optimal versions saves a few nanoseconds. I just deleted some 'optimal code' that took a lot of mutexes and so was only faster when there is no contetion but in real world multi-writer situations the easy code wins. (Shared memory vs local socket for ipc)
2 replies →
"today, performance is mostly about architectural choices, and it has to be given consideration right from the start"
This doesn't make sense. Why is performance (via architectural choices) more important today than then?
You can build a snappy app today by using boring technology and following some sensible best practices. You have to work pretty hard to need PREMATURE OPTIMIZATION on a project -- note the premature there
The big thing that changed is that almost all software performance today is bandwidth-bound at the limit. Not computation-bound. This transition was first noticed in supercomputing around 25 years ago.
Optimization of bandwidth-bound code is almost purely architectural in nature. Most of our software best practices date from a time when everything was computation-bound such that architecture could be ignored with few bad effects.
I agree. But I have to say, when defining the architecture, there are things known that will be terrible bottlenecks later. They should be avoided. Just as the previous comment, about defining proper indices in a database. Optimization means making something that is already “good” and correct better. There is no excuse to make a half ass, bug ridden shitty software, under the excuse “optimization is for later” that is technical debt generation: 2 very different things.
> You can build a snappy app today by using boring technology and following some sensible best practices.
If you are building something with similar practical constraints for the Nth time this is definitely true.
You are inheriting “architecture” from your own memory and/or tools/dependencies that are already well fit to the problem area. The architectural performance/model problem already got a lot of thought.
Lots of problems are like that.
But if you are solving a problem where existing tools do a poor job, you better be thinking about performance with any new architecture.
In the 1970s computer systems spanned fewer orders of magnitude. Operations generally took somewhere between maybe 1 and 10^8 CPU cycles. Today, the range is closer to 10^-1 to 10^13.
> Why is performance (via architectural choices) more important today than then?
There were fewer available layers of abstraction.
Whether you wrote in ASM, C, or Pascal, there was a lot less variance than writing in Rust, JavaScript, Python.
I only use mature optimizations, so I'm good.
Thinking about the overall design, how its likely to be used, and what the performane and other requirements are before aggregating the frameworks of the day is mature optimization.
Then you build things in a reasonable way and see if you need to do more for performance. It's fun to do more, but most of the time, building things with a thought about performance gets you where you need to be.
The I don't need to think about performance at all camp, has a real hard time making things better later. For most things, cycle counting upfront isn't useful, but thinking about how data will be accessed and such can easily make a huge difference. Things like bulk load or one at a time load are enormous if you're loading lots of things, but if you'll never load lots of things, either works.
Thinking about concurrency, parallelism, and distributed systems stuff before you build is also pretty mature. It's hard to change some of that after you've started.
That first sentence.
I want it in a t-shirt. On billboards. Everywhere :)
What's the problem with SOLID? It's very very rare that I see a case where going against SOLID leads to better design.
SOLID tend to encourage premature abstraction, which is a root of evil that is more relevant today than optimization.
SOLID isn't bad, but like the idea of premature optimization, it can easily lead you into the wrong direction. You know how people make fun of enterprise code all the time, that's what you get when you take SOLID too far.
In practice, it tends to lead to a proliferation of interfaces, which is not only bad for performance but also result in code that is hard to follow. When you see a call through an interface, you don't know what code will be run unless you know how the object is initialized.
In a way, SOLID is premature optimization. You are optimizing abstractions before knowing how the code is used in practice. Lots of code will be written and never changed again, but a minority will see changes quite a bit. Concentrate there. Like you don't need to optimize things that aren't in hot code (usually, omg experience will tell you that all rules have exceptions, including the exceptions).
> Lots of code will be written and never changed again, but a minority will see changes quite a bit. Concentrate there
I think the most important principle above all is knowing when not to stick to them.
For example if I know a piece of code is just some "dead end" in the application that almost nothing depends on then there is little point optimizing it (in an architectural and performance sense). But if I'm writing a core part of an application that will have lots of ties to the rest, it totally does make sense keeping an eye on SOLID for example.
I think the real error is taking these at face value and not factoring in the rest of your problem domain. It's way too simple to think SOLID = good, else bad.
here's a nice critique of SOLID principles:
https://www.tedinski.com/2019/04/02/solid-critique.html
They start by indicating people don't understand, “A module should have only one reason to change.”. Reading more of that article, it's clear the author doesn't understand much about software engineering and sounds more like a researcher who just graduated from putting together 2+2.
4 replies →
This doesn't seem to be a critique of the principles so much as a critique of their phrasing.
Even his "critique" of Demeter is, essentially, that it focuses on an inconsequential aspect of dysfunction—method chaining—which I consider to be just one sme that leads to the larger principle which—and we, apparently, both agree on this—is interface design.
It only applies to the object oriented programming paradigm
Negative.
The only part of SOLID that is perhaps OO-only is Liskov Substitution.
L is still a good idea, but without object-inheritance, there's less chance of shooting yourself in the foot.
That's understating the problem. It mandates OOP.
If you follow SOLID, you'll write OOP only, with always present inheritance chains, factories for everything, and no clear relation between parameters and the procedures that use them.
This is only superficially true. Here's a fair discussion that could serve as a counterpoint: https://medium.com/@ignatovich.dm/applying-solid-principles-...
It causes excessive abstraction, and more verbose code.
L and I are both pretty reasonable.
But S and D can easily be taken to excess.
And O seems to suggest OO-style polymorphism instead of ADTs.
This is similar to my view. All these "laws" should alwaye be used as guidance not as actual laws. Same with O. I think its good advice to design software so adding features that are orthogonal to other features don't require modifying much code.
That's how I view it. You should design your application such that extension involves little modifying of existing code as long as it's not necessary from a behavior or architectural standpoint.
1 reply →
It's a very interesting topic. Even when designing a system, how to modularize, it's healthy to wait until the whole context is in sight. It's a bit of a black art, too early or too late you pay some price.
> making your code more complicated and still slower than if you thought about performance at the start.
Not if your optimization for performance is some Rube Goldberg assemblage of microservices and an laundry list of AWS services.
Exactly. Today "premature optimization" almost always means unnecessary infra or abstractions or some other complexity- not DS&A choices.
I still don't understand microservices for anything short of a NAG of four level architecture.
Sounds like we agree.
Bunch of stuff is done for us. Using postgres having indexes correct - is not premature optimization, just basic stuff to be covered.
Having double loop is quadratic though. Parallelism is super fun because it actually might make everything slower instead of faster.
Is this less relevant today where if you do go down the wrong architecture it is much cheaper to rewrite everything again in something that is better?
All aphorisms are awful.
I like it as a way to remind myself to not get caught up in the minutiae.
Completely agreed here [1].
And as I point out, what Knuth was talking about in terms of optimization was things like loop unrolling and function inlining. Not picking the right datastructure or algorithm for the problem.
I mean, FFS, his entire book was about exploring and picking the right datastructures and algorithms for problems.
[1] https://news.ycombinator.com/item?id=47849194
Strange, this is actually one of the most important things I learnt the hard way as an analyst having had the misfortune of being forced to do local-level data engineering after being tasked with doing extensive BI on heavy data volumes (I don't have engineering education) without cloud ETL. Had someone told me this simple statement before time it would have saved me a LOT of pain and effort.
If you done enough premature optimization, you will know it’s usually wasted, detrimental and bad trade
"Premature optimization is the root of all evil"
Decades in, this is the worst of all of them. Misused by laziness or malice, and nowhere near specific enough.
The graveyard of companies boxed in by past poor decisions is sprawling. And the people that made those early poor decisions bounce around field talking about their "successful track record" of globally poor and locally good architectural decisions that others have had to clean up.
It touches on a real problem, though, but it should be stricken form the record and replaced with a much better principle. "Design to the problem you have today and the problems you have in 6 months if you succeed. Don't design to the problems you'll have have next year if it means you won't succeed in 6 months" doesn't roll off the tongue.
On your last bit, I definitely agree... personally I've leaned more and more into KISS above all else... simple things that are easy to replace are easily replaced only when you need to. Similarly, I also tend to push for initial implementations of many/most things in a scripted language first, mostly for flexibility/simplicity to get a process "right" before worrying about a lot of other things.
One thing that came out of the no-sql/new-sql trends in the past decade and a half is that joins are the enemy of performance at scale. It really helps to know and compromise on db normalization in ways such as leaning on JSON/XML for non-critical column data as opposed to 1:1/children/joins a lot of the time. For that matter, pure performance and vertical scale have shifted a lot of options back from the brink of micro service death by a million paper cuts processes.
Found the overbuilder!
>> Today, late optimization is just as bad as premature optimization, if not more so.
You are right about the origin of and the circumstances surrounding the quote, but I disagree with the conclusion you've drawn.
I've seen engineers waste days, even weeks, reaching for microservices before product-market fit is even found, adding caching layers without measuring and validating bottlenecks, adding sharding pre-emptively, adding materialized views when regular tables suffice, paying for edge-rendering for a dashboard used almost entirely by users in a single state, standing up Kubernetes for an internal application used by just two departments, or building custom in-house rate limiters and job queues when Sidekiq or similar solutions would cover the next two years.
One company I consulted for designed and optimized for an order of magnitude more users than were in the total addressable market for their industry! Of that, they ultimately managed to hit only 3.5%.
All of this was driven by imagined scale rather than real measurements. And every one of those choices carried a long tail: cache invalidation bugs, distributed transactions, deployment orchestration, hydration mismatches, dependency array footguns, and a codebase that became permanently harder to change. Meanwhile the actual bottlenecks were things like N+1 queries or missing indexes that nobody looked at because attention went elsewhere.
Thank your for posting this. I disagreed with OP but couldn't _quite_ find the words to describe why. But your post covers what i was trying to say.
I was quite literally asked to implement an in-memory cache to avoid a "full table scan" caused by a join to a small DB table recently. Our architect saw "full table scans" in our database stats and assumed that must mean a performance problem. I feel like he thought he was making a data-driven profiling decision, but seemed to misunderstand that a full-table scan is faster for a small table than a lookup. That whole table is in RAM in the DB already.
So now we have a complex Redis PubSub cache invalidation strategy to save maybe a ms or two.
I would believe that we have performance problems in this chunk of code, and it's possible an in-memory cache may "fix" the issue, but if it does, then the root of the problem was more likely an N+1 query (that an in-memory cache bandaids over). But by focusing on this cache, suddenly we have a much more complex chunk of code that needs to be maintained than if we had just tracked down the N+1 query and fixed _that_
> All of this was driven by imagined scale rather than real measurements
Yes. When I was a young engineer, I was asked to design something for a scale we didn’t even get close to achieving. Eventual consistency this, event driven conflict resolution that… The service never even went live because by the time we designed it, everyone realized it was a waste of time.
I learned it makes no sense to waste time designing for zillions of users that might never come. It’s more important to have an architecture that can evolve as needs change rather than one that can see years into the future (that may never come).
The wheel is a premature optimization to someone who never figured out how to build one.
You ARE going to need it.
I would venture that this statement is not true for library authors. Performance is a large factor in competitive advantage, especially in domains like image analysis or processing large corpuses of text etc.
In these domains, algorithm selection, and fine tuning hot spots pays off significantly. You must hit minimum speeds to make your application viable.
I would add this: https://magarshak.com/blog/perfection-is-the-enemy-of-the-go...
Unfortunately people do keep repeating it to excuse the fact that they don't know how to optimize in the first place.
Anyone who has done optimization even a little knows that it isn't very difficult, but you do need to plan and architect for it so you don't have to restructure you whole program to get it to run well.
Mostly it's just rationalization, people don't know the skill so they pretend it's not worth doing and their users suffer for it.
If software and website were even reasonably optimized people could just use a computer as powerful as a rasberry pi 5 (except for high res video) for most of what they do day to day.
I definitely hate SOLID more.