In practice though, for most enterprise web services, a lot of real world performance comes down to how efficiently you are calling external services (including the database). Just converting a loop of queries into bulk ones can help loads (and then tweaking the query to make good use of indexes, doing upserts, removing unneeded data, etc.)
I'm hopeful that improvements in LLMs mean we can ditch ORMs (under the guise that they are quicker to write queries and the inbetween mapping code with) and instead make good use of SQL to harness the powers that modern databases provide.
Also, before jsonb existed, you'd often run into big blobs of properties you don't care to split up into tables. Now it takes some discipline to avoid shoving things into jsonb that shouldn't be.
In my demo app, the CPU hotspots were entirely in application code, not I/O wait. And across a fleet, even "smaller" gains in CPU and heap compound into real cost and throughput differences. They're different problems, but your point is valid. Goal here is to get more folks thinking about other aspects of performance especially when the software is running at scale.
Apart from that my experience over the last 20 years was that a lot of performance is lost because of memory allocation (in GCed languages like Java or JavaScript). Removing allocation in hot loops really goes a long way and leads to 10 or 100 fold runtime improvements.
Parts of the GC language crowd in particular have come to hold some false optimistic beliefs about how well a GC can handle allocations. Also, Java and C# can sneak in silly heap allocations in the wrong places (e.g. autoboxing). So there is a tendency for programs to overload the GC with avoidable work.
Maybe we can ditch active models like those we see in sqlalchemy, but the typed query builders that come with ORMs are going to become more important, not less. Leveraging the compiler to catch bad queries is a huge win.
My experience with something like the latest Claude Code models these days has been that they are pretty good at SQL. I think some combination of LLM review of SQL code with smoke tests would do the trick here.
This is usually the first thing I look for when someone is complaining about speed. Developers often miss it because they are developing against a database on their local machine which removes any of the network latency that exists in deployed environments.
I think Java (or other JVM languages) are then best positioned, because of jooq. Still the best SQL generation library I've used.
There's a balance with a DB. Doing 1 or 2 row queries 1000 times is obviously inefficient, but making a 1M row query can have it's own set of problems all the same (even if you need that 1M).
It'll depend on the hardware, but you really want to make sure that anything you do with a DB allows for other instances of your application a chance to also interact with the DB. Nothing worse than finding out the 2 row insert is being blocked by a million row read for 20 seconds.
There's also a question of when you should and shouldn't join data. It's not always a black and white "just let the DB handle it". Sometimes the better route to go down is to make 2 queries rather than joining, particularly if it's something where the main table pulls in 1000 rows with only 10 unique rows pulled from the subtable. Of course, this all depends on how wide these things are as well.
But 100% agree, ORMs are the worst way to handle all these things. They very rarely do the right thing out of the box and to make them fast you ultimately end up needing to comprehend the SQL they are emitting in the first place and potentially you end up writing custom SQL anyways.
They store up conserved programming time and then spend it all at once when you hit the edge case.
If you never hit the case, it's great. As soon as you do, it's all returned with interest :)
Oracle is a prime example of this. Stored procedures are the place to put all business logic according to Oracle documentation.
This caused backslash from escaping developers who then declared business logic should never be inside the database. To avoid vendor lock-in.
There's no ideal solution, just tradeoffs.
I mean, that already happens. It's quite rare to see someone migrate from one database to another. Even if they stuck to pure SQL for everything, it's still a pretty daunting process as Postgres SQL and MSSQL won't be the same thing.
I'm not discounting the level of effort involved, but I think the reason you don't see this often is because it is rare that simply changing DBMS systems is beneficial in and of itself.
And even if it was frictionless (ie: if we had discovered ORM Samarkanda), the real choices are so limited that even if you did it regularly, you would soon run out of DBMSs to try.
Absolutely not.
That which is asserted without evidence can be dismissed without evidence.
When using JDBC I found myself quickly in implementing a poor mans ORM.
I recently fixed a treesitter perf issue (for myself) in neovim by just dfsing down the parse tree instead of what most textobject plugins do, which is:
-> walk the entire tree for all subtrees that match this metadata
-> now you have a list of matching subtrees, iterate through said subtree nodes, and see which ones are "close" to your cursor.
But in neovim, when I type "daf", I usually just want to delete the function right under my cursor. So you can just implement the same algorithm by just... dfsing down the parse tree (which has line numbers embedded per nodes) and detecting the matches yourself.
In school, when I did competitive programming and TCS, these gains often came from super clever invariants that you would just sit there for hours, days, weeks, just mulling it over. Then suddenly realize how to do it more cleverly and the entire problem falls away (and a bunch of smart people praise you for being smart :D). This was not one of them - it was just, "go bypass the API and do it faster, but possibly less maintainably".
In industry, it's often trying to manage the tradeoff between readability, maintainability, etc. I'm very much happy to just use some dumb n^2 pattern for n <= 10 in some loop that I don't really care much about, rather than start pulling out some clever state manipulation that could lead to pretty "menial" issues such as:
- accidental mutable variables and duplicating / reusing them later in the code
- when I look back in a week, "What the hell am I doing here?"
- or just tricky logic in general
I only noticed the treesitter textobject issue because I genuinely started working with 1MB autogen C files at work. So... yeah...
I could go and bug the maintainers to expose a "query over text range* API (they only have query, and node text range separately, I believe. At least of the minimal research I have done; I haven't kept up to date with it). But now that ties into considerations far beyond myself - does this expose state in a way that isn't intuitive? Are we adding composable primitives or just ad hoc adding features into the library to make it faster because of the tighter coupling? etc. etc.
I used to think of all of that as just kind of "bs accidentals" and "why shouldn't we just be able to write the best algorithms possible". As a maintainer of some systems now... nah, the architectural design is sometimes more fun!
I may not have these super clever flashes of insight anymore but I feel like my horizons have broadened (though part of it is because GPT Pro started 1 shotting my favorite competitive programming problems circa late 2025 D: )
After all, even if one has some slow and beastly, unoptimized Spring Boot container that chews through RAM, its not that expenseive (in the grand scheme of things) to just replicate more instances of it.
Or even the local filesystem :)
CPU calls are cheap, memory is pretty cheap, disk is bad, spinning disk is very bad, network is 'good luck'.
You can O(pretty bad) most of the time as long as you stay within the right category of those.
The String.format() problem is most immediately a bad compiler and bad implementation, IMO. It's not difficult to special-case literal strings as the first argument, do parsing at compile time, and pass in a structured representation. The method could also do runtime caching. Even a very small LRU cache would fix a lot of common cases. At the very least they should let you make a formatter from a specific format string and reuse it, like you can with regexes, to explicitly opt into better performance.
But ultimately the string templates proposal should come back and fix this at the language level. Better syntax and guaranteed compile-time construction of the template. The language should help the developer do the fast thing.
String concatenation is a little trickier. In a JIT'ed language you have a lot of options for making a hierarchy of string implementations that optimize different usage patterns, and still be fast - and what you really want for concatenation is a RopeString, like JS VMs have, that simply references the other strings. The issue is that you don't want virtual calls for hot-path string method calls.
Java chose a single final class so all calls are direct. But they should have been able to have a very small sealed class hierarchy where most methods are final and directly callable, and the virtual methods for accessing storage are devirtualized in optimized methods that only ever see one or two classes through a call site.
To me, that's a small complexity cost to make common string patterns fast, instead of requiring StringBuilder.
They tried, its opponents dilluted it to the point of uselessness and now will forever use this failed attempt as a wedge.
I'm sorry, I don't believe Java will get sensible String templates in our life time.
I love how Zig, D and Rust do exactly what you say: parse the format string at compile time, making it super efficient at runtime (no parsing, no regex, just the optimal code to get the string you need).
I say this but I write most of my code in Java/Kotlin :D . I just wish I could write more low-level languages for super efficient code, but for what I do, Java is more than enough.
Also C++, which works the same way.
I was listening to someone say they write fast code in Java by avoiding allocations with a PoolAllocator that would "cache" small objects with poolAllocator.alloc(), poolAllocator.release(). So just manual memory management with extra steps. At that point why not use a better language for the task?
Well, the whole thing was standard Java OOP, except they also had a bunch of functional programming stuff on top of that. I can relate to that -- I think they were university students when they started, and I definitely had an OOP and FP phase. But then they just... kept it, 10+ years later.
So while it's true that you can write C in any language... those kind of folks don't tend to use Java in the first place ;)
--
(Except Notch? Well, his code looks like C, not sure if it's actually fast! I really enjoyed his 4 kilobyte java games back in the day, I think he published the source for each one too.)
EDIT: Found it!
https://web.archive.org/web/20120317121029/http://www.mojang...
Edit 2: This one has a download, still works!
https://web.archive.org/web/20120301015921/http://www.mojang...
I've spent a fair few years developing lowish (10-20us wire to wire) latency trading systems and the majority of the code does not need to go fast. It's just wasted effort, a debugging headache, and technical debt. So the natural trade off is a bit of pain to make the hot path fast through spans, unsafe code, pre-allocated object pools, etc and in return you get to use a safe and easy programming language everywhere else.
In C# low latency dev is not even that painful, as there are a lot of tools available specifically for this purpose by the runtime.
Even JavaScript is much better for this, much, much better.
And the manually manager int array acts more like system memory, it's not continuous, so you could have point i 0 and 2 and the data would be: [1, 2, 3, x, x,x, 3, 2, 1] (3D points).
So I am not describing a struct of arrays.
A project might also grow into these requirements. I can easily imagine that something wasn't problematic for a long time but suddenly emerged as an issue over time. At that point you wouldn't want to migrate the whole codebase to a better language anymore.
This is a wrong question to ask in this context. The right question to ask is when actually exceptional flow becomes a performance bottleneck. Because, obviously, in a desktop or even in a server app validating single user input even 99% of wrong inputs won’t cause any trouble. It may become a problem with bulk processing, but then, and I have to repeat myself here, it is no longer a number parsing problem, it’s a problem of not understanding what your input is.
This is actually the perfect situation: you are allowed to do it carefully and manually for 1% of code on the hot path, but you don't have to worry about it for the 99% of the code that's not.
Doing it to avoid memory pressure generally means you simply have a bad algorithm that needs to be tweaked. It's very rarely the right solution.
Orders by hour could be made faster. The issue with it is it's using a map when an array works both faster and just fine.
On top of that, the map boxes the "hour" which is undesirable.
This is how I'd write it
long[] ordersByHour = new long[24];
var deafultTimezone = ZoneId.systemDefault();
for (Order order : orders) {
int hour = order.timestamp().atZone(deafultTimezone).getHour();
ordersByHour[hour]++;
}
It's also not less readable, just less familiar as Java devs don't tend to use arrays that much.
Practically speaking, that would be pretty unusual. I don't think I've ever seen that sort of construct in my day to day coding (which could realistically have more than 1B elements).
It does seem like java missed their chance, it's a shame.
I wish Java had a proper compiler.
https://foojay.io/today/how-is-leyden-improving-java-perform...
I long ago concluded that Java was not a client or systems programming language because of the implementation priorities of the JVM maintainers. Note that I say priorities--they are extremely bright and capable engineers that focus on different use cases, and there isn't much money to be made from a client ecosystem.
AOT options like GraalVM Native Image can help cold starts a lot, but then half your favorite frameworks breaks and you trade one set of hoops for another. Pick which pain you want.
There are JITs that use dynamic profile guided optimization which can adjust the emitted binary at runtime to adapt to the real world workload. You do not need to have a profile ahead of time like with ordinary PGO. Java doesn't have this yet (afaik), but .NET does and it's a huge deal for things like large scale web applications.
https://devblogs.microsoft.com/dotnet/bing-on-dotnet-8-the-i...
The folks on embedded get to play with PTC and Aicas.
Android, even if not proper Java, has dex2oat.
if it is valuable, i'd be surprised you can't freeze/resume the state and use it for instantaneous workload optimized startup.
Too many folks have this mindset there is only one JVM, when that has never been the case since the 2000's, after Java for various reasons started poping everywhere.
in practice, for web applications exposing some sort of `WarmupTask` abstraction in your service chassis that devs can implement will get you quite far. just delay serving traffic on new deployments until all tasks complete. that way users will never hit a cold node
There are options to turn on which cause the JVM to save off and reload compiled classes. It pretty massively improves performance.
You can get even faster if you do that plus doing a jlink jvm. But that's more of a pain. The AOT cache is a lot simpler to do.
Because in my experience as of 2026, Java programs are consistently among the most painful or unpleasant to interact with.
> But people just want to compare it to building a cli tool in go or rust.
This seems like the key. HN is definitely biased towards simpler, smaller tools. (And that's not a bad thing!). The most compelling JVM stories I hear are all from much larger scale enterprise settings.
Kafka being a good example. It's very good at what it does, but painful to manage and usually not worth the pain for anyone who's not in a mega enterprise.
A second bug is that Character.isDigit() returns true for non-ASCII Unicode digits as well, while Integer.parseInt() only supports ASCII digits.
Another bug is that the code will fail on the input string "-".
Lastly, using value.isBlank() is a pessimization over value.isEmpty() (or just checking value.length(), which is read anyway in the next line), given that the loop would break on the first blank character. It makes the function not be constant-time, along with the first point above that the length of the digit sequence isn’t being limited.
[0]
public int parseOrDefault(String value, int defaultValue) {
if (value == null || value.isBlank()) return defaultValue;
for (int i = 0; i < value.length(); i++) {
char c = value.charAt(i);
if (i == 0 && c == '-') continue;
if (!Character.isDigit(c)) return defaultValue;
}
return Integer.parseInt(value);
}Rest of advice is great: things compilers can't really catch but a good code reviewer should point out.
And aside from algorithms, it usually comes down to avoiding memory allocations.
I have my go-to zero-alloc grpc and parquet and json and time libs etc and they make everything fast.
It’s mostly how idiomatic Java uses objects for everything that makes it slow overall.
But eventually after making a JVM app that keeps data in something like data frames etc and feels a long way from J2EE beans you can finally bump up against the limits that only c/c++/rust/etc can get you past.
I’ve heard about HFT people using Java for workloads where micro optimization is needed.
To be frank, I just never understood it. From what I’ve seen heard/you have to write the code in such a way that makes it look clumsy and incompatible with pretty much any third party dependencies out there.
And at that point, why are you even using Java? Surely you could use C, C++, or any variety of popular or unpopular languages that would be more fitting and ergonomic (sorry but as a language Java just feels inferior to C# even). The biggest swelling point of Java is the ecosystem, and you can’t even really use that.
But on Java specifically: every Java object still has a 24-byte overhead. How doesn't that thrash your cache?
The advice on avoiding allocations in Java also results in terrible code. For example, in math libraries, you'll often see void Add(Vector3 a, Vector3 b, Vector3 our) as opposed to the more natural Vector3 Add(Vector3 a, Vector3 b). There you go, function composition goes out the window and the resulting code is garbage to read and write. Not even C is that bad; the compiler will optimize the temporaries away. So you end up with Java that is worse than a low-level imperative language.
And, as far as I know, the best GC for Java still incurs no less than 1ms pauses? I think the stock ones are as bad as 10ms. How anyone does low-latency anything in Java then boggles my mind.
public int parseOrDefault(String value, int defaultValue) {
if (value == null || value.isBlank()) return defaultValue;
for (int i = 0; i < value.length(); i++) {
char c = value.charAt(i);
if (i == 0 && c == '-') continue;
if (!Character.isDigit(c)) return defaultValue;
}
return Integer.parseInt(value);
}
It doesn't excuse the "use exceptions for control flow" anti-pattern, but it is a quick patch.
> StringBuilder works off a single mutable character buffer. One allocation.
It's one allocation to instantiate the builder and _any_ number of allocations after that (noting that it's optimized to reduce allocations, so it's not allocating on every append() unless they're huge).
This one is so prevalent that JVM has an optimization where it gives up on filling stack for exception, if it was thrown over and over in exact same place.
The rest were all very familiar. Well, apart from the new stuff. I think most of my code was running in java 6...
Who’s there?
long pause
Java
Java is only fast-ish even on its best day. The more typical performance is much worse because the culture around the language usually doesn't consider performance or efficiency to be a priority. Historically it was even a bit hostile to it.
Man that code looks awful. Really reminds me of why I drifted away from Java over time. Not just the algorithm, of course; the repetitiveness, the hoops that you have to jump through in order to do pretty "stream processing"... and then it's not even an FP algorithm in the end, either way!
Honestly the only time I can imagine the "process the whole [collection] per iteration" thing coming up is where either you really do need to compare (or at least really are intentionally comparing) each element to each other element, or else this exact problem of building a histogram. And for the latter I honestly haven't seen people fully fall into this trap very often. More commonly people will try to iterate over the possible buckets (here, hour values), sometimes with a first pass to figure out what those might be. That's still extra work, but at least it's O(kn) instead of O(n^2).
You can do this sort of thing in an elegant, "functional" looking way if you sort the data first and then group it by the same key. That first pass is O(n lg n) if you use a classical sort; making a histogram like this in the first place is basically equivalent to radix sort, but it's nice to not have to write that yourself. I just want to show off what it can look like e.g. in Python:
def local_hour(order):
return datetime.datetime.fromtimestamp(order.timestamp).hour
groups = itertools.groupby(sorted(orders, key=local_hour), key=local_hour)
orders_by_hour = {hour: len(list(orders)) for (hour, orders) in groups}
(Another edit: originally I missed that the code was only trying to count the number of orders in each hour, rather than collecting them. I fixed the code above, but the discussion makes less sense for the simplified problem. In Python we can do this with `collections.Counter`, but it wouldn't be unreasonable to tally things up in a pre-allocated `counts_by_hour = [0] * 24` either.)
----
Edit:
> String.format() came in last in every category. It has to... StringBuilder was consistently the fastest. The fix: [code not using StringBuilder]... Use String.format() for the numeric formatting where you need it, and let the compiler optimize the rest. Or just use a StringBuilder if you need full control.
Yeah, this is confused in a way that I find fairly typical of LLM output. The attitude towards `String.format` is just plain inconsistent. And there's no acknowledgment of how multiple `+`s in a line get optimized behind the scenes. And the "fix" still uses `String.format` to format the floating-point value, and there's no investigation of what that does to performance or whether it can be avoided.
Same for Java, I have yet to in my entire career see enterprise Java be performant and not memory intensive.
At the end of the day, if you care about performance at the app layer, you will use a language better suited to that.
Factories! Factories everywhere!
https://gwern.net/doc/cs/2005-09-30-smith-whyihateframeworks...
Most of this stuff is just central knowledge of the language that you pick up over time. Certainly, AI can also pick this stuff up instantly, but will it always pick the most efficient path when generating code for you?
Probably not, until we get benchmarks into the hot path of our test suite. That is something someone should work on.
any other resources like this?
That said, the article does have the "LLM stank" on it, which is always offputting, but the content itself seems solid.
Gradle does suck and maven is ok but a bit ugly.
I use LLMs to maintain them now. I keep the build files simple. It was an inconvenience before, but a trifle now.
* Most mature Java project has moved to Kotlin.
* The standard build system uses gradle, which is either groovy or kotlin, which gets compiled to java which then compiles java.
* Log4shell, amongst other vulnerabilities.
* Super slow to adopt features like async execution
* Standard repo usage is terrible.
There is no point in using Java anymore. I don't agree that Rust is a replacement, but between Python, Node, and C/C++ extensions to those, you can do everything you need.
Maven on the other hand, is just plain boring tech that works. There's plenty of documentation on how to use it properly for many different environments/scenarios, it's declarative while enabling plug-ins for bespoke customisations, it has cruft from its legacy but it's quite settled and it just works.
Could Maven be more modern if it was invented now? Yeah, sure, many other package managers were developed since its inception with newer/more polished concepts but it's dependable, well documented, and it just plain works.
Like I said, it's not hypermodern with batteries included, and streamlined for what became more common workflows after it was created but it doesn't need workarounds, it's not complicated to define a plugin to be called in one of the steps of the lifecycle, and it's provided as part of its plugin architecture.
I can understand spending many hours fighting Gradle, even I with plenty of experience with Gradle (begrudgingly, I don't like it at all) still end up fighting its idiocies but Maven... It's like any other tool, you need to learn the basics but after that you will only fight it if you are verging away from the well-documented usage (which are plenty, it's been battle-tested for decades).
It gets a reaction, though, so great for social media.
Programming in Rust is a constant negotiation with the compiler. That isn't necessarily good or bad but I have far more control in Zig, and flexibility in Java.
It's cool when your tooling warns you about potential bugs or mistakes in implementation, but it's still your responsibility to write the correct code. If you pick up a hammer and hit your finger instead of the nail, then in most cases (though not always) it’s your own fault.
I am talking about this bug. It looks like it is still unfixed, in the sense, there is a PR fixing it, but it wasn't merged. LOL.
Regardless of whether this specific bug would be caught by Rust compiler, Bun in general is notorious for crashing, just look at how many open issues there are, how many crashes.
Not saying that you cannot make a correct program in Zig, but I prefer having checks that Rust compiler does, to not having them.