The entire point of LLM-assisted development is to audit the code generated by AI and to further instruct it to either improve it or instruct it to fix the shortcomings - kind of being a senior dev doing a code review on your colleague's merge request. In fact, as developers, we usually read code more than we write it, which is also why you should prefer simple and verbose code over clever code in large codebases. This seems like it would be instead aimed at pure vibecoded slop.
> Do you debug JVM bytecode? V8's internals?
People do debug assembly generated by compilers to look for miscompilations, missed optimization opportunities, and comparison between different approaches.
This describes where we are at now. I don't think it's the entire point. I think the point is to get it writing code at a level and quantity it just becomes better and more efficient to let it do its thing and handle problems discovered at runtime.
Intellisense + Intellicode + Roslynator (extension) combined were really the height of productivity in Visual Studio. Now they've driven a steam-roller over all of that, forced CoPilot down our throats.
I LIKE CoPilot's "chat" interface, and agents are fine too (although Claude in VS Code is tons better), but CoPilot auto-complete is negative value and shouldn't be used.
Huh I'm the opposite. I find the copilot chat slow and low value compared to ChatGPT. But I use the tab autocomplete a lot.
Otoh I disabled all the intellisense stuff so I don't have the issues described in TFA: tab is always copilot autocomplete for whatever it shows in grey.
I hate the time unpredictability of it. Intellij also has AI completion suggestions, and sometimes they're really useful. But sometimes when I expect them, they don't come. Or they briefly flash and then disappear.
What would be nice is if you could ask for a suggestion with one key, so it's there when I want it, and not when I don't. That would put me in control. Instead I feel subjected to these completely random whims of the AI.
FPS is a poor metric anyway, things like this should be measured in frame time instead - but either are meaningless numbers without knowing the hardware it runs on.
Sure, I usually measure performance of methods like these in terms of FLOP/s; getting 50-65% of theoretical peak FLOP/s for any given CPU or GPU hardware is close to ideal.
Yes but the primary issue was that 4chan was using over a decade old version of the library that contained a vulnerability first disclosed in 2012: https://nvd.nist.gov/vuln/detail/CVE-2012-4405
No benchmarks. No FLOPs. No comparison to commodity hardware. I hate the cloud servers. "9 is faster than 8 which is faster than 7 which is faster than 6, ..., which is faster than 1, which has unknown performance".
How: You've ran the test on a bunch of hosts and create a spec from ranges.
Why: you might be concerned with network connectivity (you don't get to choose which data center you launch in and it might not be exactly equal), noisy neighbors on shared hosts, etc. if you're measuring for networking, you probably are spinning ups separate accounts/using a bank of accounts and something in every az until you find what you're looking for.
I’ve had terrible luck benchmarking EC2. Measurements are too noisy to be repeatable. The same instance of the wrong type can swing by double digit percentages when tested twice an hour apart.
That didn't stop people from throwing a fit over master-slave terminology in software (having nothing to do with slavery), going so far as to rename long-standing development branch names, as well as put significant effort into removing such terms from the code itself and any documentation.
How does that work? You transfer the money in BTC, you exchange that into real money, you open a new bank account, you deposit that money, the bank's anti-money-laundering detection sees a large deposit to a newly open account and triggers an alert, the bank locks you out of the account and asks you for a proof of income/tax payment, you have no explanation of where that money came from legally, they freeze your account and report you to their local revenue services, you're SOL.
It's way harder to support Linux than Windows from a developer's perspective. Proprietary vs. open source drivers, approach to driver updates (rolling release vs. stable distros), 5 trillion incompatible glibc versions, X11 vs. Wayland etc, janky sound systems with varied support across Linux distributions (Pulse, Alsa, PipeWire), no ABI compatibility guarantee etc.
> A picking texture is a very simple idea. As the name says, it’s used to handle picking in the game, when you click somewhere on the screen (e.g. to select an unit), I use this texture to know what you clicked on. Instead of colors, every object instance writes their EntityID to this texture. Then, when you click the mouse, you check what id is in the pixel under the mouse position.
Unrelated, but why? Querying a point in a basic quad tree takes microseconds, is there any benefit to overengineering a solved problem this way? What do you gain from this?
I assume for the picking system you're rendering each entity/block as a different color (internally) and getting the pixel color under the mouse cursor?
"Color" is pretty much just "integer" to the GPU. It doesn't care if the 32-bit value a shader is writing to its output buffer is representing RGBA or a memory pointer.
It aligns with what appears on the screen accurately and without needing any extra work to make sure there's a representation in a tree that's pixel-accurate. It's also pretty low overhead with the way modern GPU rendering works.
What if you have a collision system where collision filters can exclude collisions based on some condition in such a way that their bounding boxes can overlap? For instance an arrow that pierces through a target to fly through it and onto another target? How do you accurately store the Entity ID information for multiple entities with a limited number of bits per pixel?
Entities that can't be picked, don't write to the texture, entities that can be picked, write to the texture their id. Whatever is closer to the camera will be the id that stays there (same as a color pixel, but instead of the object color you can think object id).
So you are limited at one ID per pixel, but for me that works.
Right, it's the same z-buffer problem of deciding what pixel color is visible, with a non-blending buffer update mode.
To be totally coherent, you have to draw the entity ID in the same order you would draw the visible color, in cases where entities could "tie" at the same depth.
A point in screen space is a line in world space after inverse camera projection, so this way you get the line-to-closest-geometry test in O(1), after the overhead of needing to render the lookup texture first.
> Hopefully, this post helps illustrate the unreasonable effectiveness of SQLite as well as the challenges you can run in with Amdahl's law and network databases like postgres.
No, it does not. This article first says that normally you would run an application and the database on separate servers and then starts measuring the performance of a locally embedded database. If you have to keep the initial requirement for your software, then SQLite is completely out of equation. If you can change the requirement, then you can achieve similar performance by tuning the local PGSQL instance -- and then it also becomes a valuation of features and not just raw throughput. I'm not saying SQLite is not an option either, but this article seems confusing in that it compares two different problems/solutions.
Right - but SQLite handily beats the case where postgres is on the same box as well. And it's completely reasonable to test technology in the configuration in which it would actually run.
As an industry, we seem to have settled on patterns that actually are quite inefficient. There's no problem that requires the solution of doing things inefficiently just because someone said databases should run on a different host.
I sometimes dream of a local-first world in which all software works with local DB and only writes to the cloud as an afterthought, maybe as a backup or a way to pick up work on another machine. It just boggles my mind that more software nowadays relies on an always on internet connection for no good reason other then the design itself.
I think people's reaction to cloud vendors is to go local first. But, there's a middle ground VPS, rented server, even self hosting.
My problem with local first is it's fine for solo apps with the occasional sync. But doesn't work for medium to large datasets and the stuff I work in is generally real-time and collaborative. To me multiplayer is one of the strengths of the web.
> If you have to keep the initial requirement for your software, then SQLite is completely out of equation.
No it isn't? You can run a thin sqlite wrapping process on another server just fine. Ultimately all any DB service is, PostgreSQL included, is a request handler and a storage handler. SQLite is just a storage handler, but you can easily put it behind a request handler too.
Putting access to sqlite behind a serial request queue used to be the standard way of implementing multi-threaded writes. That's only spitting distance away from also putting it behind TCP.
You could do that, but you'd run into exactly the same bottleneck the author describes with a remote Postgres instance. The workload exposes high contention on hot rows. If transactions are kept open for several milliseconds due to this being a remote network call between client and DB server, throughput will be equally limited also when using SQLite.
Exactly. People forget that “SQLite can’t do X” often really means “SQLite doesn’t ship with X built in.” If you wrap it with a lightweight request handler or a queue, you essentially recreate the same pattern every other DB uses. The fact that PostgreSQL bundles its own coordinator doesn’t make SQLite fundamentally incapable. It just means you choose whether you want that layer integrated or external.
Far from it, as now your not just dealing with network but also with raft consensus... So each write is not just a network trip, its also 2x acknowledging. And your reads go over the leader, what can mean if somebody accessed node 1 app but node 2 is the leader, well, ...
Its slower on reads and writes, then just replications that PostgreSQL does. And i do not mean async but even sync PostgreSQL will be faster.
The reason dqlite exists is because canonical needed something to synchronize their virtualization cluster (lxd), and they needed a db with raft consensus, that is a lib (as not a full blown server install like postgres). Performance was not the focus and its usage is totally different then most people needs here.
Nit: dqlite is a library, it is not a network-exposed database like rqlite is. Sure, it requires connecting to other nodes over the network, but local access is via in-process. In contrast one connects with rqlite over the network - HTTP specifically.
Paradoxically, raw throughput matters a lot more if you are going to scale on a single box. SQLite is 10x PG on a single box in this example. Considering databases tend to be the bottle neck that can take you an order of magnitude further. PG on the same server will also be slower the more complex the transaction as unix sockets are still going to be considerably slower than a function call.
The other thing to point out is in this article is that the PG network example CANNOT scale horizontally due to the power law. You can throw a super cluster at the problem and still fundamentally do around 1000 TPS.
Also important is just how fast cheap hardware has gotten which means vertical scaling is extremely effective. People could get a lot farther with sqlite in wal mode on a single box with an nvme drive than they imagine. Feels like our intuition has not caught up with the material reality of current hardware.
And now that there are solid streaming backup systems, the only real issue is redundancy not scaling.
> If you have to keep the initial requirement for your software, then SQLite is completely out of equation.
It'd be a very short article if so, don't you think? Full article would be something like: "Normally you'd have a remote connection to the database, and since we're supposed to test SQLite's performance, and SQLite is embedded, it doesn't compare. Fin"
The table of data at the end of the article has 7 lines, only one has data for both DBs. What was the point of setting up the comparison if there is no comparison made?
> Do you debug JVM bytecode? V8's internals?
People do debug assembly generated by compilers to look for miscompilations, missed optimization opportunities, and comparison between different approaches.
reply