- Author here. Let me know if you have any questions about the post or about pgrust.
Let me take a shot at answering what I think will be the most common question: how can I trust pgrust? Our #1 priority right now is correctness. Over the past two weeks, I've done a mix of formal verification and differential fuzz testing. We've been able to prove over 1000 user facing functions have the exact same logic in both pgrust and postgres (see the proofs directory if you're curious). For cases where formal verification is not easy, we've taken the c implementation of a function and the rust implementation of a function and ran millions of inputs through each of them and confirmed they gave the same results every time.
We've only covered about 15% of the surface area so far, but in the process, we've discovered ~100 bugs in pgrust and ~20 bugs in Postgres itself. My favorite postgres bug we found is this one[0]. Postgres has a quadtree implementation. Due to floating point rounding, it was possible for a point to be neither above, nor below, nor even with the center point of the quadtree.
We've also entered engagements with Antithesis[1] to do Jepsen style fault testing and Aretta[2] to do more serious formal verification.
If you want to support the project, the easiest way is to give us a star on GitHub[3]
[0] https://www.postgresql.org/message-id/19597-39c532e61d78dff6...
- How do you know if you're making the right optimizations?
I struggle with this a lot with Marginalia's index. Where I identify a hot method in a prod profiler run, try to replicate it on a test machine where I can never get the same cache characteristics because everything in this space is like an onion of caching layers that you affect the real performance of the system. I may get it to run significantly faster, but that only sometimes makes the production profiler sample move its needle.
e.g. I've recently been experimenting with using a cursed hybrid model in Marginalia's index, where based on a mincore probe, I switch between mmap and io_uring for reading a cluster of pseudoadjacent data. There are real tanglible benefits both in the test machine and in prod with this, but the numbers do not agree at all about how the needle moves :P
- I would probably dig into the reasons for the differences in the benefit on the test machine and in prod
I had an issue like this for optimizing pgrust. I had an optimization that showed no impact on my test machine (c8g.4xl) and showed a 20% improvement when ran on my mac. It turns out the issue was the instruction cache on the c8g.4xl was being saturated on the test machine but not on my laptop, moving the bottleneck to a different place
If you can consistently reproduce the performance difference, you're already half way there
- I'm pretty sure the reason for the difference is that production machine exists in a state of mixed memory residency and low grade resource contention that is incredibly hard to replicate in a test scenario (as the moment you start making queries the pages warm up, and the test becomes unreliable).
The hard part about optimizing this type of code, IMO, is that there are so many cache layers, both in the CPU and the OS and sometimes in the storage medium. You can warm all of those caches up, but then you're testing a nonsense scenario that will basically never happen in a realistic scenario, where wall clock time is what matters, and not much the CPU is working or how many IOPS you're pushing.
- The floating point comparison bug is nightmare fuel. I could look at that for years and never spot the mistake.
- On the bright side it could probably run for years without hitting the mistake as well. But it is nice to get it out of there.
- Fuzzers are brilliant at this and produce all kinds of insane floating point inputs.
- If someone wanted to use this as a real-time WAL-tracking read-only mirror of a live production database, for analytics work, is it ready for that use case yet?
- You can try it. We're happy to help you with it, but expect there to be issues to work through. You would want to do it for something non-critical
- This is a great project. Thank you!
A question on 20s postgresql time - It does not look like you are accounting for reading data from disk? Wouldn't the aggregation query have to load data from disk first? Or is it somewhat guaranteed that the table is already in memory? The Rust version is clearly in memory (I am no rust expert, so that may not even be actually in memory, if its a generator).
- > A question on 20s postgresql time - It does not look like you are accounting for reading data from disk
I choose the data size so that it would fit in memory on the machine I was testing on. fwiw, there's still a ton of overhead Postgres has that the toy example does not. For example Postgres will serialize the numbers into tuples and need to deserialize them to execute the query. That's why it's not an apples-to-apples comparison
- Is `pgrcolumnar` the default storage layout for tables? It would be cool if the same storage engine outperforms vanilla Postgres under both OLTP and OLAP workloads.
AlloyDB from Google Cloud uses columnar storage like a secondary index, while the relations are still stored in TOAST.
- pgrcolumnar is not the default storage method. Right now, it's exposed as a table access method. There's lots of design space for how to do this so I want to avoid pre-committing to anything
- what is your vision of this project? Do you think pgrust will eventually be prod ready?
- I want to build the best database possible. While Postgres is great, there are a lot of core issues that have been around for over a decade. We're working hard to get pgrust production-ready, and it will definitely be production-ready in the near future. I wouldn't be putting hundreds of thousands of dollars into this project if I didn't think we could build a production-ready database.
- “Show me the prompt.”
- Cool project but .. reality is that people will generally not choose pgrust over Postgres, even 5-10 years from now. The problem is not that it may be technically superior and faster by then, it's that it's not built by the trusted Postgres team. There's a lot more to trust than development velocity or performance. It's also about the longevity and continuity of a critical piece of technology.
- There's also the fact that it's basically vibecoded.
- People will use it if it really delivers gain and can solve problems PG can't.
- It takes years to test it thoroughly in real environments. You don't want your data to be silently corrupted.
- Their ai agents will if we keep writing about it
Project managers and Human Resources rolling out overengineered projects will outnumber current software engineers 10 to 1
- That's a pretty gloomy view
- They will be software engineers or the people doing the software work
And different people will be doing product management and HR
all because today’s software engineers don’t want to babysit AI agents and choose antiquated libraries so that their resume said they'd been using a popular framework for a couple years
Nobody else is playing that outdated game, its just a rotation
- What are we going to do about this?
- this thread is talking about using a 300x faster library and theorizing a resistance to devs using it
just use the faster library? leverage compute resources more effectively and justify value to an organization better than the next person
otherwise, what needs to be done? I don’t see a problem with any of this aside from organizations risking less experienced people doing less efficient things in other parts of the stack
- So one trick you can do is make a ramfs/tmpfs and start Postgres on it.
You need a server with enough ram to fit it all.
But it kind of make the database fly.
- Question, does having it in pure rust, opens possibility of embedding pgrust directly into binary, making it an alternative to SQLite/turso?
- You have no idea how long I have been waiting for adaptive planning. One of my biggest annoyances with the Postgres core team has been their reluctance to implement any sort of adaptive planning despite it, at this point, being a well-established technique that has been implemented in multiple production databases. I hope this, at the very least, proves the viability of this model outside of academic/niche contexts.
- I would be interested about a more detailed architecture overview of the io scheduler (like this: https://www.scylladb.com/2021/04/06/scyllas-new-io-scheduler...) and the thread scheduler.
PostgreSQL has historically been bad at managing the noisy neighbor problem, but with thread pools, and io priorities, it can be solved.
Has this been tackled here ?
- I'll need to write up how the scheduler works at some point, but it's heavily based on these papers[0][1]. It solves two different problems. First, it lets us throttle resource-intensive queries. Second, it enables work stealing. If you have idle cores on your machine, we'll assign those cores to running queries to help speed them up. That means if you have an over-provisioned machine, we'll make use of the extra capacity to speed your queries up.
[0] https://15721.courses.cs.cmu.edu/spring2016/papers/p743-leis...
[1] https://db.in.tum.de/~kohn/papers/query-scheduling-sigmod21....
- I am very disappointed to see the direction: It is moving from a "interesting attempt to recreate system software" to "building flashy but useless demo"
Everyone who knows a bit about databases knows the difference between execution models and what kind of optimization it brings.
- Surly AI could also write a clearer headline. For the millions running it in production for decades, using a great echosystem of help support, books, consultants, and managed hosting providers; the is a noteworthy difference between the official release and a partially compatable rewrite.
- 300x faster is nice, but I mostly clicked to see if operator fusion finally explains why my GROUP BY still feels like it is doing the work by hand.
- I think in addition to making it faster, it would be useful if it could be made "leaner," e.g. can run better on lower-spec hardware than PG.
- does one not imply the other? if it can run faster in the same hardware, it should also run as fast on lower spec hardware
- No, it doesn't. Different algorithms can vary pretty wildly in performance based on the design of the hardware they run on. For instance cache sizes can make one implementation of a sort on a certain sized dataset faster or slower than another. You can have a theoretically fast algorithm that just isn't as cache efficient as a theoretically slower one in big-O terms. All levels of the memory hierarchy as well as storage have specific bandwidths and latencies that inform the real world performance results. Parallelism is another issue. Many cores can do work very fast when you're careful about how you split up work between them, taking into consideration the synchronization latency and individual cache sizes and so on. The best approach for doing work on 64 cores can be dramatically different from what works best on 1 or even 2.
- That is not a given. A database server can run faster on better hardware because it more effectively and aggressively caches things in memory, which can hurt it on lower-spec systems. Or it can better utilize SIMD instructions that are not present on the low end. Or it is more effective at utilizing more threads, but is slower when run at a low threadcount, etc etc etc.
- If I'm not mistaken ClickHouse's min spec is quite steep.
- I’m curious to see how this compares to pgColumnar or other OLAP extensions.
- At least in terms of speed, we're much faster on clickbench: https://benchmark.clickhouse.com/#system=+_b|pnc|pgrs|gQ|saB...
- The project has 2 commits.
2.
Commit #1's message is "hey claude, do a breakthrough" from a week ago and is 1.5M lines. Commit #2 is "blog post" from 4 days ago.
My head is spinning. I don't mind AI stuff or AI enabled stuff but there's gotta be some bar for ending up on HN, and also personal accountability: the lack of humility and honesty sets a new low for me. There is no "we" who "released pgrust 0.2". It's one person cosplaying a serious engineering team doing a mountain of work. The bus factor is 1, and its one you can't trust on the basics.
ex. the first 1/3 of the blog post is bloviating about how a rust for loop is faster at summing 500M numbers on the heap than loading the numbers from a table and summing them.
It leaves me in quite some anguish. This site kept me well-informed and growing for 16 years. It is no longer reliable for that as long as things like this can be the #1 post with 60 comments, with the author here, and no one mentioned any of this.
- The second line of commit #2's message is "You can find the actual git history at the v0.2 github tag." which in turn has almost 6000 commits.
I think it's a weird way to handle git history versus squashing feature branches into single commits, but it's not just one Claude session slapped up on GitHub. The OP also has a post above about their exhaustive testing which has uncovered a goodly number of bugs in Postgres itself, too. I think it's fair to say they're putting in a good amount of work on this.
- [flagged]
- pgrust looks interesting. Could it be used as a library by someone who wants a new DB for each integration test in their Rust test suite?
- One of the new features we recently built is "test mode". This brings cloning a template db from 100ms down to <10ms making it much better for tests.
If you're interested in trying it out, please reach out to me at malis@pgrust.com
- Thanks to the authors for choosing a license that respect users freedom, on top of being an awesome technical project.
- While I do like pgrust's license, I do feel like it's kind of wrong to port in such a direct way and change the license. I guess this isn't a fork, but it kind of is? It looks like according to this post [0], they did a Claude Fable + Opus re-write. I know that legally this is seemingly a valid way to do things and avoid copyright, but it feels wrong to me. I don't even necessarily think that my feelings are correct, but standing on the backs of giants and using an LLM to "reimplement" the code (not a clean room implementation) feels like it's it _shouldn't_ be a valid way to avoid copyright or allow for license changes.
Now, I don't know how MIT -> AGPL re-licensing specifics work, but still.
- My understanding is that they used c2rust, and then told the LLM/Agents to make the code more idiomatic rust, while also using the PG test suite as a feedback mechanism. This is almost certainly a derived work (and thus a fork, and should thus have the original license and copyright preserved).
For example, if I compiled PG into x86-64 assembly, and then decompiled it into C (via, say, IDA), and then polished that decompiled C code into very readable C code, it is still a derived work. For some reason, people think that if you include an LLM or Agent, copyright can be ignored, and plagiarism is now no longer possible.
It is similar to the crypto-folks thinking that if you use crypto, you no longer have to pay taxes, because the internet/computers make all inconvenient realities go away.
Honestly, such flagrant and arrogant copyright violations make it hard for me to take the project seriously, because it seems like a desperate stunt for attention (which itself may be a solid business move, but that is besides the point).
Put differently, if one were to fork pgrust, strip away the new license and copyright, and restore the original PG license and copyright (while also adding malisper+team to that copyright), they would face no legal consequences at all. In fact, they would probably be a less legal risk than the pgrust team.
- Does that license protect anything legally? This is written by AI under some initial guidance. Prompts may be protected but the output of unlicensed knowledge model corpus?
- pgrust seems to have good momentum. AGPL is an odd license for a non web project. Postgres is MIT-like, and that drove it's adoption.
Have pgrust folks reconsidered this? Else, IMO we can have an independant rust port of pgrust, which can be MIT, which will garner more attention.
- Author here. At least for databases, AGPL (or stricter) has become standard. The issue is it's so easy for megacorps (Amazon, Google, etc) to take a permissively licensed product and monetize it at the expense of the original standard.
For instance, Mongo, Cockroach, and Materialize have all gone source available. We picked AGPL because it's the best balance between open source and prevents Amazon from just repackaging it and selling it.
If AGPL is an issue for anyone, we would be happy to dual-license under a commercial license.
- There is nothing wrong with wanting to be compensated for your work, but for people like myself which use a cloud managed DB solution (GCP CloudSQL PostgreSQL) it means something like this would never be available.
I consider AGPL a poison pill in my work. That is not true with a suitable commercial license, although I expect a lot more commercial product (support/features/etc). As you note, your objective is to prevent commercialization of your software, but radically speeding up analytics is primarily a concern of large organizations so it seems like a mismatch in purpose.
- The whole point is to force the cloud providers to go through a contract negotiation for licensing rather than keep the door open for them. Nothing prevents the parent project from specially licensing to individual parties.
- > which use a cloud managed DB solution (GCP CloudSQL PostgreSQL) it means something like this would never be available
That's up to GCP then. I'm sure OP would be willing to license pgrust to them for an appropriate price.
- What's the incentive for third parties to contribute, knowing just one guy stands to benefit if the project gets monetized in such a manner?
- what exactly are your concern of using agpl db server? It doesn't force you to make rest of your system agpl.
- >we would be happy to dual-license under a commercial license
- > If AGPL is an issue for anyone, we would be happy to dual-license under a commercial license.
You should be setting up a CLA process then ASAP if you are looking to take outside contributions.
- Sure, that's your prerogative, and kudos for not talking up open source. I'm not amazon size so can't use it, and AGPL is a no go for DB, don't want to be forced to open source my app because I use this!
Will await a MIT based fork myself.
- Why would AGPL force you to open source your app? Unless you literally compile your app with pgrust by modifying the pgrust source code, you're safe. Clients aren't bound by the AGPL because they aren't derived works.
- Here we go again.
AGPL is untested in courts. There is no definitive definition of what could be considered within the blast radius such that it would require AGPL licensing.
There's a reason AGPL is banned at Google and most sane companies. It's simply too dangerous.
You can't simply say "clients aren't bound" because it depends.
I'd rather see the BSL used here to be perfectly honest. At least it's simple.
- Oh bummer. I was really excited about pgrust but AGPL is a dealbreaker. Not for me personally, but it will never see wide adoption because it’s a banned license in most corporate environments. Lack of path to wide adoption means it’s dead in the water.
It’s weird because those who actually care about optimized pg gains are most likely large corporate customers. Why make a product targeting them and license it in such a way they’ll never use it?
This also hard blocks upstreaming any beneficial features into core Postgres.
- If it is indeed 300* faster, I'm sure more rational corporations will rethink their license policy or be left in the dust.
- License policies are made by lawyers not by programmers. And their competition will be in the exact same boat (different lawyers though). AGPL is so toxic that it tends to be checked for during M&A processes so even if the current batch of lawyers is ok with it there is a chance that a later batch of lawyers is not. Given that the target audience for this project is the larger companies you are going to end up with a very nice project and zero actual users or you will end up with AWS et all stealing your work. Databases are very hard to do successfully commercially, at a minimum you should dual license them (AGPL for 'home' use and commercial licensing for parties that will want to buy the upside but they'll demand support and other stuff besides).
- A project can probably use the EUPL instead of the AGPL (EUPL is to AGPL, as MPLv2[0] is to GPL). Basically, EUPL is file-based, not project based, and so it is not aggressively viral. You can use the EUPL code any way you want, as long as you make the original code available, plus any modification to the original files.
[0]: With Exhibit B, which prevents relicensing to GPL. It is also analogous to CDDL.
- This is the case when you have your own datacenter. This isn't as big a problem in the cloud. There are ways to write licenses that prevent cloud providers from stealing while allowing customers to use the software and being required to pay for it. The problem with the AGPL has to with its viral nature, not its provisions to prevent cloud vendor theft.
- Its 300x faster for certain tests. I could pretty easily craft tests that do this on two different systems. The author mentioned this when they talked about being able to fit an entire ResultSet into memory. That's the real trick with performance. Very few workloads are CPU bound anymore (linear algebra on the CPU for example). Almost all workloads are memory bound. So its all about moving data from memory to network, back to memory and back to network, over and over again through your microservices or DBs. If the entire working set can fit in memory, you get at least a 10x performance boost. If you have to keep even a part of the working set on disk, its a huge performance loss. And the larger fraction of the working set on disk, the worse the performance loss.
PS Learn how DBs do joins for more information. Specifically the differences between hash joins, merge joins and nested loop joins. They are basically fancy ways to page part of your working set to disk at huge performance penalties.
PPS As memory gets more expensive, these techniques get more valuable. When it gets cheap, they lose value.
- They could simply spend a few months and a few million tokens and get their own port, no?
I doubt even 30000x faster would prompt a policy change.
- Are you suggesting the AI just rewrites the whole thing under a different license? There’s no way that’s not more dicey than the AGPL license.
- That's exactly what they did here, I don't see the difference.
- This is the difference. This guy took an existing source base, had Claude find specific bugs, then had Claude fix a specific bug which was then reviewed by a person. We also don't know if these changes introduce new problems yet. You are suggesting letting Claude write an entirely new source base. That's light-years away from what happened here.
- They did it on permissively licensed code would be the difference.
- No. This is system code. You let an LLM loose on it, it probably fixed 20 bugs and introduces 200 more plus 5 different performance regressions. In all fairness, your average app programmer would have the same problems. That's why it takes so long to learn to be a system programmer and why it takes so long to do anything on a systems source base. For reference, systems are OSes, DBs and compilers (although compilers are very different in many ways).
Also, you can successfully sell a systems project that is only 10% faster. 30000x faster and they are throwing illegal and debauched things through your window to get access to your improvements.
- > No. This is system code. You let an LLM loose on it, it probably fixed 20 bugs and introduces 200 more plus 5 different performance regressions.
What does it being system code have to do with anything?
- The brittleness and difficult of writing the code. The extra complexity. The cost of bugs. And ease of creating performance degradations. If you are even asking that question, you probably don't have the experience to even have an opinion on this.
Most systems code has to be "canonical". That means it can't be improved. Imagine the most difficult PR review you have ever had. Now give the reviewers OCD and an extra 30 IQ points. And every single code change has to live up to this standard. That's what it has to do with this.
- What do you mean, "No"? Letting an LLM loose on Postgres is exactly what happened here: https://malisper.me/pgrust-rebuilding-postgres-in-rust-with-...
- So, keyword 'rational', I'm not sure any sufficiently large company is a rational actor.
Yes, at [tech corp dayjob], any dependency is likely to be banned for arbitrary reasons if you bring it to the attention of the wrong people. It doesn't have to go against any of our policies e.g. don't mention anything with GPL in the name around the "risk" people. In fact, do not ever talk to the "risk" people and hope they don't talk to you.
Latest news: Apparently, devtools are a legal risk. Basic reverse-engineering of client-side JS is now banned.
The delusions really seem to scale with headcount.
- AGPL seems like the right choice to me. I’m tired of companies like PlanetScale taking PostgreSQL, building a business on top of it, and then acting like PostgreSQL is theirs to control.
- we have not once claimed postgres is under our control. i don't think you understand how open source works but thats ok.
- > It’s weird because those who actually care about optimized pg gains are most likely large corporate customers. Why make a product targeting them and license it in such a way they’ll never use it?
Wait, sorry, you're asking why make something enterprise customers might pay for, and then not give it away to them for free?
- Good news, we now know up front what an independent port would cost, and it’s not much. So no reason AWS, Google, and friends couldn’t bang out their own port if they want, binary-compatible with this one.
What we don’t know for sure is whether there is any copyright for LLM-generated code. The license might be irrelevant!
- 2 commits in the repo both generated by claude. This is AI slop, I wonder where you see good momentum?
- main indeed has two commits, but it clearly states the location of the rest of the commits, so I wouldn't be critical of main itself.
Now we see https://github.com/malisper/pgrust/tree/v0.2 has almost 6000 commits in it, with the very first one on 2026-07-02. That's a lot of token momentum!hey claude, do a breakthrough You can find the actual git history at the v0.2 github tag. Co-Authored-By: Fable <noreply@anthropic.com>It's easy to claim AI slop nowadays, but you should still mistrust-but-verify.
- What do we mean by "easy to claim"? It is written by AI. The 6000 commits are by Claude.
- 4K stargazers in a week.
- Everything around Rust is political, so the license choices are also about political statement.
- Most official Rust projects are dual MIT/Apache licensed by convention [1] (and most Rust libraries from third parties I've seen that are open source MIT follow the MIT/Apache dual license), so seems like this library shouldn't just be AGPL for a typical political choice of a Rustacean?
- official projects are usually run by sensible people who want to do things and aren't leading with ideology.
We're literally talking about an "X but in Rust" project already...
- I'll take the 300x slower non-vibe-coded pg, thanks!
- They disabled Postgres parallelism to benchmark too. Sigh.
- We disabled parallelism in the blog post for demonstration purposes. The 300x slower refers to the clickbench numbers[0] where parallelism is enabled
[0] https://benchmark.clickhouse.com/#system=+liH|pgrs|gQ&type=-...
- Who is "we"?
- The pgrust team asked me to look at their results on a review system, and I confirmed the ClickBench speedup there. Regular PostgreSQL is really terrible at some of these queries. Unfortunately fixing that is hard to do in core itself because columnar storage lives outside of the main tree, and some optimization problems only show up when layered on columnar.
- Me and Jason, the two people working on the project
- hi
- [dead]
- [dead]
218 points by poly2it 11 hours ago | 96 comments