Hi! I'm the author and the person who was on a road trip when I decided to do this experiment back in July. I'm super excited to finally get the blog post out, and even better, alongside a post describing the root cause process that went into the project. You can read about Tailscale's story here: https://tailscale.com/blog/sqlite-wal-reset-bug
I'm only sad that I didn't put SQLite under test earlier in the year, or I would have found this issue right away. If you look at the workload[1], you can see how simple it is. Exactly the same kind of workload we write every day to help our customers find bugs.
If you have any questions about our process or how debugging with Antithesis works, please let me know! Thanks for reading!
How difficult would it have been to isolate that problem if you didn't already know the SQLite subsystem it was in? This feels, to someone relatively ignorant of the SQLite / Tailscale / Antithesis architectures as a "hindsight is 20/20" kind of thing, but I'm open to learning more.
Great question! The general approach we take with transactional systems like this is to put reachability statements throughout the complex stateful machinery and then stress-test them in Antithesis.
The workload I put in place does exactly that[1]. It runs a write workload from multiple processes concurrently on the same SQLite database to cause writes to build up in the WAL, and runs checkpoints concurrently. This exercises the portion of the WAL code that, from a trivial read-through, is most likely to contain bugs (and turns out, did!).
Said differently, this is exactly the approach we take with all stateful transactional systems. I am only sad that I didn't do this experiment months earlier, as it would have saved Tailscale and the SQLite team a lot of time.
> This exercises the portion of the WAL code that, from a trivial read-through, is most likely to contain bugs (and turns out, did!).
Suuure. So why exactly haven't you found this bug already a long time ago? I mean it is trivially obvious that this is where bugs in SQLite would be, right?
SQLite is open source and is one of the most popular databases, surely testing it to find bugs would also be obvious?
We put many things to the test, but it costs money, and until very recently we haven't had a way to do this without a lot of human oversight. Now that we have https://github.com/antithesishq/antithesis-skills and LLMs are getting pretty reliably good at writing workloads, we are looking at ways to open up our technology to the OSS community. We have already started down that path with Etcd last year (https://etcd.io/blog/2025/autonomus_testing_with_antithesis/) and are working with various OSS groups to put more systems under test in a responsible way.
I'm not gonna lie, though, you're hitting a sore point. I'm absolutely kicking myself for not putting SQLite under test when I joined. The first few projects I put under test after joining were all projects that had a few properties:
1. The project had a path to a commercial deal
2. The project likely had interesting bugs to find
TBH, I didn't expect that one of the most widely used and well-tested pieces of software in existence would have a bug our fuzzer could find in 15 mins. (To be clear, this is not a hit on SQLite, just that our fuzzer is stupid good at finding race conditions. SQLite remains one of my favorite pieces of software ever.)
With this in mind, are there other similar projects that I'm not thinking of that I should put under test? Before you ask, I'm already scaling up my SQLite harness to cover more of its state space and run against master. Maybe I'll find something novel in SQLite after all! Never too late to check.
Brutally honest answer: it only recently became easy to get random systems under test with Antithesis without a bunch of manual/expert effort. Expect us to be reporting a lot of bugs against open source projects in the near future.
Nice article, but could you get someone to fix the blog layout? Having the author card pinned to the window rather than being able to scroll past it is distracting.
I went clicking through to see if I could find the prompt they fed the AI to locate the issue / write the test suite.
I couldn't find it, so its unclear if the prompt was completely "make a test suite" or was lead towards finding it in the first place, which wouldn't be a fair test.
The closest I mention of the prompt I could find was:
> Then I asked it to write a simple workload which exercised the WAL insert and checkpoint code. Notably, this is a completely generic workload.
Hey, author here! Good question. The prompt was to use the Antithesis skills[1] to write a workload that exercises SQLite's insertion and checkpoint machinery across multiple processes using a specific version of SQLite (3.51.2). The agent who did the work was aware of the bug and knew I wanted to see if Antithesis could reproduce it. However, I specifically designed the workload the same way we approach testing other WAL checkpoint-style systems: by putting them under stress. The workload itself does not know about the bug.
My goal, which I believe I accomplished, was to determine if Antithesis would have found this bug if we had put the SQLite WAL under test.
Which is: "Given the knowledge of a kind of bug that can happen in one Raft implementation, can a similar workload find similar bugs in other Raft implementations?" Turns out it can.
Yeah 100%! And I'm sorry if I sound a little more critical and less eager, its just thats theres a world of difference between a priori finding the bug, and reproducing it, and the impression of the article (from my read) was the former.
But please keep writing, I know its super hard to put yourself out there and make content!
Their founder guy says[1] they found some other bugs while doing this that they will report soon, so if that is true, seems more plausibly like a pretty generic thing.
Looking forward to seeing the other bugs they found.
We wanted to publish this ASAP since people are talking about this bug today. In a follow-up, we will show how our automated causality analysis could have saved Tailscale and SQLite from 6 months of root cause analysis. Stay tuned!
as a long time lurker who usually enjoys antithesis posts, this was a pretty existentially sad read. we all know that company blog posts here always tend to be ads at their core, but historically afaicr antithesis generally toes the line well of keeping the "thing we sell is really good" as a side dish to the "i (author) did very interesting deep thinking and/or hard work to accomplish something cool" entree.
this one falls very flat on that front imo. the "antithesis finds rare bugs" is the overwhelming majority of the actual content here, where as the usual meat of the blog, the hard work, the deep thinking, really distills down to "typed this into claude mobile, typed that into claude mobile". yes, its a statement to the power of the technologies of today that's all it takes to find this bug, but man it's a sad signal for those of us looking for compelling technical reads instead of b2b sales pitches.
Carl (author) here. This is a poor reading of the blog post. Notably, I built a general-purpose workload that simply exercises the write/checkpoint machinery. The exact same workload we use to break all kinds of transactional systems. This generic workload found the bug without any special knowledge of the bug, due to how Antithesis works.
I think the issue is that you knew there was a bug to be found in the write and checkpoint interactions, which then determined which workload you built.
My understanding is that this workload is very uncommon: the Tailscale blog says they used a custom unusual configuration to have many checkpoints like this. So without knowing the bug, it seems unlikely one would build this workload and then find the bug. If that makes sense.
Edit: just want to say that you being able to repro it is awesome, but that the overall claim seems a little overstated to me.
I can see how the post comes across that way. I may need to edit it to somehow be clearer about how I approached the workload development. I'll refer you to my other replies to comments for more information on my approach:
Hi! I'm the author and the person who was on a road trip when I decided to do this experiment back in July. I'm super excited to finally get the blog post out, and even better, alongside a post describing the root cause process that went into the project. You can read about Tailscale's story here: https://tailscale.com/blog/sqlite-wal-reset-bug
I'm only sad that I didn't put SQLite under test earlier in the year, or I would have found this issue right away. If you look at the workload[1], you can see how simple it is. Exactly the same kind of workload we write every day to help our customers find bugs.
If you have any questions about our process or how debugging with Antithesis works, please let me know! Thanks for reading!
[1]: https://github.com/antithesishq/sqlite/blob/3.51.2-instrumen...
How difficult would it have been to isolate that problem if you didn't already know the SQLite subsystem it was in? This feels, to someone relatively ignorant of the SQLite / Tailscale / Antithesis architectures as a "hindsight is 20/20" kind of thing, but I'm open to learning more.
Great question! The general approach we take with transactional systems like this is to put reachability statements throughout the complex stateful machinery and then stress-test them in Antithesis.
The workload I put in place does exactly that[1]. It runs a write workload from multiple processes concurrently on the same SQLite database to cause writes to build up in the WAL, and runs checkpoints concurrently. This exercises the portion of the WAL code that, from a trivial read-through, is most likely to contain bugs (and turns out, did!).
Said differently, this is exactly the approach we take with all stateful transactional systems. I am only sad that I didn't do this experiment months earlier, as it would have saved Tailscale and the SQLite team a lot of time.
[1]: https://github.com/antithesishq/sqlite/blob/3.51.2-instrumen...
> This exercises the portion of the WAL code that, from a trivial read-through, is most likely to contain bugs (and turns out, did!).
Suuure. So why exactly haven't you found this bug already a long time ago? I mean it is trivially obvious that this is where bugs in SQLite would be, right?
SQLite is open source and is one of the most popular databases, surely testing it to find bugs would also be obvious?
We put many things to the test, but it costs money, and until very recently we haven't had a way to do this without a lot of human oversight. Now that we have https://github.com/antithesishq/antithesis-skills and LLMs are getting pretty reliably good at writing workloads, we are looking at ways to open up our technology to the OSS community. We have already started down that path with Etcd last year (https://etcd.io/blog/2025/autonomus_testing_with_antithesis/) and are working with various OSS groups to put more systems under test in a responsible way.
I'm not gonna lie, though, you're hitting a sore point. I'm absolutely kicking myself for not putting SQLite under test when I joined. The first few projects I put under test after joining were all projects that had a few properties: 1. The project had a path to a commercial deal 2. The project likely had interesting bugs to find
TBH, I didn't expect that one of the most widely used and well-tested pieces of software in existence would have a bug our fuzzer could find in 15 mins. (To be clear, this is not a hit on SQLite, just that our fuzzer is stupid good at finding race conditions. SQLite remains one of my favorite pieces of software ever.)
With this in mind, are there other similar projects that I'm not thinking of that I should put under test? Before you ask, I'm already scaling up my SQLite harness to cover more of its state space and run against master. Maybe I'll find something novel in SQLite after all! Never too late to check.
Other embedded databases like Dolt? DuckDB? git?
Brutally honest answer: it only recently became easy to get random systems under test with Antithesis without a bunch of manual/expert effort. Expect us to be reporting a lot of bugs against open source projects in the near future.
Nice article, but could you get someone to fix the blog layout? Having the author card pinned to the window rather than being able to scroll past it is distracting.
I went clicking through to see if I could find the prompt they fed the AI to locate the issue / write the test suite.
I couldn't find it, so its unclear if the prompt was completely "make a test suite" or was lead towards finding it in the first place, which wouldn't be a fair test.
The closest I mention of the prompt I could find was:
> Then I asked it to write a simple workload which exercised the WAL insert and checkpoint code. Notably, this is a completely generic workload.
With a skeptical lens, unclear.
Hey, author here! Good question. The prompt was to use the Antithesis skills[1] to write a workload that exercises SQLite's insertion and checkpoint machinery across multiple processes using a specific version of SQLite (3.51.2). The agent who did the work was aware of the bug and knew I wanted to see if Antithesis could reproduce it. However, I specifically designed the workload the same way we approach testing other WAL checkpoint-style systems: by putting them under stress. The workload itself does not know about the bug.
My goal, which I believe I accomplished, was to determine if Antithesis would have found this bug if we had put the SQLite WAL under test.
[1]: https://github.com/antithesishq/antithesis-skills
Thanks for clarifying! It would be really interesting if Antithesis finds the bug when:
1. The specific bug isnt mentioned 2. (If youre game) a model with a knowledge-cutoff date before the report is used
This is a solid idea. However, something that is maybe similar to this idea is this: https://antithesis.com/blog/2026/finding-bugs-in-raft-implem...
Which is: "Given the knowledge of a kind of bug that can happen in one Raft implementation, can a similar workload find similar bugs in other Raft implementations?" Turns out it can.
Yeah 100%! And I'm sorry if I sound a little more critical and less eager, its just thats theres a world of difference between a priori finding the bug, and reproducing it, and the impression of the article (from my read) was the former.
But please keep writing, I know its super hard to put yourself out there and make content!
Thank you! I appreciate the note of support! I'd love to write more blog posts in the future.
Their founder guy says[1] they found some other bugs while doing this that they will report soon, so if that is true, seems more plausibly like a pretty generic thing.
Looking forward to seeing the other bugs they found.
1. https://news.ycombinator.com/item?id=49278351
looks like the bug fixes are in individual commits here: https://github.com/sqlite/sqlite/compare/master...antithesis...
I'd be interested in seeing what the SQLite people have to say about them if/when they are submitted upstream
Those are the commits from upstream that resolved the bug. Notice that most of them were committed by Dr. Richard Hipp himself.
oh! damn! I did not notice that. That does answer my question.
We wanted to publish this ASAP since people are talking about this bug today. In a follow-up, we will show how our automated causality analysis could have saved Tailscale and SQLite from 6 months of root cause analysis. Stay tuned!
as a long time lurker who usually enjoys antithesis posts, this was a pretty existentially sad read. we all know that company blog posts here always tend to be ads at their core, but historically afaicr antithesis generally toes the line well of keeping the "thing we sell is really good" as a side dish to the "i (author) did very interesting deep thinking and/or hard work to accomplish something cool" entree.
this one falls very flat on that front imo. the "antithesis finds rare bugs" is the overwhelming majority of the actual content here, where as the usual meat of the blog, the hard work, the deep thinking, really distills down to "typed this into claude mobile, typed that into claude mobile". yes, its a statement to the power of the technologies of today that's all it takes to find this bug, but man it's a sad signal for those of us looking for compelling technical reads instead of b2b sales pitches.
related: Tailscale Traces Database Corruption to 16y/o SQLite WAL-Reset Bug (tailscale.com) https://news.ycombinator.com/item?id=49272832
[flagged]
Carl (author) here. This is a poor reading of the blog post. Notably, I built a general-purpose workload that simply exercises the write/checkpoint machinery. The exact same workload we use to break all kinds of transactional systems. This generic workload found the bug without any special knowledge of the bug, due to how Antithesis works.
I think the issue is that you knew there was a bug to be found in the write and checkpoint interactions, which then determined which workload you built.
My understanding is that this workload is very uncommon: the Tailscale blog says they used a custom unusual configuration to have many checkpoints like this. So without knowing the bug, it seems unlikely one would build this workload and then find the bug. If that makes sense.
Edit: just want to say that you being able to repro it is awesome, but that the overall claim seems a little overstated to me.
I can see how the post comes across that way. I may need to edit it to somehow be clearer about how I approached the workload development. I'll refer you to my other replies to comments for more information on my approach:
https://news.ycombinator.com/item?id=49278424 https://news.ycombinator.com/item?id=49278521
Also appreciate the nice words at the end :) I'm feeling a bit ganged up on.
I read your ad.
It states you replicated the bug once the SQLite team fixed it, and published it.
Not sure what’s difficult about replicating behaviour when it’s spelled out for you.
this is a deliberate mischaracterisation of the content of the article
You’re mischaracterising an ad, that’s also funnily enough full of it.
Okay, so use the same methodology to find a new, unknown bug? That'd be actually impressive.
As-is, this is just 20/20 hindsight with concerns about leading the AI on through the prompt hand-waved away. Come on.
As the post (and linked repo) pretty clearly indicate, zero info about the existence of the bug went into this work.
Oh btw, we also found some other ones… Stay tuned!
So you found this before SQLite published the fix?