Breaking the WAL

28 points by wwilson 2 hours ago | 27 comments

carlsverre 1 hour ago |

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...

MPSimmons 1 hour ago | |

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.

carlsverre 1 hour ago | | |

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...

minimaltom 1 hour ago |

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.

carlsverre 1 hour ago | |

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

minimaltom 1 hour ago | | |

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

gamegoblin 1 hour ago | |

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

biorach 1 hour ago | | |

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

uhohherewegoaga 1 hour ago |

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.

wwilson 1 hour ago |

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!

john_strinlai 2 hours ago |

related: Tailscale Traces Database Corruption to 16y/o SQLite WAL-Reset Bug (tailscale.com) https://news.ycombinator.com/item?id=49272832

grebc 1 hour ago |

[flagged]

carlsverre 1 hour ago | |

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.

returningfory2 1 hour ago | | |

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.

grebc 1 hour ago | | |

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.

Mawr 1 hour ago | | |

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.

wwilson 1 hour ago | |

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!

grebc 1 hour ago | | |

So you found this before SQLite published the fix?