Jepsen: TigerBeetle 0.16.11
jepsen.io
jepsen.io
Fuzzer Blind Spots (Meet Jepsen!) – https://tigerbeetle.com/blog/2025-06-06-fuzzer-blind-spots-m...
This report found a number of issues, which might be a cause for concern. But I think it's a positive because they didn't just fix the issues, they've expanded their internal test suite to catch similar bugs in future. With such an approach to engineering I feel like in 10 years TigerBeetle would have achieved the "just use Postgres" level of default database in its niche of financial applications.
Also great work aphyr! I feel like I learned a lot reading this report.
Yes, we have around 6,000+ assertions in TigerBeetle. A few of these were overtight, hence some of the crashes. But those were the assertions doing their job, alerting us that we needed to adjust our mental model, which we did.
Otherwise, apart from a small correctness bug in an internal testing feature we added (only in our Java client and only for Jepsen to facilitate the audit) there was only one correctness bug found by Jepsen, and it didn’t affect durability. We’ve written about it here: https://tigerbeetle.com/blog/2025-06-06-fuzzer-blind-spots-m...
Finally, to be fair, TigerBeetle can (and is tested) to survive more faults than Postgres can, since it was designed with an explicit storage fault model and using research that was not available at the time when Postgres was released in ‘96. TB’s fault models are further tested with Deterministic Simulation Testing and we use techniques such as static memory allocation following NASA’s Power of Ten Rules for Safety-Critical Code. There are known scenarios in the literature that will cause Postgres to lose data, which TigerBeetle can detect and recover from.
For more on this, see the section in Kyle’s report on helical fault injection (most Raft and Paxos implementations were not designed to survive this) as well as a talk we gave at QCon London: https://m.youtube.com/watch?v=_jfOk4L7CiY
I have followed TigerBeetle with interest for a while, and thank you for your inspirational work and informative presentations.
However, you have stated in several occasions that the lack of memory safety in Zig is not a concern since you don't dynamically allocate memory post startup. However, one of the defects uncovered here (#2435) was caused by dereferencing an uninitialized pointer. I find this pretty concerning, so I wonder if there is something that you will be doing differently to eliminate all similar bugs going forward?
Which is, yeah, one of the bigger technical challenges for us --- we ship language-native libraries for Go,node,Java,C#,Python and Rust, and, like in the Tolstoi novel, each one is peculiar in its own way. What's worse, they aren't directly covered by our deterministic simulator. That's one of the major reasons why we invest in full-system simulation with jepsen, antithesis and vortex (https://tigerbeetle.com/blog/2025-02-13-a-descent-into-the-v...). We are also toying with the idea of generating _more_ of that code, so there's less room for human error. Maybe one day we'll even do fully native client (eg, pure Java, pure Go), but we are not there yet.
One super-specific in-progress thing is that, at the moment, the _bulk_ of the client testing is duplicated per client, and also the _bulk_ of the testing is example-based. Building simulator/workload is a lot of work, and duplicating it for each client is unreasonable. What we want to do here is to use multi-process architecture, where there's a single Zig process that generates the workloads and generates interesting sequences of commands for clients, and than in each client we implement just a tiny "interpreter" for workload language, getting a test suite for free. This is still WIP though!
Regarding the broader memory safety issue in the database. We did have a couple of memory safety bugs, which were caught early in testing. We did have one very bad aliasing bug, which would have been totally prevented by Rust, which slipped through the bulk of our testing and into the release (it was caught in testing _after_ it was introduced): https://github.com/tigerbeetle/tigerbeetle/pull/2774. Notably, while the bug was bad enough to completely mess up our internal data structure, it was immediately caught by an assert down the line, and downgraded from correctness issues to a small availability issues (just restarting the replica would fix it). Curiously, the root cause for that bug was that we over-complicated our code. Long before the actual bug we felt uneasy about the data structure in question, and thought about refactoring it away (which refactor is underway. Hilariously, it looks that just "removing" the thing without any other code changes improves performance!).
So, on balance, yeah, Rust would've prevented a small number of easy bugs, and on gnarly bug, but then the entire thing would have to look completely different, as the architecture of TigerBeetle not at all Rust-friendly. I'd be curious to see someone replicating single-thread io-uring no malloc after startup architecture in Rust! I personally don't know off the top of my head whether that would work or not.
Anyways, really interesting report and project. I also like your youtube show - keep up the great work! :)
What are you referencing here?
This surveys the excellent storage fault research from UW-Madison, and in particular:
“Can Applications Recover from fsync Failures?”
“Protocol-Aware Recovery for Consensus-Based Storage”
[disks are] somewhere between non-byzentine fault tolerance and
Byzantine fault tolerance ... you expect the disk to be almost
an active adversary ...
...
so you start to see just a single disk as a distributed system
Combined with https://www.youtube.com/watch?v=tRgvaqpQPwE it seems like the author/presenter is conflating local/disk-related properties/details with distributed/system-based requirements/guarantees. If consensus requires a node to have durably persisted some bit of state before it sends a particular message to other nodes in the distributed system, then it doesn't matter how that persistence is implemented, it only matters how that persistence is observable, disks and FS caches and etc. aren't requirements, they're just one of many possible implementation choices.
It’s a mindbender of a paradigm-shift for how to think about local recovery actions in the context of the global consensus protocol!
> Disks and flash devices exhibit a subtle and complex failure model: a few blocks of data could become inaccessible or be silently corrupted [8, 9, 32, 59]. Although such storage faults are rare compared to whole-machine failures, in large-scale distributed systems, even rare failures become prevalent [60, 62]. Thus, it is critical to reliably detect and recover from storage faults.
It's not true in general that a node in a distributed system binds its persisted state to a local disk, or flash device, or any specific implementation of any specific kind of storage system. The storage layer is the responsibility of the node to manage, and irrelevant to the wider distributed system in which the node is a participant. Any of those kinds of storage faults need to be accommodated by the node that utilizes those storage layers, but their specific details don't need to be communicated beyond the specific node where they apply. And it's not at all critical for those nodes to detect and recover from any faults in their storage layer; those faults can easily be communicated to the broader distributed system, which necessarily must be able to handle node-specific failures like those without breaking everything down!
> how to think about local recovery actions in the context of the global consensus protocol!
Local recovery actions, or any other kinds of node-specific details, have no relevance or influence on the information communicated thru the the global consensus protocol. A node can persist its state to a local disk, or to RAM, or to an S3 object, or anything else, and none of these details matter at all to the details that node communicates to other nodes in its cluster.
tl;dr: nodes don't necessarily persist state to local disk
In fact, an early version of TigerBeetle implemented one of these from Cowling and Liskov’s VSR’12 paper.
However, since then, we invested in TigerBeetle’s stable storage, for reasons which I won’t go into further here.
If you are curious to learn more about all the stable storage techniques in TigerBeetle in particular (again, this is not trying to suggest that VSR can’t also run without stable storage, or to deny other techniques such as object storage or tiering!), but if these things are interesting for you, and if you do want to learn more, then I’d recommend diving into “Durability and the Art of Consensus”.
To be clear, “Durability and the Art of Consensus” is not a paper.
> tl;dr: nodes don't necessarily persist state to local disk
In the consensus literature, this is sometimes referred to as “diskless crash recovery”. For example, see work by Dan Ports.
> Local recovery actions, or any other kinds of node-specific details, have no relevance or influence on the information communicated thru the the global consensus protocol.
This goes directly against the central finding of PAR, which gives counter-examples where your statement does not universally hold true.
Again, it’s not intuitive. And that’s why it won FAST18—because it says “everything we know is wrong”. Complete red pill and mindbender.
More broadly, there's no concept of "crash recovery" at the system level, which has any meaningful utility. Nodes are either there or they're not there, exactly how they crash or recover from their crashes are irrelevant to the overall distributed system, insofar as if a crashed-and-recovered node comes back online, it's gonna need to re-sync with its peers before it can talk to anyone else, and that's not anything to do with "crash recovery" related to local disk or anything like that (which is all implementation details of the node itself) -- it's just normal node synchronization, orthogonal to any state storage stuff of the node.
PAR is something that your system can maybe implement, it's not any kind of rule or definition or requirement that all systems of some classification must satisfy..!
https://csrc.nist.gov/projects/automated-combinatorial-testi...
My intro to other categories of test generation was usually this paper:
https://cs.stanford.edu/people/saswat/research/ASTJSS.pdf
Maybe see of your team can build combinatorial- or path-based testing in Zig next.
https://csrc.nist.gov/CSRC/media/Projects/Automated-Combinat...
Are they enabled in production? Are there some expensive ones that aren’t?
It’s not expensive.
Because we batch, this naturally separates the control plane from the data plane, amortizing assertions against the (larger) buffers now flowing through the data plane.
We do also have some intensive online verification checks, and these are gated behind a comptime flag.
Finally, we compile Zig with ReleaseSafe and further have all Zig’s own assertions enabled. For example, checked arithmetic for bounds overflow, which is not something you see enabled by default in safe builds for most languages, but which is critically important for safety.
The reason why all this is so important, is because if your program does something wrong in production, with people’s money, you want to know about it immediately and shutdown safely.
In other words, production is where you most need the safety, not in development (although you obviously want them there too to find bugs faster). But again, it’s the bugs that make it to production that we’re trying to catch with assertions.
> it’s the bugs that make it to production that we’re trying to catch with assertions.
Nicely put, I think I’ll steal this!
Question about how people end up using TigerBeetle. There's presumably a lot of external systems and other databases around a TigerBeetle install for everything that isn't an Account or Transfer. What's the typical pattern for those less reliable systems to square up to TigerBeetle, especially to recover from consistency issues between the two?
The typical pattern in integrating TigerBeetle is to differentiate between control plane (Postgres for general purpose or OLGP) and data plane (TigerBeetle for transaction processing or OLTP).
All your users (names, addresses, passwords etc.) and products (descriptions, prices etc.) then go into OLGP as your "filing cabinet".
And then all the Black Friday transactions these users (or entities) make, to move products from inventory accounts to shopping cart accounts, and from there to checkout and delivery accounts—all these go into OLTP as your "bank vault". TigerBeetle lets you store up to 3 user data identifiers per account or transfer to link events (between entitites) back to your OLGP database which describes these entities.
This architecture [1] gives you a clean "separation of concerns", allowing you to scale and manage the different workloads independently. For example, if you're a bank, it's probably a good idea not to keep all your cash in the filing cabinet with the customer records, but rather to keep the cash in the bank vault, since the information has different performance/compliance/retention characteristics.
This pattern makes sense because users change their name or email address (OLGP) far less frequently than they transact (OLTP).
Finally, to preserve consistency, on the write path, you treat TigerBeetle as the OLTP data plane as your "system of record". When a "move to shopping cart" or "checkout" transaction comes in, you first write all your data dependencies to OLGP if any (and say S3 if you have related blob data) and then finally you commit your transaction by writing to TigerBeetle. On the read path, you query your system of record first, preserving strict serializability.
Does that make sense? Let me know if there's anything here we can drill into further!
[1] https://docs.tigerbeetle.com/coding/system-architecture/
It looks like the segfaults on the JNI side would not have been protected if Rust or some other memory safe language were being used - the lack of memory safety bugs gives some decent proof that TigerBeetle's approach to Zig programming (TigerStyle iirc, lol) does what it sets out to do.
EDIT: But, yeah, totally, if not for TigerStyle, we'd die to nasal demons!
Special kudos to the founders who are sharing great insights in this thread.
Appreciate your kind words too, and look forward also to sharing something new in our talks at SD25 in Amsterdam soon!
This is an interesting meta pattern where doing something _harder_ actually simplifies the system.
Another example is that, because we assume that the disk can fail and need to include repair protocol, we get state-synchronization for a lagging replica "for free", because it is precisely the same situation as when the entire disk gets corrupted!
In TigerBeetle, we take advantage of some special properties to make the state machine checking part linear-time. We let TigerBeetle tell us exactly which transactions happen. We can do this because it's a.) strong serializable, b.) immutable (in that we can inspect DB state to determine whether an op took place), and c.) exposes a totally ordered timestamp for every operation. Then we check that that timestamp order is consistent with real-time order, using a linear-time cycle detection approach called Elle. Having established that TigerBeetle's claims about the timestamp order are valid, we can apply those operations to a simulated version of the state machine to check semantic correctness!
I'd like to generalize this to other systems, but it's surprisingly tricky to find all three of those properties in one database. Maybe an avenue for future research!
I think it should be http://pmg.csail.mit.edu/papers/vr-revisited.pdf (http scheme not https) ?
And now I have some Friday evening reading material.
The VSR 2012 paper is one of my favorites as is “Protocol-Aware Recovery for Consensus-Based Storage”, which is so powerful.
Hope you enjoy the read!
After reading the Jepsen report on TigerBeetle, the related blog post, and briefly reviewing the Antithesis integration code on GitHub workflow, I'm trying to better understand the testing scope.
My core question is: could these bugs detected by the Jepsen test suite have also been found by the Antithesis integration?
This question comes from a few assumptions I made, which may be incorrect:
- I thought TigerBeetle was already comprehensively tested by its internal test suite and the Antithesis product.
- I had the impression that the Antithesis test suite was more robust than Jepsen's, so I was surprised that Jepsen found an issue that Antithesis apparently did not.
I'm wondering if my understanding is flawed. For instance:
1. Was the Antithesis test suite not fully capable of detecting this specific class of bug?
2. Was this particular part of the system not yet covered by the Antithesis tests?
3. Am I fundamentally comparing apples and oranges, misunderstanding the different strengths and goals of the Jepsen and Antithesis testing suites?
I would greatly appreciate any insights that could help me understand this better. I want to be clear that my goal is to educate myself on these topics, not to make incorrect assumptions or assign responsibility.
There are almost certainly blind spots in the Jepsen test generators too--that's part of why designing different generators is so helpful!
1. Some sort of environment, which can run the system. The simplest environment is to spin up a real cluster of machines, but ideally you want something fancier, to improve performance, control over responses of external APIs, determinism, reproducibility, etc. 2. Some sort of load generator, which makes the system in the environment do interesting thing 3. Some sort of auditor, which observes the behavior of the system under load and decides whether the system behaves according to the specification.
Antithesis mostly tackles problem #1, providing a deterministic simulation environment as a virtual machine. The same problem is talked by jepsen (by using real machines, but injecting faults at the OS level), and by TigerBeetle's own VOPR (which is co-designed with the database, and for that reason can run the whole cluster on just a single thread). There there approaches are complimentary and are good at different things.
For this bug, the critical part was #2, #3 --- writing workload verifier and auditor that actually can trigger the bug. Here, it was aphyr's 1600 lines of TigerBeetle-specfic Clojure code that triggred and detected the bug (and then we patched _our_ equivalent to also trigger it. Really, what's buggy here is not the database, but the VOPR. Database having bugs is par of course, you can't just avoid bugs through the sheer force of will. So you need testing strategy that can trigger most bugs, and any bug that slips through is pointing to the deficiency in the workload generator.)
https://github.com/jepsen-io/tigerbeetle/blob/main/src/jepse...
To understand why the query engine bug slipped through, see: https://tigerbeetle.com/blog/2025-06-06-fuzzer-blind-spots-m...
Otherwise check out https://tigerbeetle.com/company if you want more about the corporate side.
At a national level, we’re working with the Gates Foundation to integrate TigerBeetle into their non-profit central bank switch that will be powering Rwanda’s National Digital Payments System 2.0 later this year [1].
At an enterprise level, TigerBeetle already powers customers processing 100M+ transactions per month in production, and we recently signed our first $2B fintech unicorn in Europe with a few more in the US about to close. Because of the move to realtime transaction processing around the world [2] there’s been quite a bit of interest from companies wanting to move to TigerBeetle for more performance.
Finally, to your question, some of the founders of Clear Street, a fairly large brokerage on Wall Street have since invested [3] in TigerBeetle.
[1] https://mojaloop.io/how-mojaloop-enables-rndps-2-0-ekash/
[2] https://tigerbeetle.com/blog/2024-07-23-rediscovering-transa...
"Invested" in terms of "giving you money" or in terms of "Now uses the database themselves"? I read it as the first, but I think the question is about usage, not investments.
- The investment in TigerBeetle was done personally, not through Clear Street.
- I'm no longer actively involved day-to-day as CTO at Clear Street, but while I was, TigerBeetle was a solution we very much had in mind as our volumes were increasing.
That said, roadmaps change, priorities shift, etc. If TigerBeetle existed when we started Clear Street, I very much would have used it, and saved me from many headaches.
Form the developer perspective, the big thing is that we don't have any dependencies, so updating compiler for us is just a small amount of work once in a while, and not your typical ecosystem-wide coordination problem. Otherwise, Zig's pretty much "finished" for our use-case, it more or less just works.
But, no. On the commercial side, I don't think we've had one conversation with a prospect or CTO or engineering team where they were concerned that we picked a systems language for the next thirty years. And while Zig is a beautiful, perfect replacement for C, I think the real reason the question has never come up, is that our customers come to us instead of us to them. We're not trying to convince anyone. They're already appreciating the extensive end-to-end testing we do on everything we ship.
However, I should emphasize again, that given all the assertions, fuzzing and DST we do, Zig's quality can't be overstated. It holds up.
We don't publish the pre-compiled libs yet, but it will be available soon when we stabilize the API. For now, it can be built locally from source.
Example of using the C client: https://github.com/tigerbeetle/tigerbeetle/blob/main/src/cli...
Or possibly my mental model of how physical disks and the driver stack behave these days is outdated.
For more on our open source thinking and how this is orthogonal to business model (and product!), see our interview with the Changelog: https://m.youtube.com/watch?v=Yr8Y2EYnxJs
1. AWS will take your initial and ongoing investment in the implementation but they don't have to share theirs with you. Specifically, they will take your improvements but their own improvements (say some performance optimizations) they can keep to themselves. It's good business sense if it allows them to further differentiate their "improved" offering from your "vanilla" service.
2. Competing on the the interface in this case really means competing on related services like management, etc. So your thesis is that you will provide a better/cheaper managed service than AWS. Even if that's true (a big if), most of the time the decision which service to use will have little to do with technical merit. I.e. we already use AWS, have SLA painfully negotiated, get volume discounts, etc. Do we really want to go through all of this with another vendor just for one extra service.
Just a couple of thoughts that will hopefully help you sharpen your thesis.
In practice all I've seen from AWS is just to add integrations with their internal orchestrators and not much else. Back when I was at Redis Labs, AWS added TLS support to Redis and was dying to get that upstreamed (so that they wouldn't have to maintain the patch), except that as far as I understood nobody upstream wanted that code. In other words, hypothetical improvements by AWS (and other Clouds) are extremely overrated. When it comes to tigerbeetle, I would put the chance that they introduce bugs and vulnerabilities much higher than the possibility they add any meaningful improvement over what the actual experts (the tigrebeetle team) have already done.
> Do we really want to go through all of this with another vendor just for one extra service.
That's a great point, and in fact I've seen AWS purposefully offer insane (in Europe maybe we would say anti-competitive) discounts precisely to prevent Redis Labs from gaining market share. I'm sure they will try the same with TB once it becomes mainstream enough. What TB has that Redis doesn't have is the fact that it's a database designed for truly mission-critical stuff (i.e. counting the money) and maybe customers will be willing to go through the extra motions to ensure they get the best service they can (assuming TB will be able to provide that).
Interesting, in a recent thread (I think it was about Redis going back open source) an AWS employer was bragging about substantial concurrency optimizations they implemented in Valkey. At the time I thought it could have been a great differentiator to keep proprietary but perhaps they decide to sacrifice it to help make sure Valkey takes over the Redis midshare.
We also don't expect (or need) anyone to contribute improvements upstream to us. That's open source!
Finally, open source is not the same thing as product. There are thousands of companies around the world who make high quality products that people pay for. TigerBeetle is no different.
Yes, TigerBeetle specializes only for transaction processing (OLTP). It’s not a general-purpose (OLGP) DBMS.
That said, we have customers from energy to gaming, and of course fintech.
That's like saying that rice noodles are no good for making risotto. At the core they are both rice...
Another paradigm is OLAP, in which aggregation of large datasets is the principal concern.