TigerBeetle Core System Architecture: Deconstructing Performance Engineering
ixuvo.com
ixuvo.com
The two "sources" are fake links that lead to 404s
* A Database Without Dynamic Memory Allocation[1]
* A Programmer-Friendly I/O Abstraction Over io_uring and kqueue[2]
* Writing High-Performance Clients for Tigerbeetle[3]
[1]: https://tigerbeetle.com/blog/2022-10-12-a-database-without-d...
[2]: https://tigerbeetle.com/blog/2022-11-23-a-friendly-abstracti...
[3]: https://tigerbeetle.com/blog/2023-02-21-writing-high-perform...
Sort of like a new paradigm where opinionated custom databases could be created with arbitrary entry logic built on this stack.
Seriously Can't wait to see this.
Congrats on launching [1] Tigerbeetle Cloud. Should have submitted that as well but I thought this was more interesting. May be another time.
That’s the dream. To serve the world’s transactions and data (with a whole lotta beetles!). We’re working to make it reality.
Appreciate the congrats! You make today a double whammy! :P
I'm pretty sure you can't just do a precise 128 byte align, if one of the element is an image blob or varchar(1000).
And isn't there a benefit for small allocations on advanced memory allocations that you can't leverage if all is working in big page allocations? Do you implement memory allocations from scratch or leveraging existing allocator on top of these memory blocks strategy somehow?
For example, if you take a look at our LSM compaction, regardless of the table size, we compact at the 512 KiB block granularity, and everything is streaming.
The same principle applies everywhere.
In our experience writing TigerStyle (and for all our internal code and tooling, not only TB as DBMS), we’ve never had a scenario where static allocation was not applicable or didn’t produce a better design.
You also tend to become more memory efficient, not less. Again, since you’re streaming. (You’re not allocating a massive buffer, just because a file is multi-GiB.)
The runtime allocation is type-aware, workload-aware, and schedule-aware. The last is most important. There are two places that can act as a sink for heavy memory demands: storage (i.e. paging to disk) and network (e.g. streaming results). These have their own limitations because I/O bandwidth is finite. Effectively, your allocation rate is equivalent to available I/O bandwidth.
The most powerful lever you have to manage this is total control of the schedule. Demands on memory are created by a set of operations or queries visible to the software. The scheduler doesn't incrementally execute these operations randomly, it continuously selects execution based on the availability of memory or bandwidth to absorb the allocation demand of the operation. The scheduler has the ability to control the allocation rate to instantaneously match availability.
This is essentially the very old idea of "optical buffering" -- treating fiber optic cables as RAM -- taken to its logical architectural conclusion.
The caveat is that this requires direct I/O in userspace, which places limits on software architecture. But if you care about performance, you'd be using this type of software architecture regardless.
EDIT: I guess part of the question is about the problems with clusters with >130ms latency and if there are challenges you consider easy.
That said, network latency usually follows a distribution. For example, the median might be 130 ms while p99 is 200 ms. So one important goal is to avoid being affected by the high-latency tail.
In consensus and replication systems such as TigerBeetle, you can reduce the impact quite a bit by taking advantage of the fact that you only need a quorum. We have six replicas, and under normal operation we only need acknowledgements from three (including the primary, since we use flexible quorums). That means the primary only has to wait for the two fastest replicas to respond. This is very effective at reducing tail latency.
Then, to get as close as possible to speed-of-light latency, you want to avoid adding unnecessary latency inside the system itself. We've done quite a few algorithmic optimizations there over the past year. For example, introducing radix sort and tournament trees to make CPU processing more efficient.
The basic idea is pretty simple. In VSR, there are two main phases:
1. Leader election
2. Normal replication / request processing
Before Heidi Howard’s insight, these two phases typically used the same quorum size - for example, 4 out of 6 replicas.
The key observation was that the two phases can actually use different quorum sizes, as long as the relevant quorums still intersect.
With 6 replicas, we could use a quorum of 4 for view change and a quorum of 3 for normal processing, because 4+3>6. This guarantees that every view-change quorum intersects every processing quorum. Therefore, if an operation was committed by a processing quorum, at least one replica participating in the subsequent view change knows about that operation. Combined with the protocol's view-change/log-selection rules, this ensures that committed operations are preserved when the new leader takes over.
If this interests you, Heidi gave a talk about this at systems distributed: https://youtu.be/P0cAG-RM1_c which will be released soon.
What are some interesting problems or things you can think of to work on that would give someone new a nice amount of exposure to this kind of programming?
I'd check out tigerstyle.dev, pick up Zig, and then make an HTTP server or file format parser. Those are great ways to learn and experience this kind of programming. At some point, you start to realize that it's just easier to build API services in this way.
But for sure, you can learn so much in JVM-based languages. They make you appreciate low-level techniques all the more!
When tiger beetle becomes pluggable to different use cases, how should I think about "do I want tiger beetle?"
We still write, read (and have an independent engineer review) each line of code by hand.
We go faster like that, but, most of all, it’s the guarantee we make to our users, also to continue to invest in our own understanding, because second order that’s valuable for the kind of high performance safety work we do.
Long term, I’m sure LLMs will improve, but right now they’re just not there.
Can someone explain why single-threaded execution loop is more aligned with the physical realities of modern hardware ?
It really depends on the problem space. For example, many OLAP workloads (analytical) contain large amounts of parallelizable work, then multi-core execution is absolutely the way to go. That also aligns well with the direction CPU technology is taking, with core counts continuing to increase.
For the transactional workloads we see at TigerBeetle, and in other transactional systems I've worked with, the picture is quite different. We see a lot of read-modify-write operations combined with a power-law distribution of the data.
Take a simple banking example: some accounts, such as those belonging to large online retailers, see much more activity than the average individual account. You might have 80 - 90% of transfers touching a relatively small number of these hot accounts.
Operations on the same account must be serialized to preserve correctness. That means this part of the workload cannot be meaningfully parallelized. In fact, attempting to parallelize it can make performance worse because of lock contention and coordination overhead, something the "Universal Scalability Law" captures quite well (but is also easy to test out yourself with a simple experiment).
Instead, we focus on batched execution. We carefully structure execution to make effective use of CPU caches and efficient algorithms, so that a single batch can be processed extremely efficiently without any coordination. Batch execution also allows to amortize I/O and replication.
That being said, there are areas where we could use multi-threading (e.g. compaction) that are not on the hot execution path.
"Multiprocessing" (multiple CPU cores independently executing and only able to coordinate via some message-passing system -- a definition which encompasses both multi-core CPUs and horizontally scaled distributed systems, contrasting slightly with the normal definition) is challenging for a few reasons.
Firstly, the details are an open math question, but I'll blindly state that some problems aren't amenable to parallelization. I.e., no algorithm can meaningfully improve performance via parallelization no matter the implementation. Think through how you would more quickly compute hash(hash(hash(hash(...)))) for example. The serial dependency makes things challenging. That isn't too dissimilar to the problem TB faces.
Secondly, message passing is expensive. If the only way two CPU cores can coordinate is through a multi-level cache, at best you're incurring ~tens of nanoseconds of latency per message. Contrast that with a base rate of 512 bytes processed per nanosecond with enough attention to detail on typical modern server hardware (4 pipelined AVX512 instructions at 2GHz). Messages are several orders of magnitude worse than your normal work, so if you need very many of them then you're hosed from a performance perspective (worse with longer delays, like networked computers). Even very parallelizable problems at an abstract level can suffer performance losses by trying to add even one extra core. This blog post [0] doesn't perfectly capture the idea, but it's close (and a fun read regardless).
Thirdly, message passing is an insanely complicated programming abstraction to reason about. My first two points were more about what TB was saying -- realities of modern hardware -- but the programmer experience is important too (even if you don't believe that post-2020, the LLM experience doesn't differ much from the human experience; bad code begets more bad code, slowly). The core mechanism for correctness in most software is being able to reason about "this thing is true, therefore that thing is true" and iterating. You rely on invariants like "this is sorted" to build other working theories. The invariants in multiprocessing code are much more nebulous and less amenable to accidental discovery, also less amenable to being able to build or compose them into other stronger invariants as you add code. The main reason for that is that you know almost nothing about the relative order of those messages with respect to each processer's view of which instructions happened when (and for purely multi-core "multiprocessing" the story is even worse; while my description of message-passing being the core primitive is correct, that's not what's exposed to you as a programmer, and different memory models can have even weirder interleavings than your code would naively suggest -- i.e., your code is being decomposed into smaller subunits than even a single assembly instruction, and the message passing happens at that level). The combinatorial explosion (an exponential explosion really, but big numbers either way) of states you might be interacting with makes it very difficult to understand _anything_ about the system you're examining. That's why you see a handful of primitives used over and over -- if you can decompose your problem into a parallel map plus an associative reduce then you can probably figure out some way to make it better through parallelization (not always, especially if the framework is too generic, see the linked blog post [0] if you weren't enticed to read it previously). If you can't decompose it into know primitives then it's an open research problem every time.
The crux of that third point (and we could definitely add more explanation and additional problems) is that there's a huge cost to multiprocessing. You have to be buying something substantial to even want to reach for it, else you have to be in one of the "easy" problem spaces where somebody else has done the hard work (e.g., stateless webservers).
TB isn't that. Their whole raison d'être is state management, and not in a way that's easily amenable to parallelization.
[0] https://adamdrake.com/command-line-tools-can-be-235x-faster-...
So batching requests is always something I think should increase performance by a lot, but most server implementations make this pretty difficult, but the thing I struggle the most to understand is how to keep the latency down if you have multiple clients request all batched together? The total amount of latency for all clients is always the latency for the slowest.
But if your application then creates another transfer against the client, and another, while the first request is inflight, then the client will autobatch under the hood and send these off as a batch when the first request returns.
You get this sweetspot then between latency and throughput. And your latency is not spiking as your load increases, since your throughput is now able to keep up.
It is trivial to ensure this doesn't happen at compile-time. If you are doing DMA as in the article these guarantees are required.