Having built a ticketing system that sold some Oasis level concerts there's a few misconceptions here:
Selling an event out takes a long time to do frequently because tickets are VERY frequently not purchased--they're just reserved and then they fall back into open seating. This is done by true fans, but also frequently by bots run by professional brokers or amateur resellers. And Cloudflare and every other state of the art bot detection platform doesn't detect them. Hell, some of the bots are built on Cloudflare workers themselves in my experience...
So whatever velocity you achieve in the lab--in the real world you'll do a fraction of it when it comes to actual purchases. That depends upon the event really. Events that fly under the radar may get you a higher actual conversion rate.
Also, an act like Oasis is going to have a lot of reserved seating. Running through algorithms to find contiguous seats is going to be tougher than this example and it's difficult to parallelize if you're truly giving the next person in the queue the actual best seats remaining.
There are many other business rules that accrue after years of features to win Oasis like business unfortunately that will result in more DB calls and add contention.
> Selling an event out takes a long time to do frequently because tickets are VERY frequently not purchased--they're just reserved and then they fall back into open seating.
TigerBeetle actually includes native support for "two phase pending transfers" out of the box, to make it easy to coordinate with third party payment systems while users have inventory in their cart:
> Also, an act like Oasis is going to have a lot of reserved seating. Running through algorithms to find contiguous seats is going to be tougher than this example and it's difficult to parallelize if you're truly giving the next person in the queue the actual best seats remaining.
It's actually not that hard (and probably easier) to express this in TigerBeetle using transfers with deterministic IDs. For example, you could check (and reserve) up to 8K contiguous seats in a single query to TigerBeetle, with a P100 less than 100ms.
> There are many other business rules that accrue after years of features to win Oasis like business unfortunately that will result in more DB calls and add contention.
As you move "the data to the code" in interactive transactions with multiple queries, to process more and more business rules, you're holding row locks across the network. TigerBeetle's design inverts this, to move "the code to the data" in declarative queries, to let the DBMS enforce the transactional business rules directly in the database, with a rich set of debit/credit primitives and audit trail.
If only. But you also need to fix the internal concurrency control of the DBMS storage engine. TB here is very different to PG.
For example, if you have 8K transactions through 2 accounts, a naive system might read the 2 accounts, update their balances, then write the 2 accounts… for all 8K (!) transactions.
Whereas TB does vectorized concurrency control: read the 2 accounts, update them 8K times, write the 2 accounts.
This is why stored procedures only get you typically about a 10x win, you don’t see the same 1000x as with TB, especially at power law contention.
Huge fan of what tiger beatle promotes. Even in simple system/projects batching and reducing contention can be massive win. Batching + single application writer alone in something like sqlite can get you to pretty ridiculous inserts/updates per second (although transactions become at the batch level).
I sometimes wonder how many fewer servers we would need if the aproaches promoted by Tiger Style were more widespread.
What datasteucture does Tiger Beatle use for it's client? I'm assuming its multi writer single reader. I've always wondered what the best choice is there. A reverse LMAX disruptor (multiple producers single consumer).
Agree with the above, we built and run a ticketing platform, the actual transaction of purchasing the ticket at the final step in the funnel is not the bottleneck.
The shopping process and queuing process puts considerably more load on our systems than the final purchase transaction, which ultimately is constrained by the size of the venue, which we can control by managing the queue throughput.
Even with a queue system in place, you inevitably end up with the thundering heard problem when ticket sales open, as a large majority of users will refresh their browsers regardless of instructions to the contrary
You would use TigerBeetle for everything: not only the final purchase transaction, but the shopping cart process, inventory management and queuing/reserving.
In other words, to count not only the money changing hands, but also the corresponding goods/services being exchanged.
These are all transactions: goods/services and the corresponding money.
Does that mean that there is some smoke and mirrors when, eg Taylor Swift, says they sold out the concert in minutes? Or are the mega acts truly that high demand?
You can get the seats into "baskets" (reserved) in minutes. In my experience they will not sell out for some time as they usually keep dropping back into inventory. "Sold Out" is a matter of opinion. There are usually lots of single seats left sometimes for weeks or months. The promoter decides when to label the event as "sold out".
I recently did performance testing of Tigerbeetle for a financial transactions company. The key thing to understand about Tigerbeetle's speed is that it achieves very high speeds through batching transactions.
----
In our testing:
For batch transactions, Tigerbeetle delivered truly impressive speeds: ~250,000 writes/sec.
For processing transactions one-by-one individually, we found a large slowdown: ~105 writes/sec.
This is much slower than PostgreSQL, which row updates at ~5495 sec. (However, in practice PostgreSQL row updates will be way lower in real world OLTP workloads due to hot fee accounts and aggregate accounts for sub-accounts.)
One way to keep those faster speeds in Tigerbeetle for real-time workloads is microbatching incoming real-time transactions to Tigerbeetle at an interval of every second or lower, to take advantage of Tigerbeetle's blazing fast batch processing speeds. Nonetheless, this remains an important caveat to understand about its speed.
> One way to keep those faster speeds in Tigerbeetle for real-time workloads is microbatching incoming real-time transactions to Tigerbeetle at an interval of every second or lower, to take advantage of Tigerbeetle's blazing fast batch processing speeds.
We don’t recommend artificially holding transfers just for batching purposes.
René actually had to implement a batching worker API to work around a limitation in Python’s FastAPI, which handled requests per process, and he’s been very clear in suggesting that such would be better reimplemented in Go.
Unlike most connection-oriented database clients, the TigerBeetle client doesn’t use a connection pool, because there’s no concept of a “connection” in TigerBeetle’s VSR protocol.
This means that, although you can create multiple client instances, in practice less is better. You should have a single long-lived client instance per process, shared across tasks, coroutines, or threads (think of a web server handling many concurrent requests).
In such a scenario, the client can efficiently pack multiple events into the same request, while your application logic focuses solely on business-event-oriented chains of transfers. Typically, each business event involves only a handful of transfers, which isn't a problem of underutilization, as they'll be submitted together with other concurrent events as soon as possible.
However, if you’re dealing with a non-concurrent workload, for example, a batch process that bills thousands of customers for their monthly invoices, then you can simply submit all transfers at once.
> For processing transactions one-by-one individually
If you're artificially restricting the load going into TigerBeetle, by sending transactions in one-by-one individually, then I think predictable latency (and not TPS) would be a better metric.
For example, TB's multi-region/multi-AZ fault-tolerance will work around gray failure (fail slow of hardware, as opposed to fail stop) in your network links or SSDs. You're also getting significantly stronger durability guarantees with TB [0][1].
It sounds like you were benchmarking on EBS? We recommend NVMe. We have customers running extremely tight 1 second SLAs, seeing microsecond latencies, even for one at a time workloads. Before TB, they were bottlenecking on PG. After TB, they saturated their central bank limit.
I would also be curious to what scale you tested? We test TB to literally 100 billion transactions. It's going to be incredibly hard to replicate that with PG's storage engine. PG is a great string DBMS but it's simply not optimized for integers the way TB is. Granted, your scale likely won't require it, but if you're comparing TPS then you should at least compare sustained scale.
There's also the safety factor of trying to reimplement TB's debit/credit primitives over PG to consider. Rolling it yourself. For example, did you change PG's defaults away from Read-Committed to Serializable and enable checksums in your benchmarks? (PG's checksums, even if you enable them, are still not going to protect you from misdirected I/O like the recent XFS bug.) Even the business logic is deceptively hard, there are thousands of lines of complicated state machine code, and we've invested literally millions into testing and audits.
Finally, it's important that your architecture as a whole, the gateways around TB, designs for concurrency first class, and isn't "one at a time", or TigerBeetle is probably not going to be your bottleneck.
We didn't observe any automatic batching when testing Tigerbeetle with their Go client. I think we initiated a new Go client for every new transaction when benchmarking, which is typically how one uses such a client in app code. This follows with our other complaint: it handles so little you will have to roll a lot of custom logic around it to batch realtime transactions quickly.
I'm a bit worried you think instantiating a new client for every request is common practice. If you did that to Postgres or MySQL clients, you would also have degradation in performance.
PHP has created mysqli or PDO to deal with this specifically because of the known issues of it being expensive to recreate client connects per request
We shared the code with the Tigerbeetle team (who were very nice and responsive btw), and they didn't raise any issues with the script we wrote of their Tigerbeetle client. They did have many comments about the real-world performance of PostgreSQL in comparison, which is fair.
Thanks for the code and clarification. I'm surprised the TB team didn't pick it up, but your individual transfer test is a pretty poor representation. All you are testing there is how many batches you can complete per second, giving no time for the actual client to batch the transfers. This is because when you call createTransfer in GO, that will synchronously block.
For example, it is as if you created an HTTP server that only allows one concurrent request. Or having a queue where only 1 worker will ever do work. Is that your workload? Because I'm not sure I know of many workloads that are completely sync with only 1 worker.
To get a better representation for individual_transfers, I would use a waitgroup
var wg sync.WaitGroup
var mu sync.Mutex
completedCount := 0
for i := 0; i < len(transfers); i++ {
wg.Add(1)
go func(index int, transfer Transfer) {
defer wg.Done()
res, _ := client.CreateTransfers([]Transfer{transfer})
for _, err := range res {
if err.Result != 0 {
log.Printf("Error creating transfer %d: %s", err.Index, err.Result)
}
}
mu.Lock()
completedCount++
if completedCount%100 == 0 {
fmt.Printf("%d\n", completedCount)
}
mu.Unlock()
}(i, transfers[i])
}
wg.Wait()
fmt.Printf("All %d transfers completed\n", len(transfers))
This will actually allow the client to batch the request internally and be more representative of the workloads you would get. Note, the above is not the same as doing the batching manually yourself. You could call createTransfer concurrently the client in multiple call sites. That would still auto batch them
Yeah it was back in February in your community Slack, I did receive a fairly thorough response from you and others about it. However then there were no technical critiques of the Go benchmarking code, just how our PostgreSQL comparison would fall short in real OLTP workloads (which is fair).
Interesting, I thought I had heard that this is automatically done, but I guess it's only through concurrent tasks/threads. It is still necessary to batch in application code.
But nonetheless, it seems weird to test it with singular queries, because Tigerbeetle's whole point is shoving 8,189 items into the DB as fast as possible. So if you populate that buffer with only one item your're throwing away all that space and efficiency.
We certainly are losing that efficiency, but this is typically how real-time transactions work. You write real-time endpoints to send off transactions as they come in. Needing to roll more than that is a major introduction of complexity.
We concluded where Tigerbeetle really shines is if you're a large entity like a central bank or corporation sending massive transaction files between entities. Tigerbeetle is amazing for moving large numbers of batch transactions at once.
We found other quirks with Tigerbeetle that made it difficult as a drop-in replacement for handling transactions in PostgreSQL. E.g. Tigerbeetle's primary ID key isn't UUIDv7 or ULID, it's a custom id they engineered for performance. The max metadata you can save on a transaction is a 128-bit unsigned integer on the user_data_128 field. While this lets them achieve lightning fast batch transaction processing benchmarks, the database allows for the saving of so little metadata you risk getting bottlenecked by all the attributes you'll need to wrap around the transaction in PostgreSQL to make it work in a real application.
> you risk getting bottlenecked by all the attributes you'll need to wrap around the transaction in PostgreSQL to make it work in a real application.
The performance killer is contention, not writing any associated KV data—KV stores scale well!
But you do need to preserve a clean separation of concerns in your architecture. Strings in your general-purpose DBMS as "system of reference" (control plane). Integers in your transaction processing DBMS as "system of record" (data plane).
Dominik Tornow wrote a great blog post on how to get this right (and let us know if our team can accelerate you on this!):
> We didn't observe any automatic batching when testing Tigerbeetle with their Go client.
This is not accurate. All TigerBeetle's clients also auto batch under the hood, which you can verify from the docs [0] and the source [1], provided your application has at least some concurrency.
> I think we initiated a new Go client for every new transaction when benchmarking
The docs are careful to warn that you shouldn't be throwing away your client like this after each request:
The TigerBeetle client should be shared across threads (or tasks, depending on your paradigm), since it automatically groups together batches of small sizes into one request. Since TigerBeetle clients can have at most one in-flight request, the client accumulates smaller batches together while waiting for a reply to the last request.
Again, I would double check that your architecture is not accidentally serializing everything. You should be running multiple gateways and they should each be able to handle concurrent user requests. The gold standard to aim for here is a stateless layer of API servers around TigerBeetle, and then you should be able to push pretty good load.
We didn't rule out using Tigerbeetle, but the drop in non-batch performance was disappointing and a reason we haven't prioritised switching our transaction ledger from PostgreSQL to Tigerbeetle.
There was also poor Ruby support for Tigerbeetle at the time, but that has improved recently and there is now a (3rd party) Ruby client: https://github.com/antstorm/tigerbeetle-ruby/
I think the drop in non-batch performance was more a function of the PoC than of TB. Would love to see what our team could do for you here! Feel free to reach out to peter@tigerbeetle.com
It seems to me that, in practice, you'd want the "LiveBatcher" to have some durability as well. Is there a scenario where a customer could lose their place because of a horribly timed server shutdown, where those transfers hadn't even been sent to TigerBeetle as pending yet? Or am I misunderstanding the architecture here?
Edit: Yes, I think I misunderstood something here. The user wouldn't even see their request as having returned a valid "pending" ticket sale since the batcher would be active as the request is active. The request won't return until its own transfer had been sent off to TigerBeetle as pending.
The short answer is that we tried, back in 2020, while working on a central bank payment switch by the Gates Foundation. We found we were hitting the limits of Amdahl's Law, given Postgres' concurrency control with row locks held across the network as well as internal I/O, leading to the design of TigerBeetle. To specialize not for general purpose but only for transaction processing.
On the one hand, yes, you could use a general purpose string database to count/move integers, up to a certain scale. But a specialized integer database like TigerBeetle can take you further. It's the same reason, that yes, you could use Postgres as object storage or as queue, or you could use S3 and Kafka and get separation of concerns in your architecture.
I did a talk diving into all this recently, looking at the power law, OLTP contention, and how this interacts with Amdahl's Law and Postgres and TigerBeetle: https://www.youtube.com/watch?v=yKgfk8lTQuE
i am not an exact expert on the limitation you claim to have encountered on postgresql but perhaps someone with more postgresql expertise can chime in on this comment and give some insight
For updating a single resource where the order of updates matters the best throughput one can hope for is the inverse of locking duration. Typical postgres using applications follow the pattern where a transaction involves multiple round trips between the application and the database to make decisions in the code running on the application server.
But this pattern is not required by PostgreSQL, it's possible to run arbitrarily complex transactions all on server side using more complex query patterns and/or stored procedures. In this case the locking time will be mainly determined by time-to-durability. Which, depending on infrastructure specifics, might be one or two orders of magnitude faster. Or in case of fast networks and slow disks, it might not have a huge effect.
One can also use batching in PostgreSQL to update the resource multiple times for each durability cycle. This will require some extra care from application writer to avoid getting totally bogged down by deadlocks/serializability conflicts.
What will absolutely kill you on PostgreSQL is high contention and repeatable read and higher isolation levels. PostgreSQL handles update conflicts with optimistic concurrency control, and high contention totally invalidates all of that optimism. So you need to be clever enough to achieve necessary correctness guarantees with read committed and the funky semantics it has for update visibility. Or use some external locking to get rid of contention in the database. The option for pessimistic locking would be very helpful for these workloads.
What would also help is a different kind of optimism, that would remove durability requirement from lock hold time, which would then result in readers having to wait for durability. Postgres can do tens of thousands of contended updates per second with this model. See the Eventual Durability paper for details.
Yes, that's why I would expect it to smoke Postgres here, in process is orders of magnitude faster. Do you really need concurrency here when you can do 10-100k+ inserts per second?
Also surprised. My yardstick was this post which showed SQLite beating Postgres in a Django app. Benchmarking is hard, and the author said the Postgres results were not tuned to the same degree as SQLite, so buyer beware.
https://blog.pecar.me/django-sqlite-benchmark
Selling an event out takes a long time to do frequently because tickets are VERY frequently not purchased--they're just reserved and then they fall back into open seating. This is done by true fans, but also frequently by bots run by professional brokers or amateur resellers. And Cloudflare and every other state of the art bot detection platform doesn't detect them. Hell, some of the bots are built on Cloudflare workers themselves in my experience...
So whatever velocity you achieve in the lab--in the real world you'll do a fraction of it when it comes to actual purchases. That depends upon the event really. Events that fly under the radar may get you a higher actual conversion rate.
Also, an act like Oasis is going to have a lot of reserved seating. Running through algorithms to find contiguous seats is going to be tougher than this example and it's difficult to parallelize if you're truly giving the next person in the queue the actual best seats remaining.
There are many other business rules that accrue after years of features to win Oasis like business unfortunately that will result in more DB calls and add contention.
reply