Switching DBs is very invasive. Switching your application base even more so. As long as you're happy happy with PocketBase and there's nothing you're critically missing, the answer will almost certainly be: no. Is there anything you're looking for?
64 karma · joined October 31, 2024
I don't think the referenced `Store` is used in the SQLite access path and PocketBase uses SQLite in WAL mode. Besides, in cases where the `Store` is used, it uses a read/write lock to allow for concurrent reads.
Naively, I would expect SQLite to be able to delete tens-of-thousands (or even hundreds) of records per seconds, since it's simply appending deletions to the WAL.
Could you elaborate a bit more on your scaling concerns? You can certainly have lock-congestion with SQLite. That, said Postgres - while awesome - isn't the most horizontally scalable beast.
It's maybe more like you point out: realtime in the OS context vs realtime in an event processing context. The latter is certainly not defined as strictly and often just means push-based. It has been a popular moniker, e.g. in kafka-land, for a while. I'm not sure it intrinsically takes away from the OS context - it doesn't need to be a deep dish pizza situation.
If you're comparing in-process SQLite to talking to SQLite over HTTP you'll probably get a small penalty for any language. When co-located on the same machine, you can probably expect something like ~5ms for JS (just for the event-loop to do the IO and getting back to you).
However, if you have multiple processes reading and writing from the same DB it may actually aid latency due to congestion.
I ran some benchmarks (TrailBase author here and big fan of PocketBase): https://trailbase.io/reference/benchmarks#insertion-benchmar..., where you can see, e.g. the single-process drizzle (i.e. JS with in-process SQLite) performance vs over-HTTP for TrailBase. PocketBase should be similar when not fully loaded. There's also some concrete latency percentile numbers when at full-tilt: https://trailbase.io/reference/benchmarks#read-and-write-lat.... On my machine you can expect p50 to be around 15-20ms.
I'm a bit surprised on the mobile comment, since last I checked PB's UI wasn't responsive, i.e. you had to scroll a lot horizontally on mobile. Despite it's missing polish, I tried to make TB's at least work well on mobile. Could you elaborate? - thanks
Agreed.
> It sounds like your v8 worker threads are mixing read and write work so you are running the query in another sqlite thread pool to preven> In WAL mode SQLite is very good at supporting parallel reads from multiple threads. It should only block for a long time when writing (since writes require an exclusive lock.)
Agreed.
> It sounds like your v8 worker threads are mixing read and write work so you are running the query in another sqlite thread pool to prevent writes from blocking reads.
The v8 isolates run whatever you as a TrailBase user feed them. I would certainly expect writes to be a common occurrence.
> Given the additional costs of cross-thread communication I would be surprised if this approach maximizes throughput under highly concurrent loads compared to segregating write requests into a dedicated thread and running read queries synchronously from within their threadpool with a single task per thread.
Ultimately, it will depend a lot on the ratios. If you have mostly reads and the occasional write you're probably right. I did spend a bit of time exploring different execution models: https://github.com/ignatz/libsql_bench in case you're interested. There's also some prior works from the folks GIL'ed languages (especially ruby) around how to wrangle write congestion for multi-process workloads. Sadly for them, they don't have inter-thread comms in their arsenal :)
One big unknown for me is, how you'd clearly separate reads from writes. As far as I can think, you'd have to rely on users to pick the right sync or async funnel. Which may be ok at least for simple queries.
FWIW, the thing or elephant that bothered me more than inter-thread comms is the opportunity cost of not running reads in parallel. Then at the same time, the current setup does seem to manage to saturate the machines I've run on. Very high core-count machines would probably be a different story. It will certainly also depend on how much actual other work the server has to do, i.e. is it just a glorified SQLite accessor? I certainly would love to further optimize that aspect. You seem very well informed so I'd love to hear your thoughts. Hit me up, if you'd like to chat more.t writes from blocking reads.
The v8 isolates run whatever you as a TrailBase user feed them. I would certainly expect writes to be a common occurrence.
> Given the additional costs of cross-thread communication I would be surprised if this approach maximizes throughput under highly concurrent loads compared to segregating write requests into a dedicated thread and running read queries synchronously from within their threadpool with a single task per thread.
Ultimately, it will depend a lot on the ratios. If you have mostly reads and the occasional write you're probably right. I did spend a bit of time exploring different execution models: https://github.com/ignatz/libsql_bench in case you're interested. There's also some prior works from the folks GIL'ed languages (especially ruby) around how to wrangle write congestion for multi-process workloads. Sadly for them, they don't have inter-thread comms in their arsenal :)
One big unknown for me is, how you'd clearly separate reads from writes. As far as I can think, you'd have to rely on users to pick the right sync or async funnel. Which may be ok at least for simple queries.
FWIW, the thing or elephant that bothered me more than inter-thread comms is the opportunity cost of not running reads in parallel. Then at the same time, the current setup does seem to manage to saturate the machines I've run on. Very high core-count machines may be a different story. It will certainly also depend on how much actual other work the server has to do, i.e. is it just a glorified SQLite accessor? I certainly would love to further optimize that aspect. You seem very well informed so I'd love to hear your thoughts. Hit me up, if you're willing to chat more.
FWIW, it never felt like a dispute and very much agree with your suggestion. I'm also just trying to do a decent enough job, both with the benchmarks and TrailBase itself. Either way, my offer to keep an open channel stands in case you want to share experiences or are in desperate need for a beverage :)
> I feel I prefer to have a locked-down database, and implement everything "backend-side" with a kind of "admin API" which has access to everything, and checks user roles in the backend, it feels cleaner to me, is that also possible?
is different from what FireBase or TrailBase does?
Are you saying that you'd prefer to run your own backend binary (as opposed to running in an integrated runtime), do your own ACL checking, and have more of a free-form SQL-like API with the DB layer?
FWIW, I did this locally (also happy to fork PB and check it in, certainly aids transparency). I've already updated the numbers in the benchmark doc but will continue to try squeeze more out of it.
As to 4x, my concurrency levels are significantly lower which could certainly explain it. Is PB juggling multiple connections increasing write lock congestion or are you serializing access, e.g. via a worker thread? Also happy to chat more (feel free to send me an email), I certainly want PB to have the best possible representation and maybe we can even speed things up on both sides
Anyway, w/o fog I managed to run with mattn/go-sqlite3. I'm not sure this is expected but it didn't seem to make much of a difference with my setup (For transparency, I do recently have some issues with repeatability likely due to btrfs). I'm certainly not seeing 4x but around 30-35%, which is still very impressive!
That said, you absolutely can generate code from the DB schema already. There are `/examples` (just none in C# yet).
I'm optimistic that C# will become increasingly important with respect to cross-platform mobile/desktop development and thus receive higher priority treatment in general.
What you're saying makes a lot of sense. SQLite is sync and if you're program is alone accessing SQLite doing a single task, going sync is the way. If you're doing a lot of parallel work, both your JS event loop interleaving many tasks and several event loops accessing SQLite in parallel you have to make trade-offs. Specifically, `conn.query` may block for a long time w/o doing any work. Depending on your use-case it may or may not be ok to block the event-loop that entire period. TrailBase's setup is optimized to maximize throughput under highly concurrent loads, rather than minimizing latency in single-threaded workloads. That's not to say, TrailBase isn't quick. It's pretty low-latency even under load. However, if that's all you're after you're probably better off with better-sqlite3 or dropping down to C :).
Does that make sense?
Did I misunderstand? Was there something that thew you off? - always keen to improve
It would have certainly been simpler to just plot the overall runtime (width of the graph), I did think that it was quite interesting that PB's goja integration takes a while before utilizing the entire machine.
I still wanna get the mattn/go-sqlite3 driver to work and it's getting a bit late here for writing coherent text... I'll update the benchmarks ASAP
Naively, I would argue that being idiomatic in the respective ecosystem is more important than perfect consistency. Only few users will likely use 2 or more languages and probably even then there's a balance to be struck.
Hey thanks for chiming it. Huge fan of PocketBase, has been a major inspiration :applause:. For anyone driving by, certainly a more mature product.
- Based on your benchmarks repo it looks like that the tests were done against PocketBase < v0.23 but note that PocketBase v0.23+ (especially with the Create API rule dry submit removal in v0.24+) has introduced significant changes and performance improvements - ~4x times in high concurrent scenarios in our own benchmarks[0] (if you want to retest it note that the CGO driver is no longer loaded by default and will have to be registered manually; see the example "db_cgo.go" in the PocketBase benchmarks repo or in the "Custom SQLite driver" docs[1]).
You're right. I did run v0.22.21, which simply was current when I ran the benchmarks first. I absolutely will add the information, rerun, and thanks for the pointers. Glad to hear you got such a boost :clap:
Would it be faster? Maybe, I found that the dart and JS clients didn't reach their theoretical min latency of 3-5ms, so I'm inclined to believe that there's some bottlnecking on the server-side. I'd be very happy to be wrong on this.
From your perspective, would it make sense to just compare the respective dart and JS clients?
A few observations:
- Both benchmark driver and server are running locally on my laptop, completely saturating the machine.
- I've managed to achieve higher throughput with rust. At least part of it is probably that the client side is lighter on my already saturated machine.
- I've found that I'm getting roughly 3x the performance on a pretty humble 8700G desktop.
- I've found the file-system to have a non-trivial impact.
- The benchmark is only as fast as the bottleneck 50cm in front of the screen managed to make it run. The benchmark driver is here: https://github.com/trailbaseio/trailbase-benchmark/tree/main.... Dotnet is super swift, so I'd be surprised if it couldn't be further optimized (whereas Dart and Node are fairly client-side bottlenecked).