IOPS
Hi folks! I am new to Neon and currently exploring it to have a db-per-tenant. From reading the docs, I can see that Neon is superior to Aurora Serverless V2 (at least for the points I care about). But I am not able to find documentation on how many IOPs are supported per DB on Neon. Does anyone have an idea?
6 Replies
ratty-blush•6mo ago
Hi, we don't measure IOPS on Neon, so translating between Aurora and Neon can be a little tricky
foreign-sapphire•6mo ago
Does IOPS apply to Neon even when you are using Bare Metal?
ratty-blush•6mo ago
It's complicated, given we have separated storage and compute.
Our storage is bare metal nvme, although we do offload (very) cold pages to s3. Our compute also has nvme-based local file cache (LFC).
But there is an IO cost when the page isn't cached on the compute and has to be fetched from our storage nodes. We do measure this data internally but it's not factored into any kind of billing. There likely is a limit to how many getpage requests we can service but I'm not sure what that is off the top of my head - I think even our largest databases struggle to hit it.
We do also employ pre-fetching, which should be even more improved in PG18 with some more async-io functionality. Prefetching allows us to get the data cached before it's needed in the query, which improves the latencies quite significantly.
ambitious-aqua•6mo ago
Hi Conrad! Thanks for the details and the context. That's really helpful.
Hi @Conrad Ludgate ! Quick follow up: If not IOPs, are there any benchmarks on how many transactions per seconds are viable on Neon?
foreign-sapphire•6mo ago
This is a very interesting topic that might be worth writing a blog about.
Is there any pre-release notes or future changes for PG18?
And finally, how big is approximately the largest database they have to not reach those numbers? Sure, as long as it's not private or sensitive information.
ratty-blush•6mo ago
I saw this new brief document https://github.com/postgres/postgres/blob/master/src/backend/storage/aio/README.md
I'm not sure if we have any public throughput benchmarks. I'd have to check
We have some multi-TB database, I think double-digit TB now