Sorry--the throttling was at the AWS Lambda layer, not S3. We were being throttled because we'd deploy a new container image and suddenly thousands of new containers are all simultaneously trying to pull the database file from S3.

We aim to return a response in the single digit milliseconds and sometimes get tens of thousands of requests per second, so even if it only takes a second or two to fetch that file from S3, the request isn't getting served while it's happening, and new requests are coming in.

You very quickly hit your Lambda concurrency limit and get throttled just waiting for your instances to fetch the file, even though logically you're doing exactly what you planned to.

By having the file exist already in the container image, you lean on AWS's existing tools for a phased rollout to replace portions of your deployment at a time, and every one is responding in single digit milliseconds from its very first request.

EDIT: The same technique could be applied for other container management systems, but for stuff like Kubernetes or ECS, it might be simpler to use OP's method with a readiness check that only returns true if you fetched the file successfully. And maybe some other logic to do something if your file gets too stale, or you're failing to fetch updates for some reason.

Are lambdas guaranteed to be warm on the first request to serve in single digit milliseconds?

They can be with provisioned concurrency.

If provisioned concurrency is used, then fetching the database at startup would be a valid strategy given you could still achieve flat response times.

That's true. I prefer this approach because it removes that additional thing (the request to S3) that can be slow or fail at runtime. Or "initialization" time, I guess, depending on how you look at it.

I wrote a tool to handle micro blobs specifically because we were being heavily rate limited by S3 for both writes and reads. We got about 3k/s per bucket before S3 rate limiting started kicking in hard.

Granted we also used said tool to bundle objects together in a way that required sezo state to track so that we could fetch them as needed cheaply and efficiently so it wasn't a pure S3 issue.

Interesting, thanks! PUT is advertised at 3500/s, so with a combo load, you were at least within range of advertised limits. I have not approached that scale so didn't know, it was a real question!

Yeah I was processing a bunch of iceberg catalog data, it was pretty trivial to get to this point on both PUTs and GETs with our data volume, I was doing 400,000 requests/m and of course my testing was writing to one prefix :)

I asked the S3 team what “prefix” meant at reinvent, and my current understanding is “whatever starting length of key gives a reasonable cardinality for your objects”.

So if your keys are 2024/12/03/22-45:24 etc, I would expect the prefix to be first 7 characters. If your keys are UUIDs I’d assume first two or three. For ULIDs I’d assume first 10. I this there’s a function that does stat analysis on key samples to figure out reasonable sharding.

I have never seen this explained, so thank you! Sounds like it's kind of "up to S3 and probably not predictable by you" -- which at least explains why it wasn't clear!

If you don't have "a lot" of keys, then you probably have only one prefix, maybe? Without them documenting the target order of magnitude of their shards?

I would assume so, the extreme case being just one key, which of course has only one partition. But see https://youtu.be/NXehLy7IiPM (2024 Reinvent S3 deep dive) - there’s still replication happening on single objects. So it’s still sort of sharded, but I do think key partitions where groups of keys have shared choke points based on sort order exist.

Yep. Works similarly with google cloud storage buckets. It seems like the indexing function they use for splitting/distributing/sharding access looks at your objects keys and finds a common prefix to do this.

The problem with a date based key like the one you used (that's very common) is that if you read a lot of files that tend to be from the same date (for example: for data analysis you read all the files from one day or week, not files randomly distributed) all those files are going to share the same prefix and are going to be located in the same shard, reducing performance until the load is so high that Google splits that index in parts and begins to distribute your data in other shards.

For this reason they recommend to think your key name beforehand and split that prefix using some sort of random hash in a reasonable location of your key:

https://cloud.google.com/storage/docs/request-rate#naming-co...

It would be nice if S3 provided similar public guidance. For instance:

> Adding a random string after a common prefix still allows auto-scaling to work, but…

No way to know if that's true of S3's algorithm too without them revealing it.

Yep, seems to hint something in the first paragraph of a performance tip [0] but it doesn't specify how does it choose prefixes, or how many prefixes does it shard, or anything...

  0: https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance.html

Yes, I've been throttled many times by S3. My largest database is ingesting ~5PB/day and that turns into a lot of files in S3. At one point we changed our S3 key scheme to not have hashes up front, which unlocked some simplicity in control plane operations like deleting old files; we did this on the strength of the announcement from AWS that you no longer needed to get clever with prefixes.

This was incorrect at our scale, and we had to switch back.