Will R2 beta launch without signed requests working? Will R2 support signing content-length? This was a weird one that AWS didn’t support immediately.

Content-length header*

Yeah, should be no problem to do with unbound. Though it would be nice to have the feature built-in. I'm thinking of building a Docker Registry on Workers and R2 with ability to push images. Since the blob/layer hash is used as the object key, I need to ensure the client isn't providing a bad hash. I could save the upload to a temporary key while I calculate the hash of the stream, but then I'd need to copy the object afterwards.

It makes sense you need to evict at some point. But you're already keeping two runtime versions in memory during those 30 seconds, so I don't understand how you risk running OOM by keeping it longer?

I believe because new requests are hitting the new runtime.

So imagine that there's a Worker in the old runtime taking an hour to shut down...

The old runtime, which is big in RAM, needs to stay up for that (so it's not the normal Worker RAM usage, it's higher). But then requests come into the new runtime...

And if a lot of requests come into the new runtime, OOM.

The number of active Workers in the old runtime will fall over time as they finish their request. You're not suddenly handling more requests just because they're split between runtimes. The way I see it, the only difference (between normal operation and runtime change) is the overhead of the new runtime which the machine can already handle (and does so for 30 seconds). But perhaps I'm missing something.

Oh yeah, you're correct. No clue, I'd be interested to know too. 🙂

Indeed. I don't have any problem with eviction after 30 seconds in these cases. I also understand you don't want to run different/multiple versions of software for longer periods of time. But I'm curious to know why the grace period isn't, say, 5 minutes as that would cover more of the long-running Workers as well.

Interesting, I always also thought its a wall time limit, not 30s of CPU time 🤯 As the docs used to say it includes IO time. (https://github.com/cloudflare/cloudflare-docs/pull/2830/files)

Read a bit further down.

It is wall time.

no, I mean I also lived in an impression its killed after 30s

Oh.

Yeah, I did too 😄

Good to know, actually.

getting 500 internal error calling

UploadPart

for any partNumber > 1, is this working yet? also 500 internal error trying to complete even a one-part multipart upload : ( now sometimes working if you upload in sequence quickly, but flaky, hmm

Beta will be without signed requests working. I believe whatever headers you tell us you signed will get validated.

Multipart is definitely implemented and functional AFAIk

How big is the part you're sending through?

yea it does get wedged sometimes though (all subsequent calls will 500), and you have to abandon that uploadId and start another. Also does not work via the aws cli. Since we cannot list uploads, can we assume multipart uploads get cleaned up (and not charged) after some time?

Naw you have to abort currently

hehe i'd better not lose track of those long uploadIds then

But we'll add listing uploads before we start charging

It's not lost

k

small potatoes: 5.1mb to 20mb