Wrong chunking of parts when uploading to S3

[incubos]

Docker Registry 2.7.0 with S3 storage backend chunkSize set to 32MB:

...
storage:
    s3: 
        ...
        chunksize: 33554432

When pushing a ≈70MB layer to the registry it is split into 32MB and ≈38MB parts which doesn't respect chunkSize configuration. Here are the intercepted part upload HTTP requests, please, focus on Content-Length headers:

// Part 1 (OK)
PUT /******/_uploads/e291fc4d-b670-4f7b-aa47-76ec637918ff/data?partNumber=1&uploadId=68c1a260-c6a7-11ea-a23b-a50eac241fcb HTTP/1.1
Host: ******
User-Agent: aws-sdk-go/1.15.11 (go1.11.6; linux; amd64)
Content-Length: 33554432
...

// Part 2 (not OK)
PUT /******/_uploads/e291fc4d-b670-4f7b-aa47-76ec637918ff/data?partNumber=2&uploadId=68c1a260-c6a7-11ea-a
23b-a50eac241fcb HTTP/1.1
Host: ******
User-Agent: aws-sdk-go/1.15.11 (go1.11.6; linux; amd64)
Content-Length: 38611098
...

[milosgajdos]

@incubos can you verify if this is broken on main branch which is a candidate for v3?

[incubos]

@incubos can you verify if this is broken on main branch which is a candidate for v3?

Could you please reference a commit containing the fix for the issue?
Unfortunately, I will not be able to verify in near future. Maybe someone else is able to?

[milosgajdos]

I havent dived into this properly. I was reviewing some open issues and asking the reporters to check on their side.

There have been a lot of updates in the past 3 years across multiple storage drivers. Now as for your question:

When pushing a ≈70MB layer to the registry it is split into 32MB and ≈38MB parts which doesn't respect chunkSize configuration.

This is the current behaviour of the S3 storage driver -- the last chunk is combined into the chunk that's about to be "flushed" to S3:

distribution/registry/storage/driver/s3-aws/s3.go

Line 1640 in 79ef555

if _, err := buf.Write(w.pending.data); err != nil {

This isn't going to change anytime soon, alas there has been some recent discussion around this issue that cropped up recently #3940

So let's keep this open until there is some "fix"