Canceled GCS Multipart upload leaks HTTP2 streams and leads to REFUSED_STREAM errors

[KJTsanaktsidis]

Client

Google Cloud Storage

Describe Your Environment

SDK is used in Hashicorp Vault running on Ubuntu 18.04, on GCE

Expected Behavior

When canceling a context passed to ObjectHandle.NewWriter(ctx), all resources associated with the request should be cleaned up and it won't affect subsequent operations

Actual Behavior

I have a minimal reproduction test case here: https://github.com/KJTsanaktsidis/refused_stream_repro

The reproduction tries to upload lots of files to GCS with ObjectHandle.NewWriter(ctx), and cancels some of the contexts at random times. Eventually, after running for a few minutes, all uploads start returning the following error:

Post https://www.googleapis.com/upload/storage/v1/b/kjs_cool_vault_bucket/o?alt=json&prettyPrint=false&projection=full&uploadType=multipart: stream error: stream ID 45457; REFUSED_STREAM

I used a packet capture to debug the communication between Vault and GCS, and I found that the SDK would often

• Create a new HTTP2 stream to perform the multipart uplaod
• Send a HEADERS frame for POST /upload/storage/v1/b/vaultgcs_backend_us1_staging/o?alt=json&prettyPrint=false&projection=full&uploadType=multipart
• Never send a subsequent DATA frame for the request body, and
• Never send a subsequent RST_STREAM frame to clean up the stream

So, eventually, the GCS server starts sending REFUSED_STREAM errors back to the client when any new upload attempt is made, because the number of forgotten-about upload streams has exceeded the servers HTTP2 MAX_CONCURRENT_STREAM value of 100.

This issue seems to be the cause of hashicorp/vault#5419

[enocom]

This is likely related to #753, as well.

[enocom]

Another related issues: golang/go#20985.

[enocom]

Thanks for the reproduction @KJTsanaktsidis. It's been very helpful.

Meanwhile, here's another related issue: golang/go#27208.

[KJTsanaktsidis]

Yup - golang/go#27208 is exactly it!

I tried my reproduction using golang.org/x/net from the provided PR from that issue golang/net#18 (see branch working_version in my repo). This seems to have fixed the issue - I no longer got STREAM_REFUSED errors!

So I think this isn't a bug in google-cloud-go at all, and should go away on go 1.12?

[enocom]

Well done!

Provided the patch lands in Go 1.12, yes, the problem should go away. I'll ping the Go issue.

[KJTsanaktsidis]

Awesome - thanks heaps for your help in joining these dots!

[enocom]

Since this is an issue outside this repo, I'm going to close it.

Thanks again for all your help getting to the bottom of this.