Description of problem:

Running 10, 100, 200, 300 concurrent builds in Online Pro production.  This cluster has 60 computes, 3 registries and uses S3 as the backend storage for the registry.

At the 300 concurrent build level a significant (20%) number of builds start failing with either 404 errors for the push or a cryptic EOF message.   Sample logs from failed builds are attached.   I'll add a link to complete build logs in a private comment.   Registry pod logs and the syslogs of the nodes where the registry is running are attached.

Version-Release number of selected component (if applicable): 3.5.5.50


How reproducible: Always when running 300 concurrent builds (5 builds per compute)


Steps to Reproduce:
1.  Online Pro production
2.  Create 300 projects with a quickstart - we used nodejs-mongodb
3.  Run all 300 builds concurrently.   Looping the projects with oc start-build will work or you can use the svt build_test tool:  https://github.com/openshift/svt/blob/master/openshift_performance/ose3_perf/scripts/build_test-README.md

Actual results:

20% of the builds fail with 404 errors on the push or with EOF messages like this:

2017-07-03T21:25:12.365716675Z error: build error: Failed to push image: Put https://172.30.118.67:5000/v2/proj44/nodejs-mongodb-example/blobs/uploads/c6348934-ce2a-480f-8307-d20a06a0a0d9?_state=RVk_IZciwmaWU3X7URbhnPJDO4CzFFnefRROMuybp-F7Ik5hbWUiOiJwcm9qNDQvbm9kZWpzLW1vbmdvZGItZXhhbXBsZSIsIlVVSUQiOiJjNjM0ODkzNC1jZTJhLTQ4MGYtODMwNy1kMjBhMDZhMGEwZDkiLCJPZmZzZXQiOjgwNjAsIlN0YXJ0ZWRBdCI6IjIwMTctMDctMDNUMjE6MjQ6MTJaIn0%3D&digest=sha256%3Ae8ccea6b2ec2c6cca16c742a2777017e86a223fb36961385d1ddfaba573de1b8: EOF


Expected results:

Less than 1% failures is the target