A performance issue was identified in Ginkgo where every Openshift-tests invocation was triggering a full stack trace for every test, every time a test ran. We added a carry patch to our forked ginkgo (which is mostly static) to disable capturing this trace and reduced total CPU used by an e2e run for 4.5 from ~10.5k core seconds to ~4.5k core seconds (60% reduction in total CPU). As e2e is one of our more significant loads this has a significant reduction in CI resources consumed. The fix is safe to use within our limited use of Ginkgo and has been backported to 4.5 and 4.4 already with a vendor bump. We should vendor bump in 4.3 as well since that's a significant amount of installs.
Ginkgo patch in https://github.com/openshift/onsi-ginkgo/commit/67da0dd32db383f566d9d2b6559bf1d0ef03b5cb Vendor bump in 4.5: https://github.com/openshift/origin/pull/24742 Vendor bump in 4.4: https://github.com/openshift/origin/pull/24743 Vendor bump in 4.3: https://github.com/openshift/origin/pull/24746 Verified in 4.5 and 4.4 builds, graph of pod core seconds from api.ci (the tall job at the same time is a 4.3 build) https://www.dropbox.com/s/jrpw3bsxclqcb64/Screenshot%202020-03-22%2018.30.38.png?raw=1
This merged but is not something QE verifies.