Inspecting the waterfall of the cxx-driver project I noticed that starting with commit 1afd5102fd63322fa448f490d5f0dbaee78374bb on May 6th all the tasks on arm64 and ppc64le build variants failed. The preceding commit which had executed the same tasks, ab8091ec66cc765ec1117b9cb05978771242b8f7, on May 5th had all of them pass. I did some comparisons between the Evergreen logs of the 1afd5102fd63322fa448f490d5f0dbaee78374bb build and the ab8091ec66cc765ec1117b9cb05978771242b8f7 build. There did not appear to be any differences to which the failure could be attributed. After that, I restarted two of the tasks which had completed successfully for the build of commit ab8091ec66cc765ec1117b9cb05978771242b8f7 and they both failed. This makes it seem very likely that some external dependency or environmental factor has changed in a way as to trigger this failure. However, a review of the drivers-evergreen-tools.git history did not reveal any recent changes that might be suspect.

Comparing the output of the Evergreen tasks for the successful executions and failing executions, it seems that the failures are associated with the estimatedDocumentCount unified spec test. The successful tasks only ran the subtests that were described as "on less than 4.9.0", while the failing tasks also attempted the subtests described as "on 4.9.0 or greater". The strange thing is that the tasks in question all run against either server 5.0 or latest. It seems effectively impossible for this failure to manifest in this way.