Description
I think the issue in #5531 was revealed by the presence of this test, but this test, trying to simulate a portion of a previous version's state to exercise a specific upgrade step on a live system is inherently flaky, because it breaks core assumptions about the state of the built-in tables on a running system. That has consequences in other areas of the code than just the balancer race condition that was identified in #5531. I still think this test should be deleted, but agree that we need a new, more narrow, test to check for that race condition that was identified in #5531.
Originally posted by @ctubbsii in #5566
I believe the test discussed in #5566 wrote to the accumulo.root table when the metadata table changes to delete the accumulo.scanref table were flushed, resulting in a new RFile for the accumulo.metadata table. A new test should do something similar by causing changes to the root table, but should not do so by attempting to remove a built-in table, which breaks other assumptions on the live system.