When stressing TestKVNemesisSingleNode with a short timeout (2-5 minutes), it's easy to see the test time out. This appears to be due to thrashing within the Range merge transaction.

The Range merge transaction does not use the standard transaction retry mechanism (i.e. epochs). Instead, it uses a completely separate transaction when restarting due to retry errors. This leaves room for thrashing and livelock if multiple transactions keep stepping on each other's toes.

make roachprod-stress PKG=./pkg/kv/kvnemesis TESTS=TestKVNemesisSingleNode TESTTIMEOUT=2m TESTFLAGS='-v -show-logs' STRESSFLAGS='-stderr -maxfails 1' CLUSTER=<cluster-name>