@dt saw on one of his clusters that a historical read issued by a samplerProcessor was stuck in an indefinite loop of hitting an intent during evaluation, not pushing the intent's txn, and retrying evalution. Some debugging revealed that this endless loop was taking place on a follower replica.

Follower replicas do not have their lockTable enabled, which means that they will not add discovered locks to the table. The result of this is that the intents were being ignored during each "concurrency retry" and never pushed. Because the request was performing a follower read, it was not prevented from evaluating during its next iteration, so the loop never terminated. This was a counter-example to:

We need to fix this by either pushing immediately when an intent is discovered on a replica without an enabled lockTable or throwing a NotLeaseholderError in this situation.