It should be old_value=false, new_value=true, so when the compatibility is set to match one of the older versions, we will have this feature disabled.

ok, done.

Methods should have the following semantics:

x getX() {
  return x;
}

void cancel() {
  x.reset()/clear();
}

X detachX() {
  auto ret = move(x);
  x.reset()/clear();
  return move(ret);
}

I.e., detach is get + cancel.

So, getParallelReadingExtension now does what it is supposed to do. But detachParallelReadingExtension acts like cancel.

Thank you for your patient explanation of the meaning of detach. However, it may not fit the semantics here.

The original parent part reading step needs to pass ParallelReadingExtension to the projection reading step, but it doesn't necessarily require "detach" because the original parent part reading step may still need ParallelReadingExtension, and at the same time, the projection reading step may need to clear the received ParallelReadingExtension.

Therefore, perhaps clearParallelReadingExtension is the most semantically appropriate here.

Please add a test case to ensure memory-bound merging and projections can coexist.

Okay, I’ve added the test cases.

I’ve noticed that when parallel replicas support projections while aggregation_in_order is active, it often leads to errors, causing different replicas to select different coordinator modes. Therefore, it might be a good idea to disable parallel replicas from reading projections when aggregation_in_order is active

When connecting to the initiator, we read the initiator_protocol_version value from the socket. In newer versions, it is set to the current DBMS_PARALLEL_REPLICAS_PROTOCOL_VERSION, but previous versions had different values. Apparently, the test for backward compatibility of the protocol was broken some time ago, so the issue does not reproduce in the CI. I will fix it shortly here, and you'll be able to see the problem.

What I mean is that this test will also succeed if we set enable_parallel_replicas=0, which is not good because it tests the behaviour of parallel replicas specifically. So, I'd like to ensure that the test will fail with disabled parallel replicas (because we have some traces of using parallel replicas in the reference file).