I'm not sure this comparison makes sense. The existence of a low-fee-rate descendant doesn't make a transaction worse for miners, but it would cause the feerate to look worse in this comparison.

So, you mean the scenario where we have a high fee-rate transaction that is spent by one or more low fee-rate transactions? For instance suppose we have two transactions in our mempool: tx1a, 1KB w/ 1mBTC fee, which is spent by tx2, 10KB w/ 1mBTC fee. We get tx1b, 10KB w/ 2.1mBTC fee. Since the overall feerate of tx1b is higher than tx1a+tx2, it'll be accepted, even though a miner might have rather mined tx1a instead, ignoring tx2 (for now).

I agree that's less than optimal. If we make the assumption that there's always more demand for blockchain space than supply, it might be reasonable for the replacement logic criteria to be whether or not we're increasing the fee-rate of any subset of the mempool. (while still paying more fees than the replaced transactions)

Without taking CPFP into account, you could simply use the same kind of priority heap logic as in CreateNewBlock() on the list of transactions that would be replaced by the new transaction. You'd iterate through the heap from highest priority to lowest, stopping once you had found as many bytes worth of transactions as the candidate replacement. If the fee-rate of the replacement is higher than the fee-rate of those transactions, accept the replacement into the mempool.

To take CPFP into account... Thinking about it more the mempool package tracking is probably backwards from what we want. Right now it tracks descendants, when really what we want to know is "what's the total fee-rate if I mine this transaction, and all ancestors?" If we kept track of "packages" that way, we'd be able to do the comparison by determining if the total fee-rate of the new package created by the replacement is higher than the fee-rates of all the packages invalidated by it. I actually did some work on something along these lines a few years ago, though didn't finish it - the implementation is a lot simpler now that we have strict ancestor limits. (when limiting, just throw away the tx associated with the lowest fee-rate package, which is guaranteed to have no descendants)

For now though I'd be inclined to merge this PR as-is, as all the above options are pretty complex. I also don't see any way this code could be used in an DoS attack.

Right, the distinction between ancestor and descendant package is what I was getting at.

Descendant packages are I think the right thing to use for mempool limiting. I don't follow what you're saying about a limiting algorithm that uses fee with ancestors -- it's entirely possible that the worst thing under that sort would have descendants in the mempool.

(FYI I have a branch that implements ancestor packages. I might propose merging it at some point if it looks like we can take advantage of it in the mining code, but I'm not ready to advocate that now.)

Anyway in this code, we are comparing the feerate of the replacing transaction (with uncalculated/unknown fee rate including its ancestors) to an estimate of the feerate of the descendants of all the conflicting transactions. This strikes me as incorrect on both fronts; by not putting any bounds on the ancestor fee rate, we might be accepting a replacement transaction that is unlikely to confirm anytime soon. On the other hand, by looking at feerate with children (and overcounting those children, at least in the current implementation), we might be making it so that miners would prefer the original transactions to the replacing one.

I don't know that either of these issues constitutes an attack, but I do think it's useful to try to help users avoid shooting themselves in the foot, say by accidentally adding an input that is part of a long unconfirmed chain (causing the replacing transaction to be worse), and to give miners code that doesn't require further optimization to do the economically rational thing.

So with that in mind, how about this:

• Require that any new inputs that show up in the replacing transaction be already confirmed. In the future, if we do merge something like ancestor package tracking and better mining code, we could update this test appropriately.
• Require that for each entry E in setConflicts,
  feerate(replacing transaction) > max(feerate of E, feerate of E with descendants). That doesn't completely eliminate the possibility that it could be somehow worse for a miner to accept the new transaction, but it eliminates some degenerate cases (where a high fee rate transaction is dragged down by lower fee rate transactions) and is easy to calculate.

Fixed both these issues.

I decided not to do the max() version of this, as I think the requirement that the new fees > total-replaced-fees is sufficient; might help to get txs unstuck in some cases, and non-CPFP miners aren't evaluating that anyway.

When I check out the commit referenced in the README and run the test, I get an error:
AttributeError: 'Proxy' object has no attribute 'generate'

I think this function is not defined in the Proxy class? Adding it in the appropriate place in python-bitcoinlib seems to fix it.

It's not included, no. You can just do a "call" instead.

proxy.call("generate", 1)

Fixed.

I forgot to add generate() when python-bitcoinlib dropped support for calling RPC commands implicitly; replaced with .call() so as to continue to use the official v0.5.0 release rather than git master.

Wouldn't it be better to put the new code (from L987) after this check instead of before?
It would avoid doing further mempool validations for transactions that are going to be rejected by this check anyway.
I know there can be transactions that would pass this check but be rejected as a replacement, but this check already has the inputs of the transaction cached, so it seems cheaper than the RBF logic and it makes sense to me to do it first.
Note that this minor nit wouldn't change the total diff (although it can be done later with other changes too [for example, a rebased #6445 ]).

You mean after the CheckInputs() line? The RBF code is just a bunch of pointer following of in-memory data, with limited depth and breadth, so it shouldn't be expensive code to run - remember that the mempool has tx fee and size data in the CTxMemPool structs.

Never mind, I misread this line for AreInputsStandard().

The following was later removed by #7594