AJ,

Thanks for the thoughtful post. I think your observations about how we view mempool policy in the Bitcoin Core project, and how that seems to be changing in the discussions around `-mempoolfullrbf`, are on-point and provide a helpful baseline for considering future policy changes.

For a long time I viewed fullrbf as an eventuality and I considered myself to be philosophically supportive of the idea.  However, after giving this issue some thought in the past few weeks, I am reversing my thinking on this.  Concretely, I will argue that we should continue to maintain a relay policy where replacements are rejected for transactions that don't opt-in to RBF (as described in BIP 125), and moreover, that we should remove the `-mempoolfullrbf` flag from Bitcoin Core’s latest release candidate and not plan to release software with that flag, unless (or until) circumstances change on the network, which I'll discuss below.

This is, of course, a nuanced topic, and among the considerations is a philosophy of how to think about the relay policy and configuration options that we make available in Bitcoin Core (a consideration that is perhaps unique to that project, but I think relevant for this mailing list).

I'll start with some technical issues regarding the benefits of enabling fullrbf on the network.  In the current BIP 125 regime, every time a transaction is created, a choice is made whether to subject the transaction to BIP 125’s RBF rules or not (based on the sequence values of the inputs).  So given that users can already opt-in to RBF, the benefit of a “fullrbf” network policy would be if, somehow, RBF users were still denied the benefits of RBF due to the existence of other transactions that don’t opt-in.

Along those lines, Antoine Riard brought up[1] a DoS vector that is available to someone who wants to interfere with multi-party funded transactions, and suggested that fullrbf would eliminate the problem.  After exploring that question again in this thread (thanks to Greg Sanders for clarifying this to me), I understand that the issue is around ensuring that a multiparty (coinjoin-type) protocol is able to make eventual progress, by having a candidate multiparty transaction either eventually confirm or become conflicted with something that has been confirmed, in which case the double-spend information could be used to start a new coinjoin round with fewer participants.  The concern Antoine and Greg have brought up is that non-rbf transactions can persist in the mempool ~indefinitely (at a low feerate and not subject to replacement) and interfere with progress being made in a coinjoin protocol.

However, it seems to me that similar problems exist for such a protocol even in a fullrbf world, as we understand that term today.  I mentioned the ability for rbf “pinning” to interfere with relay of the multiparty transaction (even if the conflicting transaction signals for RBF – a set of large but low feerate conflicting transactions can persist in the mempool and make it difficult for the coinjoin transaction from confirming, at least without attaching a very large fee); and as Greg mentioned in a followup, the BIP 125 rule to only permit 100 transactions to be removed from the mempool at a time during a replacement can also be used to pin a coinjoin protocol in the same way as a non-rbf transaction today.  It seems to me that what these multiparty protocols actually need is some sort of "maximal rbf" network policy: a way to guarantee that a transaction which should be desirable for a miner to mine would always get to a miner and considered for inclusion in a block, no matter what the state of node’s mempools on the network.

While that sounds like a reasonable thing to want on its face (and worth working on), it's not how opt-in RBF works today, nor is it how transaction relay has ever conceptually worked.  We have not, thus far, been able to come up with a total ordering on transaction desirability.  Moreover, due to all the DoS issues that exist with transaction relay, there are plenty of seemingly legitimate ways to construct transactions that would not relay well on the network.  Relay has only ever been a best-efforts concept, where we carve out a small subset of the entire transaction universe for which we try to optimize propagation.  The idea behind this approach is that if every use case we can come up with has some way to achieve its goals using transactions that should (eventually) be able to relay, then users wouldn’t have much demand for transactions that would deviate from the supported policies, and we therefore shouldn’t need to worry too much about incentive compatibility concerns when it comes to transaction types that wouldn’t relay at all, even if they are high feerate.  (And when those situations arise where the standard transactions do not accommodate some needed use case, developers typically work to define a policy that is compatible with our anti-DoS goals to support such use cases, such as with the recent proposal for version=3 transactions [2].)

BIP 125's RBF rules themselves were an effort to carve out just a subset of situations where a transaction should evict conflicting ones -- it was not a design that anyone thought would ensure that all replacements which "should" be mined would always propagate.  And I don't believe that we know how to design policy rules that would achieve the goals of this kind of multiparty protocol in a DoS resistant way, today.  Along those lines, I would point out that even the BIP 125 design itself is not entirely incentive compatible, in that it is possible to construct a replacement transaction that would evict transactions which would be preferable to be included in a block! [3]  (This has been known for years, but fixing this has proven difficult, and the only way to fix it that I’m aware of would be to make BIP 125 RBF even more restrictive than it is today. I do think this is something that needs to be worked on.)

Given the limitations of RBF as we have it today, it appears to be incorrect that a fullrbf network policy would solve the problems Antoine raised.  And so absent any other examples, it does not seem to me that fullrbf solves any problems for RBF users, who are already free to choose to subject their transactions to BIP 125’s RBF policy.  From this perspective, "enabling fullrbf" is really just taking away user choice to opt a transaction into a non-replacement policy regime.

I think we should ask, then, whether it is reasonable on its face that users might want to opt-in to a non-replacement policy?  Or in other words, is it reasonable for a user to mark a transaction as non-replaceable and have that indication be enforced by the network? Note that these are two different questions: you could imagine a world where fullrbf is a dominant policy, but users still use the BIP 125 signaling method to indicate, in an unenforced way, their intention to not replace a transaction.  This might give useful information to the network or the recipient for how to interact with such a transaction.

And I think that it's entirely possible that users would continue to use the BIP 125 signaling to indicate that they do not intend to replace a transaction.  For better or worse, this might be because zeroconf services continue to differentiate their behavior based on such a signal (possibly in conjunction with other factors), or it could be because there are other behaviors that could be utilized more effectively if the transaction originator has made such a signal, such as the recipient chaining an unconfirmed transaction as a way to bump the fee (CPFP) [4].

If it were to be the case that users continued to use BIP 125-style signaling to indicate that they do not plan to replace a transaction, would that be harmful to the network?  This is not something we can stop in our policy rules (short of censoring such transactions, an obviously bad idea).  I think network actors can always do things that we might think are harmful for the network, but that doesn’t mean that there are no legitimate use cases for the tools that such actors might be using.  Just because someone might use some policy to adopt a zeroconf model, doesn’t mean that others aren’t using the same policy to achieve benign ends (such as better CPFP behavior).

Moreover, while users might attempt to exploit services that offer zeroconf or other differentiated behavior to non-replacement signaling transactions, they also might not -- I think predicting user behavior in this way (and specifically predicting the complexities of what a business might do and whether users might try to subvert it) is beyond the scope of what we can do as protocol developers.  Instead, I think we can try to answer a different question: if a group of users were to want the ability to opt-in to a non-replacement policy regime, is that a technically sound option for us to have on the network and enforce in software?  Specifically, does that interfere with having a sensible anti-DoS mempool acceptance algorithm, or interfere with other protocols on the network, or necessarily run counter to the interests of miners or node operators?

And I think the answer to that question, in looking at the difference between opt-in RBF and fullrbf, is no: offering the ability to opt-in to a non-replacement regime for transactions doesn't introduce any fundamental issues with software or network policy or other protocols.  In a world where we only had fullrbf, I could imagine at some point down the road proposing a non-replacement signal myself, because the complexities around transaction chains (and pinning) are more complex for the RBF case than for the non-RBF case (and BIP 125 is not always incentive compatible to begin with!).  Conceptually, this is no different to me than the version=3 transaction policy proposal that has been advancing, if we think of it as a special set of restrictions on transactions designed to accommodate a particular use case.  

Philosophically, I think we should be looking to add non-interfering use cases to what the network supports.  

To those who argue for making fullrbf a default policy on the network (or even just offering a flag for users to enable fullrbf), I pose this hypothetical: suppose we deploy the v3 transaction policy proposal (which I hope will happen in the near future).  That policy would restrict the ways that outputs of a v3 transaction can be spent while the transaction is unconfirmed, including by limiting the number and size of descendants that such a transaction can have, and limiting the types of unconfirmed ancestors that can be included.  Suppose in a few years someone proposes that we add a "-disable_v3_transaction_enforcement" flag to our software, to let users decide to turn off those policy restrictions and treat v3 transactions the same as v2, for all the same reasons that could be argued today with fullrbf: miners might earn more revenue if we allowed multiple descendant v3 transactions; it's illogical for the recipient of a v3 transaction to believe what is a fundamentally unenforceable promise of a sender to not issue more high value children that descend from an unconfirmed transaction; it's inappropriate for Bitcoin Core to dictate policy on the network and we should honor user choice to turn off that flag if that’s what users want; if users are relying on v3’s policy restrictions for security then that is an unstable model and we should assume it will get broken[5].  

It’s obvious to me that adding a flag to disable v3 policy would be subversive to making the lightning use case for v3 transactions work.  And so my response to such a hypothetical proposal would be to argue that no, we should not enable users to disable this policy, because as long as that policy is just optional and working for those who want it, it shouldn’t harm anyone that we offer a tighter set of rules for a particular use case.  Adding a way to bypass those rules is just trying to break someone else’s use case, not trying to add a new one.  We should not wield "incentive compatibility" as a bludgeon for breaking things that appear to be working and not causing others harm.

I think this is exactly what is happening with fullrbf.

In comparing v3 transaction policy with opting out of transaction replacement, there is of course one significant difference that I have ignored thus far: I think the real difference is an opinion about whether non-replacement transactions that are being used today are, overall, bad for Bitcoin, and whether lightning’s use of v3 transactions in the future would be bad for Bitcoin. If you think that zeroconf is unequivocally bad, and that no one will be able to plausibly construct a case that lightning is bad, then that qualitative judgment might sway you to not worrying about the philosophical issues I've raised above, because these situations can be distinguished.

However I am not personally willing to say that I think, overall, non-rbf-signaling transactions in use on the network today are bad for Bitcoin (or that fullrbf is definitely good – BIP 125’s rbf rules are something we’ve been trying to improve upon for years, with little success).  Nor am I convinced that someone couldn’t put together a cogent argument for lightning being bad for Bitcoin, because of its reliance on relay policies that are difficult to design and impossible to guarantee as part of its security model.  So I choose instead to merely make a judgment that seems more factually verifiable, which is that non-replacement is a policy widely in use on the network today, and we largely don't have reason to think (as far as I know!) that the network is seeing a lot of transactions that would violate that policy.

If it did turn out that users were commonly signaling non-replacement, but then signing and trying to relay doublespends, then I think that would be a very good reason for Bitcoin Core to adopt fullrbf to reflect the reality of what is happening.  In the meantime, I think it makes more sense to say that because we have BIP 125, there seems to be no need for users to signal one way and behave another, and therefore there is no need to offer software that might break a policy that is working well for some users.  Other software projects might choose differently, and it is after all a permissionless network, so if this is in fact an unstable equilibrium that will not last, then presumably someday it will be apparent it is not working and we’ll abandon it.  But I think the philosophy of transaction relay policy in Bitcoin Core should be to support disparate use cases in order to try to make everything work better, rather than break things prematurely because we guess others will break them eventually anyway.

For those that have read this long email and still favor a fullrbf network policy (or even just the ability for users to be able to turn on fullrbf for themselves), I’d ask for thoughts on the following questions, which have guided my thinking on this:

Does fullrbf offer any benefits other than breaking zeroconf business practices?  If so, what are they?

Is it reasonable to enforce BIP 125's rbf rules on all transactions, if those rules themselves are not always incentive compatible?

If someone were to propose a command line option that breaks v3 transaction relay in the future, is there a logical basis for opposing that which is consistent with moving towards fullrbf now?

Cheers,
Suhas


[1] https://lists.linuxfoundation.org/pipermail/lightning-dev/2021-May/003033.html