Hi everyone,

I'm writing to propose a very simple set of mempool/transaction relay
policies intended to aid L2/contract protocols. I realized that
the previously proposed Package Mempool Accept package RBF [1]
had a few remaining problems after digging into the RBF logic more [2].
This additional set of policies solves them without requiring a huge RBF
overhaul.

I've written an implementation (and docs) for Bitcoin Core:
https://github.com/bitcoin/bitcoin/pull/25038

(You may notice that this proposal incorporates feedback on the PR - thanks
Suhas Daftuar, Gregory Sanders, Bastien Teinturier, Anthony Towns, and
others.)

If you are interested in using package RBF/relay to bump presigned
transactions, I think you may be interested in reviewing this proposal.
This should solve Rule 3 pinning and perhaps allow us
to get rid of CPFP carve-out (yay!). I'm keen to hear if people find
the 1-anchor-output, 1000vB child limit too restrictive. Also, if you find a
pinning attack or something that makes it unusable for you, I would
really really like to know.

Note that transactions with nVersion=3 ("V3 transactions") are
currently non-standard in Bitcoin Core. That means **anything that was
standard before this policy change would still be standard
afterwards.** If you don't want your transactions to be subject to
these rules, just continue whatever you're doing and don't use
nVersion=3. AFAICT this shouldn't break anything, but let me know if
this would be disruptive for you?

**New Policies:**

This includes:
- a set of additional policy rules applying to V3 transactions
- modifications to package RBF rules

**V3 transactions:**

Existing standardness rules apply to V3 (e.g. min/max tx weight,
standard output types, cleanstack, etc.). The following additional
rules apply to V3:

1. A V3 transaction can be replaced, even if it does not signal BIP125
   replaceability. (It must also meet the other RBF rules around fees,
etc. for replacement to happen).

2. Any descendant of an unconfirmed V3 transaction must also be V3.

*Rationale*: Combined with Rule 1, this gives us the property of
"inherited" replaceability signaling when descendants of unconfirmed
transactions are created. Additionally, checking whether a transaction
signals replaceability this way does not require mempool traversal,
and does not change based on what transactions are mined. It also
makes subsequent rules about descendant limits much easier to check.

*Note*: The descendant of a *confirmed* V3 transaction does not need to be
V3.

3. An unconfirmed V3 transaction cannot have more than 1 descendant.

*Rationale*: (Upper bound) the larger the descendant limit, the more
transactions may need to be replaced. This is a problematic pinning
attack, i.e., a malicious counterparty prevents the transaction from
being replaced by adding many descendant transactions that aren't
fee-bumping.

(Lower bound) at least 1 descendant is required to allow CPFP of the
presigned transaction. The contract protocol can create presigned
transactions paying 0 fees and 1 output for attaching a CPFP at
broadcast time ("anchor output"). Without package RBF, multiple anchor
outputs would be required to allow each counterparty to fee-bump any
presigned transaction. With package RBF, since the presigned
transactions can replace each other, 1 anchor output is sufficient.

4. A V3 transaction that has an unconfirmed V3 ancestor cannot be
   larger than 1000 virtual bytes.

*Rationale*: (Upper bound) the larger the descendant size limit, the
more vbytes may need to be replaced. With default limits, if the child
is e.g. 100,000vB, that might be an additional 100,000sats (at
1sat/vbyte) or more, depending on the feerate.

(Lower bound) the smaller this limit, the fewer UTXOs a child may use
to fund this fee-bump. For example, only allowing the V3 child to have
2 inputs would require L2 protocols to manage a wallet with high-value
UTXOs and make batched fee-bumping impossible. However, as the
fee-bumping child only needs to fund fees (as opposed to payments),
just a few UTXOs should suffice.

With a limit of 1000 virtual bytes, depending on the output types, the
child can have 6-15 UTXOs, which should be enough to fund a fee-bump
without requiring a carefully-managed UTXO pool. With 1000 virtual
bytes as the descendant limit, the cost to replace a V3 transaction
has much lower variance.

*Rationale*: This makes the rule very easily "tacked on" to existing
logic for policy and wallets. A transaction may be up to 100KvB on its
own (`MAX_STANDARD_TX_WEIGHT`) and 101KvB with descendants
(`DEFAULT_DESCENDANT_SIZE_LIMIT_KVB`). If an existing V3 transaction
in the mempool is 100KvB, its descendant can only be 1000vB, even if
the policy is 10KvB.

**Package RBF modifications:**

1. The rule around unconfirmed inputs was
originally "A package may include new unconfirmed inputs, but the
ancestor feerate of the child must be at least as high as the ancestor
feerates of every transaction being replaced."

The package may still include new unconfirmed inputs. However,
the new rule is modified to be "The minimum between package feerate
and ancestor feerate of the child is not lower than the individual
feerates of all directly conflicting transactions and the ancestor
feerates of all original transactions."

*Rationale*: We are attempting to ensure that the replacement
transactions are not less incentive-compatible to mine. However, a
package/transaction's ancestor feerate is not perfectly representative
of its incentive compatibility; it may overestimate (some subset of
the ancestors could be included by itself if it has other high-feerate
descendants or are themselves higher feerate than this
package/transaction). Instead, we use the minimum between the package
feerate and ancestor feerate of the child as a more conservative value
than what was proposed originally.

2. A new rule is added, requiring that all package transactions with
mempool conflicts to be V3. This also means the "sponsoring"
child transaction must be V3.

*Note*: Combined with the V3 rules, this means the package must be
a child-with-parents package. Since package validation is only
attempted if the transactions do not pay sufficient fees to be
accepted on their own, this effectively means that only V3
transactions can pay to replace their ancestors' conflicts, and only
V3 transactions' replacements may be paid for by a descendant.

*Rationale*: The fee-related rules are economically rational for
ancestor packages, but not necessarily other types of packages.
A child-with-parents package is a type of ancestor package. It
may be fine to allow any ancestor package, but it's more difficult
to account for all of the possibilities. For example, it gets much
harder to see that we're applying the descendant limits correctly if
the package has a gnarly, many-generation, non-tree shape. I'm also
not sure if this policy is 100% incentive-compatible if the sponsor
is not a direct descendant of the sponsee.

Please see doc/policy/version3_transactions.md and
doc/policy/packages.md in the PR for the full set of rules.

**Intended usage for LN:**

Commitment transactions should be V3 and have 1 anchor output. They
can be signed with 0 fees (or 1sat/vbyte) once package relay is deployed
on a significant portion of the network. If the commitment tx must
be broadcast, determine the desired feerate at broadcast time and
spend the anchor output in a high feerate transaction. I'm going to
call the broadcasted commitment tx "the parent" and the attached
fee-bumping tx "the child."

- This child must be V3.
- This child must be at most 1000vB. Note this restricts the
  number of inputs you can use to fund the fee bump. Depending
on the output types, this is around 6-15.
- One child may fund fees for multiple commitment tx ("batched
  fee-bumping").
- To do a second fee-bump to add more fees, replace the
  *child* with a higher-feerate tx. Do not try to attach a grandchild.

Otherwise, never try to spend from an unconfirmed V3 transaction. The
descendant limits for V3 transactions are very restrictive.

**Expected Questions:**

"Does this fix Rule 3 Pinning?"
Yes. The V3 descendant limit restricts both you and your counterparty.
Assuming nodes adopted this policy, you may reasonably assume that you
only need to replace the commitment transaction + up to 1000vB.

"Only 1 anchor output? What if I need to bump counterparty's commitment tx
in mempool?"
You won't need to fee-bump a counterparty's commitment tx using CPFP.
You would just package RBF it by attaching a high-feerate child to
your commitment tx.

"Is this a privacy issue, i.e. doesn't it allow fingerprinting LN
transactions based on nVersion?"
Indeed it may be unrealistic to assume V3 transactions will be in
widespread use outside of L2. IIUC, unilateral closes are already
obvious LN transactions because of the HTLC inputs. For e.g.
cooperative closes and opens, I think it makes sense to continue using
V2. So, unless I'm missing something, this shouldn't make it worse.

"So a V3 transaction that doesn't signal BIP125 replaceability is
replaceable? Is that a backward compatibility issue?"
Yes it's replaceable. It's not an issue AFAICT because,
under previous policy, the V3 transaction wouldn't have been
in the mempool in the first place.

"Can a V2 transaction replace a V3 transaction and vice versa?"
Yes, otherwise someone can use V3 transactions to censor V2
transactions spending shared inputs. Note if the
original V3 transaction has an unconfirmed V3 parent, this would
violate the "inherited V3" rule and would be rejected.

Thanks for reading! Feedback and review would be much appreciated.

[1]:
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2021-September/019464.html
[2]:
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2022-January/019817.html

Best,
Gloria