Hmm, I don't know anything about  SIGHASH_BUNDLE. The only references online I can find are just mentions (mostly from you). What is  SIGHASH_BUNDLE?

> unless you're binding a WTXID

That could work, but it would exclude cases where you have a transaction that has already been partially signed and someone wants to, say, only sign that transaction if some 3rd party signs a transaction paying part of the fee for it. Kind of a niche use case, but it would be nice to support it if possible. If the transaction hasn't been signed at all yet, a new transaction can just be created that includes the prospective fee-payer, and if the transaction is fully signed then it has a WTXID to use.

> then you can have fee bumping cycles

What kind of cycles do you mean? You're saying these cycles would make it less robust to reorgs?

> OP_VER

I assume you mean something other than pushing the version onto the stack? Is that related to your fee account idea?


On Wed, Jan 19, 2022 at 1:32 AM Jeremy <jlrubin@mit.edu> wrote:
Ah my bad i misread what you were saying as being about SIGHASH_BUNDLE like proposals.

For what you're discussing, I previously proposed https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2020-September/018168.html which is similar.

The benefit of the OP_VER output is that SIGHASH_EXTERNAL has the issue that unless you're binding a WTXID (which is maybe too specific?) then you can have fee bumping cycles. Doing OP_VER output w/ TXID guarantees that you are acyclic.

The difference between a fee account and this approach basically boils down to the impact on e.g. reorg stability, where the deposit/withdraw mechanism is a bit more "robust" for reorderings in reorgs than the in-band transaction approach, although they are very similar.

--


On Tue, Jan 18, 2022 at 8:53 PM Billy Tetrud <billy.tetrud@gmail.com> wrote:
>  because you make transactions third party malleable it becomes possible to bundle and unbundle transactions.

What I was suggesting doesn't make it possible to malleate someone else's transaction. I guess maybe my proposal of using a sighash flag might have been unclear. Imagine it as a script opcode that just says "this transaction must be mined with this other transaction" - the only difference being that you can use any output with any encumberance as an input for fee bumping. It doesn't prevent the original transaction from being mined on its own. So adding junk inputs would be no more of a problem than dust attacks already are. It would be used exactly like cpfp, except it doesn't spend the parent. 

I don't think what I was suggesting is as different from your proposal. All the problems of fee revenue optimization and feerate rules that you mentioned seem like they'd also exist for your proposal, or for cpfp. Let me know if I should clarify further. 

On Tue, Jan 18, 2022 at 8:51 PM Jeremy <jlrubin@mit.edu> wrote:
The issue with sighash flags is that because you make transactions third party malleable it becomes possible to bundle and unbundle transactions.

This means there are circumstances where an attacker could e.g. see your txn, and then add a lot of junk change/inputs + 25 descendants and strongly anchor your transaction to the bottom of the mempool.

because of rbf rules requiring more fee and feerate, this means you have to bump across the whole package and that can get really messy.

more generally speaking, you could imagine a future where mempools track many alternative things that might want to be in a transaction.

suppose there are N inputs each with a weight and an amount of fee being added and the sighash flags let me pick any subset of them. However, for a txn to be standard it must be < 100k bytes and for it to be consensus < 1mb. Now it is possible you have to solve a knapsack problem in order to rationally bundle this transaction out of all possibilities.

This problem can get even thornier, suppose that the inputs I'm adding themselves are the outputs of another txn in the mempool, now i have to track and propagate the feerates of that child back up to the parent txn and track all these dependencies.

perhaps with very careful engineering these issues can be tamed. however it seems with sponsors or fee accounts, by separating the pays-for from the participates-in concerns we can greatly simplify it to something like: compute effective feerate for a txn, including all sponsors that pay more than the feerate of the base txn. Mine that txn and it's subsidies using the normal algo. If you run out of space, all subsidies are same-sized so just take the ones that pay the highest amount up until the added marginal feerate is less than the next eligible txn.




On Tue, Jan 18, 2022 at 6:38 PM Billy Tetrud <billy.tetrud@gmail.com> wrote:
I see, its not primarily to make it cheaper to append fees, but also allows appending fees in cases that aren't possible now. Is that right? I can certainly see the benefit of a more general way to add a fee to any transaction, regardless of whether you're related to that transaction or not. 

How would you compare the pros and cons of your account-based approach to something like a new sighash flag? Eg a sighash flag that says "I'm signing this transaction, but the signature is only valid if mined in the same block as transaction X (or maybe transactions LIST)". This could be named SIGHASH_EXTERNAL. Doing this would be a lot more similar to other bitcoin transactions, and no special account would need to be created. Any transaction could specify this. At least that's the first thought I would have in designing a way to arbitrarily bump fees. Have you compared your solution to something more familiar like that?

On Tue, Jan 18, 2022 at 11:43 AM Jeremy <jlrubin@mit.edu> wrote:
Can you clarify what you mean by "improve the situation"?

There's a potential mild bytes savings, but the bigger deal is that the API should be much less vulnerable to pinning issues, fix dust leakage for eltoo like protocols, and just generally allow protocol designs to be fully abstracted from paying fees. You can't easily mathematically quantify API improvements like that.


On Tue, Jan 18, 2022 at 8:13 AM Billy Tetrud <billy.tetrud@gmail.com> wrote:
Do you have any back-of-the-napkin math on quantifying how much this would improve the situation vs existing methods (eg cpfp)?



On Sat, Jan 1, 2022 at 2:04 PM Jeremy via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> wrote:
Happy new years devs,

I figured I would share some thoughts for conceptual review that have been bouncing around my head as an opportunity to clean up the fee paying semantics in bitcoin "for good". The design space is very wide on the approach I'll share, so below is just a sketch of how it could work which I'm sure could be improved greatly.

Transaction fees are an integral part of bitcoin.

However, due to quirks of Bitcoin's transaction design, fees are a part of the transactions that they occur in.

While this works in a "Bitcoin 1.0" world, where all transactions are simple on-chain transfers, real world use of Bitcoin requires support for things like Fee Bumping stuck transactions, DoS resistant Payment Channels, and other long lived Smart Contracts that can't predict future fee rates. Having the fees paid in band makes writing these contracts much more difficult as you can't merely express the logic you want for the transaction, but also the fees.

Previously, I proposed a special type of transaction called a "Sponsor" which has some special consensus + mempool rules to allow arbitrarily appending fees to a transaction to bump it up in the mempool.

As an alternative, we could establish an account system in Bitcoin as an "extension block".

Here's how it might work:

1. Define a special anyone can spend output type that is a "fee account" (e.g. segwit V2). Such outputs have a redeeming key and an amount associated with them, but are overall anyone can spend.
2. All deposits to these outputs get stored in a separate UTXO database for fee accounts
3. Fee accounts can sign only two kinds of transaction: A: a fee amount and a TXID (or Outpoint?); B: a withdraw amount, a fee, and an address
4. These transactions are committed in an extension block merkle tree. While the actual signature must cover the TXID/Outpoint, the committed data need only cover the index in the block of the transaction. The public key for account lookup can be recovered from the message + signature.
5. In any block, any of the fee account deposits can be: released into fees if there is a corresponding tx; consolidated together to reduce the number of utxos (this can be just an OP_TRUE no metadata needed); or released into fees *and paid back* into the requested withdrawal key (encumbering a 100 block timeout). Signatures must be unique in a block.
6. Mempool logic is updated to allow attaching of account fee spends to transactions, the mempool can restrict that an account is not allowed more spend more than it's balance.

But aren't accounts "bad"?

Yes, accounts are bad. But these accounts are not bad, because any funds withdrawn from the fee extension are fundamentally locked for 100 blocks as a coinbase output, so there should be no issues with any series of reorgs. Further, since there is no "rich state" for these accounts, the state updates can always be applied in a conflict-free way in any order.


Improving the privacy of this design:

This design could likely be modified to implement something like Tornado.cash or something else so that the fee account paying can be unlinked from the transaction being paid for, improving privacy at the expense of being a bit more expensive.

Other operations could be added to allow a trustless mixing to be done by miners automatically where groups of accounts with similar values are trustlessly  split into a common denominator and change, and keys are derived via a verifiable stealth address like protocol (so fee balances can be discovered by tracing the updates posted). These updates could also be produced by individuals rather than miners, and miners could simply honor them with better privacy. While a miner generating an update would be able to deanonymize their mixes, if you have your account mixed several times by independent miners that could potentially add sufficient privacy.

The LN can also be used with PTLCs to, in theory, have another individual paid to sponsor a transaction on your behalf only if they reveal a valid sig from their fee paying account, although under this model it's hard to ensure that the owner doesn't pay a fee and then 'cancel' by withdrawing the rest. However, this could be partly solved by using reputable fee accounts (reputation could be measured somewhat decentralized-ly by longevity of the account and transactions paid for historically).

Scalability

This design is fundamentally 'decent' for scalability because adding fees to a transaction does not require adding inputs or outputs and does not require tracking substantial amounts of new state.

Paying someone else to pay for you via the LN also helps make this more efficient if the withdrawal issues can be fixed.

Lightning:

This type of design works really well for channels because the addition of fees to e.g. a channel state does not require any sort of pre-planning (e.g. anchors) or transaction flexibility (SIGHASH flags). This sort of design is naturally immune to pinning issues since you could offer to pay a fee for any TXID and the number of fee adding offers does not need to be restricted in the same way the descendant transactions would need to be.

Without a fork?

This type of design could be done as a federated network that bribes miners -- potentially even retroactively after a block is formed. That might be sufficient to prove the concept works before a consensus upgrade is deployed, but such an approach does mean there is a centralizing layer interfering with normal mining.


Happy new year!!

Jeremy

_______________________________________________
bitcoin-dev mailing list
bitcoin-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev