Hi James,
I fully agree on the need to reframe the conversation around
mempools/fee-bumping/L2s though please see my following comments, it's
far from simple!
Layering on special cases, more carve-outs, and X and Y percentage
thresholds is going to make reasoning about the mempool harder than
it
already is.
I think that's true with a lot of (all ?) pieces of software, there is
a trend towards complexification. As new Bitcoin use-cases such as LN
or vaults appear, it's not surprising to see the base layer upper
interfaces changing to support the requirements. Same with kernels, at
beginning, you can have a basic memory support with paging/memory
rights/kernel allocators then as you start to support more
platforms/devices you might have to support swaps/DMA/VMs
management...
That we should keep the complexity reasonably sane to enable human
auditing, and maybe learn from the failures of systems engineering,
that's something to muse on.
The countervailing force here ends up being spam prevention (a la
min-relay-fee)
to prevent someone from consuming bandwidth and mempool space with a
long series of
infinitesimal fee-bumps.
I think here we should dissociate a) long-chain of transactions and b)
high-number of repeated fee-bumps.
For a) _if_ SIGHASH_ANYPREVOUT is deployed and Eltoo adopted as a
primary update mechanism for stateful L2s, one might envision
long-chain of update transactions servicing as a new pinning vector,
where all the chain elements are offering a compelling feerate/fees.
It might be solvable with smarter mempool logic sorting the elements
from the best offer to the lower ones, though that issue would need
more serious investigation.
For b) if we bound with a hard constant the number of RBF attempts, we
decrease the fault-tolerance of L2 transaction issuers. Some of them
might connect directly to the miners because they're offering higher
number of incentive-compatible RBF attempts than vanilla full-nodes.
That might provoke a more or slow centralization of the transaction
relay topology...
Instead of prompting a rebroadcast of the original transaction for
replacement, which contains a lot of data not new to the network, it
makes more sense to broadcast the "diff" which is the additive
contribution towards some txn's feerate.
In a distributed system such as the Bitcoin p2p network, you might
have transaction A and transaction B broadcast at the same time and
your peer topology might fluctuate between original send and broadcast
of the diff, you don't know who's seen what... You might inefficiently
announce diff A on top of B and diff B on top A. We might leverage set
reconciliation there a la Erlay, though likely with increased
round-trips.
It's probably uncontroversial at this
point to say that even RBF itself is kind of a hack - a special
sequence number should not be necessary for post-broadcast
contribution
toward feerate.
I think here we should dissociate the replace-by-fee mechanism itself
from the replacement signaling one. To have a functional RBF, you
don't need signaling at all, just consider all received transactions
as replaceable. The replacement signaling one has been historically
motivated to protect the applications relying on zero-confs (with all
the past polemics about the well-foundedness of such claims on other
nodes' policy).
In a sane design, no structural foresight - and certainly no wasted
bytes in the form of unused anchor outputs - should be needed in
order
to add to a miner's reward for confirming a given transaction.
Have you heard about SIGHASH_GROUP [0] ? It would move away from the
transaction to enable arbitrary bundles of input/outputs. You will
have your pre-signed bundle of inputs/outputs enforcing your
LN/vaults/L2 and then at broadcast time, you can attach an
input/output. I think it would answer your structural foresight.
One of the practical downsides of CPFP that I haven't seen discussed
in
this conversation is that it requires the transaction to pre-specify
the
keys needed to sign for fee bumps. This is problematic if you're,
for
example, using a vault structure that makes use of pre-signed
transactions.
It's true it requires to pre-specify the fee-bumping key. Though note
the fee-bumping key can be fully separated from the
"vaults"/"channels" set of main keys and hosted on replicated
infrastructure such as watchtowers.
The interface for end-users is very straightforward: if you want to
bump
fees, specify a transaction that contributes incrementally to
package
feerate for some txid. Simple.
As a L2 transaction issuer you can't be sure the transaction you wish
to point to is already in the mempool, or have not been replaced by
your counterparty spending the same shared-utxo, either competitively
or maliciously. So as a measure of caution, you should broadcast
sponsor + target transactions in the same package, thus cancelling the
bandwidth saving (I think).
This theoretical concession seems preferable to heaping more rules
onto
an already labyrinthine mempool policy that is difficult for both
implementers and users to reason about practically and conceptually.
I don't think a sponsor is a silver-bullet to solve all the L2-related
mempool issues. It won't solve the most concerning pinning attacks, as
I think the bottleneck is replace-by-fee. Neither solve the issues
encumbered by the L2s by the dust limit.
If a soft-fork is the cost of cleaning up this essential process,
consideration should be given to paying it as a one-time cost. This
topic merits a separate post, but consider that in the 5 years
leading
up to the 2017 SegWit drama, we averaged about a soft-fork a year.
Uncontroversial, "safe" changes to the consensus protocol shouldn't
be
out of the question when significant practical benefit is plain to
see.
Zooming out, I think we're still early in solving those L2 issues, as
the most major second-layers are still in a design or deployment
phase. We might freeze our transaction propagation interface, and get
short for some of the most interesting ones like channel factories and
payment pools. Further, I think we're not entirely sure how the mining
ecosystem is going to behave once the reward is drained and their
incentives towards L2 confirmations.
Still, if we think we have a correct picture of the
fee-bumping/mempools issues and are sufficiently confident with the
stability of L2 designs, I think the next step would be to come with
quantitative modelling of each resources consumed by fee-bumping (CPU
validation/bandwidth/signing interactivity for the L2s...) and then
score the "next-gen" fee-bumping primitives.
I'm not out to propose soft-forks lightly, but the current
complexity
in fee management feels untenable, and as evidenced by all the
discussion lately, fees are an increasingly crucial part of the
system.
Overall, I think that's a relevant discussion to have ecosystem-wise.
Though there is a lot of context and I don't think there is a simple
way forward. Maybe better to stick to an evolutionary development
process with those mempool/fee-bumping issues. We might envision
two-or-three steps ahead though unlikely more.
Cheers,
Antoine
[0] SIGHASH_GROUP described here
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2021-May/019031.html
and roughly roughly implemented here :
https://github.com/ariard/bitcoin/pull/1
Le jeu. 10 févr. 2022 à 14:48, James O'Beirne via bitcoin-dev
<bitcoin-...@lists.linuxfoundation.org> a écrit :
There's been much talk about fee-bumping lately, and for good reason
-
dynamic fee management is going to be a central part of bitcoin use
as
the mempool fills up (lord willing) and right now fee-bumping is
fraught with difficulty and pinning peril.
Gloria's recent post on the topic[0] was very lucid and highlights a
lot of the current issues, as well as some proposals to improve the
situation.
As others have noted, the post was great. But throughout the course
of reading it and the ensuing discussion, I became troubled by the
increasing complexity of both the status quo and some of the
proposed remedies.
Layering on special cases, more carve-outs, and X and Y percentage
thresholds is going to make reasoning about the mempool harder than
it
already is. Special consideration for "what should be in the next
block" and/or the caching of block templates seems like an imposing
dependency, dragging in a bunch of state and infrastructure to a
question that should be solely limited to mempool feerate aggregates
and the feerate of the particular txn package a wallet is concerned
with.
This is bad enough for protocol designers and Core developers, but
making the situation any more intractable for "end-users" and wallet
developers feels wrong.
I thought it might be useful to step back and reframe. Here are a
few
aims that are motivated chiefly by the quality of end-user
experience,
constrained to obey incentive compatibility (i.e. miner reward, DoS
avoidance). Forgive the abstract dalliance for a moment; I'll talk
through concretes afterwards.
# Purely additive feerate bumps should never be impossible
Any user should always be able to add to the incentive to mine any
transaction in a purely additive way. The countervailing force here
ends up being spam prevention (a la min-relay-fee) to prevent
someone
from consuming bandwidth and mempool space with a long series of
infinitesimal fee-bumps.
A fee bump, naturally, should be given the same per-byte
consideration
as a normal Bitcoin transaction in terms of relay and block space,
although it would be nice to come up with a more succinct
representation. This leads to another design principle:
# The bandwidth and chain space consumed by a fee-bump should be
minimal
Instead of prompting a rebroadcast of the original transaction for
replacement, which contains a lot of data not new to the network, it
makes more sense to broadcast the "diff" which is the additive
contribution towards some txn's feerate.
This dovetails with the idea that...
# Special transaction structure should not be required to bump fees
In an ideal design, special structural foresight would not be needed
in order for a txn's feerate to be improved after broadcast.
Anchor outputs specified solely for CPFP, which amount to many bytes
of
wasted chainspace, are a hack. It's probably uncontroversial at this
point to say that even RBF itself is kind of a hack - a special
sequence number should not be necessary for post-broadcast
contribution
toward feerate. Not to mention RBF's seemingly wasteful consumption
of
bandwidth due to the rebroadcast of data the network has already
seen.
In a sane design, no structural foresight - and certainly no wasted
bytes in the form of unused anchor outputs - should be needed in
order
to add to a miner's reward for confirming a given transaction.
Planning for fee-bumps explicitly in transaction structure also
often
winds up locking in which keys are required to bump fees, at odds
with the idea that...
# Feerate bumps should be able to come from anywhere
One of the practical downsides of CPFP that I haven't seen discussed
in
this conversation is that it requires the transaction to pre-specify
the
keys needed to sign for fee bumps. This is problematic if you're,
for
example, using a vault structure that makes use of pre-signed
transactions.
What if the key you specified n the anchor outputs for a bunch of
pre-signed txns is compromised? What if you'd like to be able to
dynamically select the wallet that bumps fees? CPFP does you no
favors
here.
There is of course a tension between allowing fee bumps to come from
anywhere and the threat of pinning-like attacks. So we should
venture
to remove pinning as a possibility, in line with the first design
principle I discuss.
---
Coming down to earth, the "tabula rasa" thought experiment above has
led
me to favor an approach like the transaction sponsors design that
Jeremy
proposed in a prior discussion back in 2020[1].
Transaction sponsors allow feerates to be bumped after a
transaction's
broadcast, regardless of the structure of the original transaction.
No rebroadcast (wasted bandwidth) is required for the original txn
data.
No wasted chainspace on only-maybe-used prophylactic anchor outputs.
The interface for end-users is very straightforward: if you want to
bump
fees, specify a transaction that contributes incrementally to
package
feerate for some txid. Simple.
In the original discussion, there were a few main objections that I
noted:
1. In Jeremy's original proposal, only one sponsor txn per txid is
allowed by policy. A malicious actor could execute a pinning-like
attack by specifying an only-slightly-helpful feerate sponsor
that
then precludes other larger bumps.
I think there are some ways around this shortcoming. For example:
what
if, by policy, sponsor txns had additional constraints that
- each input must be signed
{SIGHASH_SINGLE,SIGHASH_NONE}|ANYONECANPAY,
- the txn must be specified RBFable,
- a replacement for the sponsor txn must raise the sponsor
feerate,
including ancestors (maybe this is inherent in "is RBFable," but
I don't want to conflate absolute feerates into this).
That way, there is still at most a single sponsor txn per txid in
the
mempool, but anyone can "mix in" inputs which bump the effective
feerate of the sponsor.
This may not be the exact solution we want, but I think it
demonstrates
that the sponsors design has some flexibility and merits some
thinking.
The second objection about sponsors was
2. (from Suhas) sponsors break the classic invariant: "once a valid
transaction is created, it should not become invalid later on
unless
the inputs are double-spent."
This doesn't seem like a huge concern to me if you consider the txid
being sponsored as a sort of spiritual input to the sponsor. While
the
theoretical objection against broadening where one has to look in a
txn
to determine its dependencies is understandable, I don't see what
the
practical cost here is.
Reorg complexity seems comparable if not identical, especially if we
broaden sponsor rules to allow blocks to contain sponsor txns that
are
both for txids in the same block _or_ already included in the chain.
This theoretical concession seems preferable to heaping more rules
onto
an already labyrinthine mempool policy that is difficult for both
implementers and users to reason about practically and conceptually.
A third objection that wasn't posed, IIRC, but almost certainly
would
be:
3. Transaction sponsors requires a soft-fork.
Soft-forks are no fun, but I'll tell you what also isn't fun: being
on
the hook to model (and sometimes implement) a dizzying potpourri of
mempool policies and special-cases. Expecting wallet implementers to
abide by a maze of rules faithfully in order to ensure txn broadcast
and
fee management invites bugs for perpetuity and network behavior that
is
difficult to reason about a priori. Use of CPFP in the long-term
also
risks needless chain waste.
If a soft-fork is the cost of cleaning up this essential process,
consideration should be given to paying it as a one-time cost. This
topic merits a separate post, but consider that in the 5 years
leading
up to the 2017 SegWit drama, we averaged about a soft-fork a year.
Uncontroversial, "safe" changes to the consensus protocol shouldn't
be
out of the question when significant practical benefit is plain to
see.
---
I hope this message has added some framing to the discussion on
fees,
as well prompting other participants to go back and give the
transaction sponsor proposal a serious look. The sponsors interface
is
about the simplest I can imagine for wallets, and it seems easy to
reason about for implementers on Core and elsewhere.
I'm not out to propose soft-forks lightly, but the current
complexity
in fee management feels untenable, and as evidenced by all the
discussion lately, fees are an increasingly crucial part of the
system.
[0]:
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2022-January/019817.html
[1]:
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2020-September/018168.html
_______________________________________________
bitcoin-dev mailing list
bitcoin-...@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev
_______________________________________________
bitcoin-dev mailing list
bitcoin-...@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev
