Dave Harding and Mike Schmidt are joined by Gloria Zhao, Johan Torås Halseth, and Salvatore Ingala to discuss Newsletter #254.
The Bitcoin Optech Podcast and transcription content is licensed Creative Commons CC BY-SA 2.0
Waiting for confirmation #4: Feerate estimation (24:19)
Releases and release candidates
Notable code and documentation changes
Mike Schmidt: So, welcome everybody to Bitcoin Optech Newsletter #254 Recap on Twitter Spaces. It’s Thursday, June 8th, and we’ll do some introductions before jumping into our newsletter. Today we’ll be talking about MATT and CTV and joinpools, along with our limited weekly series about mempool and transaction selection, Waiting for confirmation #4 on feerate estimation. And then we have a release from LND and some PRs to go through, so thank you all for joining.
I’m Mike Schmidt, contributor at Bitcoin Optech and Executive Director at Brink, where we fund Bitcoin open-source developers. And unfortunately, Murch couldn’t make it this week, but fortunately, Dave Harding can. Dave?
Dave Harding: I’m Dave Harding, I’m one of the authors of this week’s newsletter and really looking forward to this discussion.
Mike Schmidt: Gloria?
Gloria Zhao: Hi, I’m Gloria, I work on Bitcoin Core. I’m funded by Brink and I authored the feerate estimation section.
Mike Schmidt: Johan?
Johan Halseth: Hi, I’m Johan, I work on Lightning and Bitcoin stuff. I’m funded by NYDIG.
Mike Schmidt: Salvatore?
Salvatore Ingala: Hi all, I’m Salvatore, I work on the Bitcoin application at Ledger, and I also wrote the initial proposal for the MATT governance.
Mike Schmidt: Awesome. Well, let’s jump into it.
Using MATT to replicate CTV and manage joinpools
Our first news item, and only news item this week, is using MATT to replicate CTV and manage joinpools. So, I’ll take a stab at setting the context for everybody here, and then we can jump in with the experts to dive a little bit deeper. So, we’ve had Salvatore on previously, and we’ve highlighted some of his work, which was MATT, and initially presented at, I think, BTC Azores last year, and it’s a proposal that seeks to enable general smart contracts in Bitcoin. And the original mailing list post outlined several potential new opcodes as part of enabling the MATT proposal, including something called OP_CHECKOUTPUTCOVENANTVERIFY (COCV), which is the piece that would enable some covenant functionality.
I think that opcode has since now been named to OP_CHECKOUTPUTCONTRACTVERIFY, and the idea of a covenant is that instead of being able to spend the coins in my UTXO to any destination, there would be some restriction put on where those coins can be sent, and that can enable certain smart contract use cases. And in Newsletter #249, Salvatore talked about how you can use some of MATT’s opcodes, including OP_COCV, to enable certain vault use cases. And now this week, Johan has taken OP_COCV and shown that you could replicate some of the functionality of the OP_CTV, OP_CHECKTEMPLATEVERIFY proposal, and in a separate post also outlined how some of these MATT opcodes could be used in conjunction with the currently disabled OP_CAT opcode to create a joinpool.
So, to kick things off, Salvatore, what would you add to that framing of the context of MATT before we move forward with Johan’s two different posts?
Salvatore Ingala: Yeah, I think it was a good introduction. I think, apart from specifying the specific opcodes, maybe one thing that could be useful for someone who’s new to the idea is what the opcodes do. If you look at the core of what the proposal does, it’s just to give a way of embedding some data, just a hash let’s say, inside a script, inside a UTXO. So you want to be able to do two things in scripts: one is to access any data which might be embedded in the current input that you’re spending, so give us access to script to this data; and the other thing that you want to do is to be able to constrain the script and the data of the output. Once you have these two primitives, you are able to do some interesting stuff, because you can have data which is dynamically computed, while normally the program you don’t care about dynamically computing that.
But for many contracts, it’s impossible to pre-enumerate all the possible futures, which was a limitation of CTV instead, where you kind of need to be able to enumerate them in advance, otherwise you cannot put those into the script, because the futures will depend on what’s parsed into the witness stack when you spend the coins. And the initial proposal that I made for MATT was not super-formally well-written, because there are some details still to be figured out exactly on how to do them, and so they were in these two opcodes. There could be ways of doing that with just a single opcode that I touched briefly in the last mailing list, and I think I’ll try to experiment soon, but that’s the core idea. Once you can have this way of embedding some data and piping this data through transaction spends, then you can build some stuff on top, and I think that’s all I will add.
Mike Schmidt: And I guess I could segue to one of the things that could be built on top, and, Johan, we can jump into your first post about replicating the functionality of CTV. Do you want to talk a little bit about that?
Johan Halseth: Yeah, I can talk about that, and I can also give some context of how I got to this point as well. So, I started researching using CTV actually as part of the next generation of Lightning Hash Time Locked Contract (HTLC) outputs for commitment transactions, because I wanted to make those more efficient, because now we have all these HTLC outputs on the Lightning commitment transactions, and I wanted to compress this into a single output. So I started using CTV, but quickly got into the same problem as Salvatore here is explaining, that you cannot really predict all possible futures of spending those HTLCs; it becomes an exponential blow up.
So, I started looking into the various ways we could solve this with the proposals on the table, and that piqued my interest because it’s very, very simple really in the COCV and CHECKINPUTCONTRACTVERIFY (CICV) opcodes. You can checksum data in the input and checksum data in the output. And then I started looking at what I could do with this and quickly realized that you can actually, as long as you can compute it in Bitcoin script, you can basically use these opcodes as a way of accessing memory in the input and the output states. So that’s super-interesting that you can find all this. You can get this very, very powerful thing in Bitcoin script with such simple opcodes.
After I did this with both still working on the HTLC proposal, so that’s coming soon, then also, like a simplified version of that again, it’s the coinpool idea. So I published that and also realized that, okay, with this, I can actually very, very simply just do what CTV does with this opcode, because it’s a strict superset really of what you could do. So, that’s how I got to that point. And yeah, a very simple opcode in itself, but very powerful properties you can get with those opcodes.
Mike Schmidt: I noticed a tool that I hadn’t heard before that you used to do some of the demonstrations here, which is Tapsim. Do you want to talk about that real quick, and then we can dig a little bit deeper on both of those?
Johan Halseth: Yeah, for sure. So, Tapsim is a tool I created basically to start playing around with these covenant proposals, because there’s so many proposals out there and they’re all very theoretical. There’s mailing list posts and some of them have BIPs, but not all of them have a real implementation. And if there’s an implementation, it’s in Bitcoin Core, and it’s very hard to access and play around with.
So, what I did was I made a fork of BTCD, which is in Go, which for me personally is much easier to work with. And then I created this tool that just inspects the VM state during script execution and added some UX sugar on top, so it’s easy to step through what the script is, what the stack and old stack and the state of the VM looks like while it’s executing. So, I packaged this into a tool that I’ve been using myself a bunch. It’s very helpful. I didn’t plan on publishing it yet because it’s a bit rough around the edges, but it’s usable and for me it’s been very, very helpful in terms of debugging and playing around with different opcodes.
Mike Schmidt: Dave, do you have questions on these two mailing list posts that we covered this week?
Dave Harding: Sure. Actually, I wanted to start by just saying I really liked Tapsim. It reminds me a lot of an old program by Kalle Alm, called btcdeb, which was very similar. It just built one copy of bitcoind and it was a debugging tool for scripts. And, way back in the day, Kalle implemented, I think it was BIP116, OP_MERKLEBRANCHVERIFY, which was going to enable MAST on Bitcoin before the idea of taproot came around. And so Kalle implemented that in Bitcoin Core and allowed you to go and play with it with btcdeb.
I think this is a really great way for evaluating proposals like this, to just implement them in bitcoind, or in an alternative implementation of the consensus rules, and then plug in some scripts and go and see how they work. I just want to say to Johan, I really, really like going and looking at these demos using Tapsim. I didn’t run them, but I just looked at the code for them, and it just helped me see how these things were working a lot better than reading pages and pages of mailing list posts, so I really appreciated that.
I guess the question I had for Johan was, I’m pretty sure I know the answer to this, but having implemented the primary behavior of CTV using COCV, would you still want CTV and script, if you were actually going to go out and build things like joinpools or vaults or other stuff; would you still want CTV and script or do you think COCV is enough?
Johan Halseth: Well, it’s a very good question. So obviously, I think as I was trying to demonstrate, is that you can basically simulate CTV in using COCV. So, I think it’s strictly more or less a superset of what you could do. But as mentioned in my post as well, CTV can actually be used to compress future spends a lot more than COCV, in some sense at least. So it’s more efficient, so I would say there are different use cases for those two proposals. I’m excited for both of them, but maybe obviously more excited about the COCV because of the powerful features it offers. Yeah, so that’s my answer there.
I just wanted to mention that btcdeb was a huge influence, or something I’ve used before, and had a lot of influence on how I created Tapsim as well, but I found it much easier to work with Go codebase basically, which is why I created Tapsim.
Dave Harding: Yeah, that makes total sense. I think it’s going to get accessible to a lot more people, which is just great for evaluating these proposals. We want to get people with hands on them and building stuff that they actually want to use. I think, as we move into the future and soft forks become kind of harder to do just because of the sheer mass of people who have to have eyeballs on this, have to be confident in it and upgrading their nodes and willing to run this, we just need people to build out these use cases, not just in theory, but as close to practice as we can.
So I guess my question for you and for Salvatore would be, are either of you guys working on getting this maybe into Bitcoin Inquisition, or just getting this implemented up for more experimentation on signets or other test networks?
Johan Halseth: I think maybe Salvatore you…
Salvatore Ingala: Yeah, so the last one or two months, I’ve been a little bit busy with the work I’m doing at Ledger with Miniscript, so I didn’t do a lot of progress. I was actually very happy to see that Johan was doing some more progress on the proposal as well. But definitely, yeah, the main thing that I would like to achieve in the near future is to actually have fully formally defined opcodes that have all the missing features that are in a way trivial, like being able to inspect the amounts, but there is some design space to fill there, let’s say. And so, there was something that I mentioned in the proposal following up on Johan’s post, which is he suggested for some reasons to actually make the two opcodes kind of symmetric so that you can have the same semantics that you have for inputs as you can have for outputs, because in the initial proposal the CICV only works for the current input and it’s a little bit simpler opcode, but they are kind of doing the same thing.
So, one could think of either making them symmetric or even coming up with just a single opcode that can work on either an input or an output, and yeah, it would be interesting to try to do that and see if the code gets too complicated, it makes sense to have two opcodes instead because it’s a little bit simpler programming. And so, that would be something that I want to experiment in the near future.
Just connecting to what was mentioned before about CTV, because in terms of functionality, CICV together with COCV, is a strict superset of what CTV enables. On one hand, one could think, okay, then we don’t need CTV; on the other hand, there are many cases where we can show that CTV is a lot more efficient, and actually that was one of the things that I wanted to show in the post with emulating OP_VAULT with MATT opcodes, because that’s one case where you can see that CTV makes the construction a lot more efficient.
So, since it doesn’t add any more powers to the script and the opcode is still very simple, I think it’s a no-brainer that if you’re happy merging MATT opcodes, then adding CTV as an optimization is a very small amount of complexity added, but it makes it a lot more efficient for some interesting use cases. So, I think it will make a lot of sense to include CTV in the proposal as well. I don’t know if I missed any of the questions.
Mike Schmidt: No, I think you addressed that. Dave mentioned Inquisition. You guys obviously have mailing list posts which are garnering feedback. I know that there’s also a Contracting Primitives Working Group, and then there’s probably some offline discussions as well. I’m curious as to, Johan, both with your example post as well as the broader MATT proposal, what has feedback been from the community; how would you summarize the community’s temperature check on this proposal and these related demos?
Johan Halseth: Well, what I would say is really that the original MATT post that Salvatore did was maybe a bit hard to grok for many; it has a lot. So, it’s very cool that you can do arbitrary computation using these opcodes, but maybe what I tried to do is to get to the heart of it, what it’s like, what can you distill these opcodes into? And what I found from doing this experiment is, basically it gives you access to some memory in the output and some memory in the input, and that’s very powerful, as Salvatore shows in his post as well. So I think maybe that’s the feedback I got is that, “Oh, okay, now I understand what these opcodes are doing”, instead of having this very, very powerful and maybe complex example of what you can do with this, trying to get into the really simple way of explaining these opcodes and how you from there can build out to something much, much more powerful.
That’s also kind of why I created Tapsim as well, so you can easily step through these scripts and so that you can understand what’s going on and how you can build from there. And also, without announcing Tapsim in any large way, people that have looked at it have given me the feedback that, “This is super, super useful, and it’s something I’ve really wanted for a long time”. So I’m very happy to hear that as well.
Mike Schmidt: Salvatore, what is your feeling on how the community’s reacted to the MATT proposal?
Salvatore Ingala: Yeah, I agree with Johan. My initial posts were a little bit hard to decode for many people, and the fact that I was not able to show code initially because I thought, if Tapsim was available when I wrote the proposal, probably it would take me less time to write some code that I can show to people, instead of actually writing a functional test in Bitcoin Core. So, that took me quite some time to find the energy and the commitment to actually put the many days in a row into this project and have some working code. And so, yeah, I’m definitely looking forward to experiment more with Tapsim as well.
I think from past experience with soft forks and past experience with discussions on covenants, there’s a little bit of, let’s say, PSTD from the Bitcoin community, where people are scared of covenants for potential risks that are being discussed, but not really materialized in any concrete scenarios of what are the dangers of covenants. And so, seeing the other side where we see what are the useful things that we can build with covenants might probably help to level up the discussions. And basically, we want more people thinking about these ideas and reasoning about them and thinking, of course, if there is serious concerns about dangers that could come up with covenants, that’s something that more research will help to figure out as well.
But yeah, my impression is that actually the more people will think about these things, the more we will realize that the scale of these things was a bit overblown, while there are many interesting things that we can build with them. And so actually, it’s not obvious at all the risks to potential, like the game theory of Bitcoin might materialize, and it could even be the opposite; by enabling more applications to be built on top of Bitcoin, more smart contracts that could be even privacy solutions or these coinpools and more reasons for people to pay fees on the base layer, that could actually even improve the game theory. It’s not obvious at all that there are dangers that are in the negative in terms of game theory, it could be an improvement as well.
So, definitely my hope is that we get more people and more minds on this problem, on how we extend Bitcoin smart contract security in a secure way, and I’m quite optimistic about the potential of this kind of approach on improving Bitcoin.
Mike Schmidt: Dave, I’m curious, we have a lot of these kinds of proposals, and I’m thinking as a general Optech Newsletter reader or a listener to the show, what would be a useful way to think about where we, as a Bitcoin community, are at with all these proposals? Is there a particular way that you think about things that you think would be useful for others to be aware of in terms of these types of proposals? It seems like every other month, there’s something interesting, a proposal or innovation that is involving either covenants or new opcodes; how do you think about it?
Dave Harding: Well to a certain degree, I just try not to think about it because like you said, there’s just so much going on, so I write the newsletter and then I just run away and put my head in the sink! But I think that we have a lot of these proposals that are significantly overlapping in the functionalities that they enable, sort of like how we see in this week’s post that COCV can emulate part of CTV, you know, it can emulate it all, I guess, maybe. But there’s trade-offs; between all these proposals, there’s trade-offs. We have a lot of ideas, they’re significantly overlapping the functionality that they can perform, but there’s trade-offs.
So CTV, like Johan said, is going to be more efficient in some cases, but it’s less flexible in other cases. You have these trade-offs, and I’m going to show my own idea here, but a little bit over a year ago, I posted to the mailing list the idea of an automatically reverting soft fork. So after, say, five years, we would activate a soft fork, and after five years, it would un-activate. Those consensus rules that we added in the soft fork would no longer be enforced at the consensus level. Some people didn’t like that idea; Matt Corallo in particular did not like that idea, and that’s fine.
The more I see these proposals that are technically sound and they have a minimum consensus footprint, the technical complications are not great, the more I feel like that might be a reasonable direction to go, is just find a bunch of these proposals, test the heck out of them on Inquisition, test the heck out of them in adding them to Bitcoin Core proper, and then activate a bunch of them and see what happens, see what opcodes people use over a period of five years, see what people build on them, and then let the ones that aren’t being used or that are being used poorly maybe just un-activate, fall out of consensus use, and then keep the ones that work really well.
So, that’s kind of where I’m leaning on a path forward, again that’s a controversial path forward, but just this idea of maybe we should think about this stuff as, let’s grab all the goodies and not worry now about trying to find which one is in some criteria best.
Mike Schmidt: I think the idea of sidechains originally was along this line, which is let these ideas proliferate and see what works. Obviously that hasn’t come to fruition in the Bitcoin community in a trustless way. You have things like Liquid that are doing some of these sorts of experiments, but yeah, I do remember that post and I’m not sure if we’ll end up there, but it’s an interesting route to get some of these things operationalized. Salvatore or Johan, do you have any closing words or call to action for the community?
Johan Halseth: Yeah, I can just add that I think there’s been a lot of covenant proposals. I think many of them could achieve the same as COCV does. But the nice thing about it, in my opinion, is the simplicity of the opcode. It’s very, very easy to reason about what it does, and still you can build all these super-interesting use cases, which is why I’m very excited about the proposal.
Salvatore Ingala: So, since you asked for closing words, I’ll try to do a pitch for the proposal, which is I think soft forks are more dangerous than covenants, and so adding a more general covenant will reduce the need for future soft forks. So, I think that’s one reason why I think simple opcodes that are general could be an interesting direction, and that’s what I’m trying to do with MATT.
Mike Schmidt: Well, thank you both for joining. You’re welcome to stay on as we go through the rest of the newsletter, but if you have something that you need to get to, you’re free to drop. Thank you for joining us.
Salvatore Ingala: Thank you. My pleasure, as usual.
Johan Halseth: Thank you.
Waiting for confirmation #4: Feerate estimation
Mike Schmidt: Next section from the newsletter was from our limited series on transaction relay and mempool inclusion and policy, titled Waiting for confirmation #four: Feerate estimation. So luckily, Gloria has returned this week to join us to talk more about this weekly series. Last week’s topic was all about techniques you could use to minimize your transaction fees, including things like modern output types, coin selection considerations, payment batching, and more. And this week, the topic is transaction feerate estimation. So maybe to lead in, Gloria, how do we think about what a transaction’s feerate should be?
Gloria Zhao: Yeah, so again, I kind of want to preface this with the hope for this series is to start conversations about how we can make things better for all the users of Bitcoin. And so it may seem weird that I’ve dedicated an entire post to feerate estimation, but I thought it’s a space that’s really ripe for innovation, and there’s a lot of work we can do to make things better, and it’s very multifaceted and interesting. So hopefully, someone’s been nerd sniped by this post.
But yeah, so the question was, how do we think about feerates? So, feerate estimation, the goal is to translate a target timeframe for which you want your transaction to get confirmed to a minimal feerate that you should pay. So obviously, if you pay, I don’t know, 1 Bitcoin on your transaction, you’re going to probably get confirmed pretty quickly. But you want to pay as little as possible, of course. And the main point of this post is to say that fee estimation is really hard for a few reasons: (1) the supply is unpredictable; (2) the demand is unpredictable; and (3) the information isn’t always public to you and can sometimes be gamed, really.
So an overarching idea for this series so far has been to create this public, efficient auction for block space. So, when I talk about information not being public or being gameable, for example, if we only looked at the fees of transactions included in blocks, basically the miners can put artificially high feerate transactions in their blocks to drive up feerates, if we had a very silly, naïve theory estimator that only did that.
Anyway, so back to supply and demand being unpredictable. Blocks don’t come every ten minutes exactly, and that is a really sucky, UI/UX problem for Bitcoin, but it’s part of what makes feerate estimation hard. For example, if the merchant gives you 30 minutes to send the payment before they give the goods to you, but the next block takes 45 minutes to be found, you can get screwed. But of course, sometimes you find three blocks in a row in the span of a few minutes, right? And these kinds of things are, I guess, a UX problem. It’s not just, how do we write a piece of software that’s really good at estimating things?
But of course, there’s also the other side of the coin, that demand is very unpredictable. Of course, we all know that there are huge fluctuations in volume and sometimes you can get blindsided by that. Sometimes your transaction can fall out of the mempool and that’s a whole other UI/UX problem. And yeah, so fee estimation is really hard. I talked about two existing fee estimators that I’m aware of: one is mempool space, which I think is pretty accurate, I think a lot of people use mempool space, I imagine. And their approach, hopefully if there’s someone on the call, you can correct me if I’m wrong here, is essentially you have a really good view of what’s in miners’ mempools and you can almost just calculate what’s going to be in the next n blocks. So, you take the mempool and you run the block assembler algorithm and you’re like, “All right, to get into the nth block, I can literally just build n blocks and then tell you what the feerates of those transactions are”.
Of course, I imagine that to do this kind of accounting for other transactions that might come in in the next timeframe is to build with a decreased block size to account for like, okay, there’s other people that might send transactions at these feerates and they’ll fit into these empty spaces in the blocks that we’re projecting. And so that’s one. It very much relies on all of your information being accurate, like what’s in your mempool actually being what miners are going to mine, and maybe that’s very appropriate for something like mempool space, where you have a lot of nodes, you have a very good idea of what transactions are in miners’ mempool, and maybe even have a good idea of what out-of-band fees they might be accepting. So that’s one pretty accurate, as far as I know, fee estimation algorithm.
Then there’s also Bitcoin Core’s, which tries to sidestep the problem of non-public information by not trying to record it. So Bitcoin Core’s fee estimation algorithm is looking at transactions as they come into your mempool and then recording when that happens and then recording when you then see them confirmed, and then it’s historical-data-based. Of course, if you have miners putting artificial high fee transactions in their blocks, then that won’t impact your fee estimation since you won’t see them that time. So, that’s trying to avoid this gameability aspect of potentially trying to just use the information available to you.
So, that’s highlighting two feerate estimation algorithms that I’m aware of. I think both of them have room for improvement; both of them are perhaps more appropriate for the user, the piece of software that they’re designed for, like Bitcoin Core hopefully is just an individual user running their Raspberry Pi node or their laptop node, or trying to have an independent fee estimator and not relying on centralized APIs. And hopefully, Bitcoin Core gives them a nice, trustless fee estimation based on public information. And then mempool space has access to more information, and hopefully can give you a more precise result, but perhaps requires you to have a very, very accurate idea of what miners are going to mine.
So, yeah, it’s a multifaceted, fascinating problem. There’s a mixture of UI/UX, maybe there’s room for data-based, intelligent modeling of forecasting demand, put some data scientists on the case. I feel like at least Bitcoin Core, I feel like there’s a lot of room for improvement, and hopefully someone comes and looks at our fee estimator and opens a PR or something; that’s been my goal.
Mike Schmidt: Gloria, you mentioned just now and also in the post that forecasting block demand space is ripe for exploration, and you mentioned some examples of that. You mentioned data science, but you also mentioned in the post about certain activity patterns that may occur or certain times of the day or business hours or external events that can mess with feerates. I know that there was the BitMEX withdraw at 8.00am Eastern time every day. I think that’s gone now, but I think that was a big, known thing. Are you aware of anybody working on some supplemental, external events or otherwise ways of doing fee estimation?
Gloria Zhao: I am not at all. I mean, Murch will tweet about it, and he’ll be able to point out patterns, and I’m sure we all do a little bit of thinking about it. I don’t know of anybody who’s actually, I don’t know, building a model or… I mean, surely maybe we can start with like a hackathon project or something to start plugging in data and seeing. Well, I think what we should do really is first try to build a framework for feerate estimation accuracy. Oh, I think Josie, Josie Bake, has built a really nice IPython notebook that draws a few graphs and has some tooling for like parsing at least the Bitcoin core fee estimation database. Those are the only people that spring to mind.
I’m very sorry if there is totally someone who’s just dropped a paper on this, for example. Anyone that knows anything, please ping me. I feel like this is something that is almost a low-hanging fruit, almost. We could definitely improve, or there’s some very obvious things that we can do to try to get started, like making things better. Just, you know, anyone listening interested, please do something.
Mike Schmidt: Dave, a lot of interesting points here from Gloria. Do you have anything to augment, or questions for Gloria on this topic?
Dave Harding: No, I think those are very well-written posts. I don’t really have any questions. Related to your previous question about people working on better feerate estimation, a few years ago, Kalle Alm, who we already mentioned in this podcast, he had a project called the Mempool Monitoring Project, where basically he just recorded every transaction that hit his node, when it hit his node, when it got confirmed, and just details like that in a historical database to provide the information for future research efforts. I don’t know where that went, but it’s the kind of thing that would be useful.
He also had the idea of, one of the things we have with feerate estimation is that it’s kind of self-recursive in the sense that when feerates go high, the feerate estimator tends to return high feerate estimates, everybody starts paying higher fees. And then when there’s a small spike in demand or a loss in supply, feerates go higher, the feerate estimator turns even higher fees, and so it just keeps increasing, increasing, increasing, and it falls off a cliff all of a sudden when demand drops just a little bit, and whatnot. And so, Kalle also had the idea, I don’t know if it was his idea, but he was working on a test implementation of having feerates set to the lower bound of what was currently in the mempool versus what the statistical-based feerate estimator does, so kind of a synthesis of the two approaches that Gloria describes in her post right now, just synthesizing this and taking the minimum of that, and return that as the feerate for transactions that could be easily fee-bumped with RBF.
I guess that would bring me to one other point I’d like to make, is that if you don’t have feerate estimation, basically what you’re stuck doing is setting a transaction at a very low feerate, waiting some amount of time for it to confirm, and then RBF fee-bumping it. You just keep doing this until all of a sudden it gets confirmed. This is, as far as I know, your only alternative for trustless fee management to running your own node with a mempool, is to just iteratively RBF fee-bump your transactions until it gets confirmed. That’s kind of a bad UX in the sense that it takes a long time before you get to the rate of the current mempool. You’re probably still going to overpay too by some amount, so I think the rate estimation is an interesting topic. It’s good to explore, it’s good to see how far we can get at making good estimates considering, like Gloria said, the unpredictability of the supply and demand.
I guess I could throw one more point here, is a few years ago, some researchers posted a paper to Bitcoin-Dev suggesting we change the way the auction works. The idea was that you overpay your fees, but you get a refund, so miners have to claim every transaction in a block at a consistent feerate; every transaction in a block pays the same feerate, but you can overpay your feerate and get a refund of the difference between what you paid and what miners claimed. The cheapest feerate claimed in a block would be the feerate that applies to all transactions in a block. Unfortunately, this does not work with Bitcoin’s UTXO model, at least not very well, you’d have to make a horrible hack of it. But man, I think that would be a great improvement. So that’s it for me.
Mike Schmidt: So how would that work then? Transaction fees would be collected by the miner and when the block is mined, essentially users who are transacting and who have transacted over that average, or whatever that cut-off is, would get paid out in the coinbase, sort of like a mining pool kind of thing, to get their refund?
Dave Harding: So, it was designed kind of in mind of the Ethereum account model. So in Ethereum, you have an account with a balance. So, if you think of the way it would work in Ethereum is that all the transactions in the block would say, okay, you can use up to 1F of my balance to pay the transaction fee to get my transaction in the block. And the miners would take all the transactions they could. Again, they would pull them by what would be most profitable for the miner, and they would choose whatever the lowest paying transaction they include in the block, and they would take 100% of that transaction’s allocated fee. But the highest transaction, they might only take 5% of its fee, and the rest would stay in the Ethereum account.
Now in Bitcoin, we don’t have accounts. That’s what makes this a really sticky proposal, is that you’d basically have to have the miner, like you say, in the coinbase transaction issue a bunch of outputs for every transaction in that block. So if you had 4,000 transactions, the coinbase transaction would have to have 4,000 outputs at about 40 bytes each, which is just insane. I think Mark Friedenbach had a proposal for how to do this, and it’s something that becomes a little bit more possible with covenants because covenants can kind of get us towards an account model. If you want that, it’s bad for privacy, it has all these problems applying to Bitcoin. I just wanted to mention it in case somebody’s listening and can think about a really, really clever way to make that possible, because it just allows you to say, “This is the highest feerate I’m willing to pay for my transaction. Get it done”, and yeah, that would be nice.
Mike Schmidt: Just pipe all the excess funds into a joinpool, there you go!
Dave Harding: You solved it, Mike, yeah!
Mike Schmidt: Gloria, anything before we move on?
Gloria Zhao: Next week’s is about DoS.
Mike Schmidt: Looking forward to it. Gloria, thanks for joining us. I know you are at an event. If you need a drop, thanks for joining; if not, happy to have you on.
Gloria Zhao: Cool. Thanks for the not dox!
Mike Schmidt: Next section of the newsletter is Releases and release candidates, and we just had one here, which is LND 0.16.3-beta, which is noted as a maintenance release. There’s a few different bug fixes and other performance fixes here. The one that I thought was notable was that LND bumps the version of its underlying BTC wallet library, which is a library that LND uses for its wallet functionality. The reason that it’s bumping the version of that library was that there was a performance issue in this BTC wallet library that caused CPU usage to spike when performing certain mempool related operations.
So, there was an optimization put in BTC wallet that added a cache, which improves this performance and solves the CPU issue, which also then obviously affected LND. So, by bumping that version of that library, LND is no longer susceptible to these CPU spikes. Dave, was there anything else notable from this LND release that you wanted to know?
Dave Harding: No, we just added a note here that I think the fix that you’re talking about was related to their mempool watching logic. So, they’re actually looking at the mempool now to find transactions, Lightning transactions, that have gone onchain for one reason or another, so they can resolve them quicker than waiting for them to confirm. And so, this is a speed up between the time that a channel closes, and you can start using your funds for something else. So, they’re working on that, they’re making LND a little bit faster its users, so that’s nice.
Mike Schmidt: Next section of the newsletter is Notable code and documentation changes. I’ll use this opportunity to solicit any questions that the audience may have; feel free to request speaker access, or comment on this Twitter thread and we can get to your question at the end of the newsletter.
Bitcoin Core #26485
First PR is Bitcoin Core #26485, and this is a change to Bitcoin Core RPC methods and how they’re called. So, in order to call certain Bitcoin Core RPCs previously, there was a parameter called options, which was required to pass certain parameters, and that particular options parameter was a big, nested JSON object. And with this PR, that nested JSON option still can be used, but there’s an additional way that you can now call certain RPCs that required that previously, and you can actually use name parameters instead of a big, old JSON object. So, this adds some flexibility for applications calling RPC, which seems nice. Dave, is there a particular use case that you think was in mind here for adding this additional way to call RPC?
Dave Harding: So, I use bitcoin-cli a lot, and one of the really big pains of using it is that you spend half your life quoting stuff, because JSON in a shell is just not fun. You have to put the outer thing in single quotes, you have to put all of your parameters in double quotes, and then you’re putting curly braces around everything, or square braces around arrays. And so, anytime somebody comes along with an idea like this, it’s just a small code change and allows you to stop quoting everything and stop – the braces aren’t so bad, but just simplifying stuff. I think that’s what it is. It’s just Bitcoin programmers are spending a lot of time using Bitcoin Core and they’re spending a lot of their lives just quoting things and then dealing with the problems that happen, the weird error messages you get in the shell when you misquote stuff.
So, I think that’s what’s happening here, is they’re just making it a little bit easier to use, especially since it’s something they use every day, tons of times every day.
Mike Schmidt: Eclair #2642, adding a closedchannels RPC. So Eclair’s adding this, we noted that a few newsletters ago, we covered CLN adding a similar RPC, and I thought it was interesting in digging into this PR, the first comment I think was in response to someone opening this PR, was t-bast saying, “Can you explain why you think this is useful?” And the person opening the PR said, “First of all, it’s going to make me as famous as Rusty Russell is via the Bitcoin Optech Newsletter”, since we had covered this CLN RPC previously. So, I thought that was kind of funny to get an Optech shoutout in the middle of this PR for the reason of opening this.
But the real reason the person gave was, the only way to figure out which peer closed the channel and what the cause and what the balance of the channel was, in addition to other information, was this person was using a Lightning explorer, and mentioned Amboss and some other ones, and noted that that is a privacy leak by using these explorers to find out some of the information that was already in his node, he just didn’t have an easy way to access it. And then t-bast added the Optech Make Me Famous label to the PR, and now this person is famous! Dave, any comments?
Dave Harding: No, except that I laughed at that remark too on the comment!
Mike Schmidt: LND #7645, making sure that any user-provided feerate in certain RPC calls is no less than a “relay feerate”. And so there’s OpenChannel, CloseChannel, SendCoins, SendMany RPCs in the LND node that are functions that a user can provide a feerate. And previously, there was no check on what that feerate would be, and thus it could be below either a relay fee or the minimum mempool fee. And so this change actually puts a check in and will provide an error if the “relay feerate” may seem slightly different than things depending on the back end.
So, there’s multiple backends, and so for bitcoind, this relay feerate that they’re referring to is actually whatever is greater, the relay fee or the minimum pool fee, and I’m not exactly sure what BTCD equivalent is for that. But there’s at least some checking going on for user-provided feerates to make sure that they’re adequate, I would say. Dave?
Dave Harding: Yeah. So, this is just something that you need to do when we’re having what some people call full mempools. So when bitcoind, and I think BTCD now as well, their mempools have a maximum size, and when they get full, they start dropping transactions at the bottom. And they can use, I think it’s BIP133, but it could be BIP130, I get this confused all the time. There’s a message that nodes can send that says, “This is the current minimum fee that I’m accepting”. And so when you’re at that point that the node is rejecting transactions below the minimum feerate, if you try to send a transaction below that, it’s just not going to propagate at all. It’s going to hit the first node and that node’s going to drop it, and it’s going to go nowhere in the network.
So, I think that’s all that’s happening here, is saying that your local node, your backend, has a minimum, and maybe all its peers have a minimum, I don’t know exactly how this is implemented, and if you’re trying to send a transaction into that, it’s just not going to work, so try with a higher fee. So, this is a nice fix, say, an edge case that users might bump into, improvement.
Next PR is LND #7726, making a change that it will always spend all HTLCs paying the local node if a channel needs to be settled onchain, even if it might cost more in transaction fees to sweep them than they are worth. And this is in contrast to a PR from Eclair from last week, that added logic not to claim an HTLC that would cost more in fees to claim than it’s worth. And noted in this PR is, “In the future, we’ll start to make more economical decisions about if we go to chain at all for small value HTLCs”. So, it’s interesting to see sort of diverging PRs with regards to claiming HTLCs on chain. We have one saying we’ll do it economically, and one saying we’ll do it no matter what. Dave, thoughts on this PR?
Dave Harding: This one was a bit confusing for me to disentangle when I was writing it up, to figure out exactly what was going on. It said it’s a small code change to LND, the actual PR itself, that you kind of have to dig in to figure out what it’s working on. But the idea there is they already had logic that wouldn’t try to claim an HTLC if it looked like it was uneconomical. But there were a bunch of cases where it could be economical because other things depended on it. And so they decided to make the call to always claim it, even though it might not be economical, just because they don’t want to lose out on those occasions where it could be a larger amount of funds depended on it that would definitely make it economical.
What they’re going to do is go back and just add cost accounting to their program to be able to figure this stuff out so they can always make economical decisions. That just means sometimes an HTLC on LN is going to be uneconomical. You make a transaction on LN two weeks ago when fees were lower, and today fees are higher and it’s just not worth claiming that money. And that’s just an interesting dynamic that we need to think about when designing Lightning software in the future, is what do we do about high-fee environments and small HTLCs not necessarily making sense.
So, I think this is an interesting, I don’t want to say it’s an open problem in LN; it’s not. I think there’s solutions out there, but just figuring out which solutions we’re going to adopt and what trade-offs we’re going to make.
Mike Schmidt: Last PR this week is LDK #2293, and the motivation here sounds like LND sometimes stops responding, leading to channels being forced closed. And the common solution for LND node operators is to restart their node or reconnect to their peers, and just try to start from some sort of a fresh state. And so LDK is mitigating this interoperability issue by disconnecting unresponsive peers after a period of time to, I guess, also cause that fresh state to be forced. And there is actually a related LND PR, which was recently merged as well. So, Dave, this sounds a little funky. It’s like the old, if it’s not working, turn it off and turn it back on again.
Dave Harding: That was exactly what I was going to say! But yeah, actually it seems like a decent solution to actually just have in your codebase, is if your peer isn’t responding, but you think they could be responding and you think there’s just something going on there, just start and stop. And the reason that works in software so often is that it just brings you back to that initial state. Programmers are a lot better about reasoning about initial states and finding bugs in initial states than they are in a piece of software that’s been running for days or weeks and is in just some state that is very rarely reached in the code, that the programmer has not thought about.
So, just in the case of unresponsiveness, just going back to that initial state, I think it’s a good solution. I don’t see any problems here. It’s an ugly hack, I guess, but if you can get over the ugliness, it seems pretty functional to me.
Mike Schmidt: We have one question, Dave, that I’ll direct at you that came from Chad Pleb. This person asks, “Is it possible to take historical blocks and say something about how efficient transaction fees were in order to approximate how much was overpaid or wasted, or using some sort of standard deviation like metrics?”
Dave Harding: So, Chad, what you need is you need to know when that transaction was first relayed and when it got entered to a block. So, you just can’t grab blocks themselves because you’re missing the information of when that person sent that transaction. What I think you might be looking at is, say, the difference between the highest feerate for a transaction in a block, and the lowest feerate transaction in a block. And if they’re all close together, then you could call that efficient; and if they’re very far apart, you could call that inefficient.
That works pretty well, if you assume that nobody was paying fees out of bound and that miners weren’t doping their own blocks. So miners can add transactions to their blocks that pay any feerate at a very low risk to the miner of somebody else grabbing that fee. So, just to be really specific here, if I’m mining a block, I can include a transaction that pays 1 BTC in fee to myself. And that makes it look like I claimed a lot of fees, but in reality I just wasted block space. So if you ignore that, if you ignore people paying fees out of bound, so people paying fees directly to miners through a credit card or through another system, then you can do that.
I don’t know exactly what information that gives you except, like you said, how inefficient it was compared to optimum. Again, I think what you really want to know is what people paid, when they paid, and how long it took them to get into a block. I guess you also want to know how long they wanted to wait, which is unknowable from public data. Sorry, it was a bit of a rambling answer. I hope that kind of answered your question at least.
Mike Schmidt: Thanks, Dave. I think that’s it for questions and that’s it for the newsletter this week. I’d like to thank Gloria for joining us and Johan and Salvatore, and to my co-host, Dave, for coming in for Murch. Thanks, Dave.
Dave Harding: Always a pleasure, Mike.
Mike Schmidt: Thanks everybody, cheers.