Thanks Christian,
On 2/13/23 7:32 AM, Christian Decker wrote:
Hi Matt,
Hi Joost,
let me chime in here, since we seem to be slowly reinventing all the
research on reputation systems that is already out there. First of all
let me say that I am personally not a fan of reputation systems in
general, just to get my own biases out of the way, now on to the why :-)
Reputation systems are great when they work, but they are horrible to
get right, and certainly the patchworky approach we see being proposed
today will end up with a system that is easy to exploit and hard to
understand. The last time I encountered this kind of scenario was during
my work on Bittorrent, where the often theorized tit-for-tat approach
failed spectacularly, and leeching (i.e., not contributing to other
people's download) is rampant even today (BT only works because a few
don't care about their upload bandwidth).
First of all let's see what types of reputation system exist (and yes,
this is my very informal categorization):
- First hand experience
- Inferred experience
- Hearsay
The first two are likely the setup we all are comfortable with: we ourselves
experienced something, and make some decisions based on that
experience. This is probably what we're all doing at the moment: we
attempt a payment, it fails, we back off for a bit from that channel
being used again. This requires either being able to witness the issue
directly (local peer) or infer from unforgeable error messages (the
failing node returns an error, and it can't point the finger at someone
else). Notice that this also includes some transitive constructions,
such as the backpressure mechanism we were discussing for ariard's
credentials proposal.
Ideally we'd only rely on the first two to make decisions, but here's
exactly the issue we ran into with Bittorrent: repeat interactions are
too rare. In addition, our local knowledge gets out of date the longer
we wait, and a previously failing channel may now be good again, and
vice-versa. For us to have sufficient knowledge to make good decisions
we need to repeatedly interact with the same nodes in the network, and
since end-users will be very unlikely to do that, we might end up in a
situation were we instinctively fall back to the hearsay method, either
by sharing our local reputation with peers and then somehow combine that
with our own view. To the best of my knowledge such a system has never
been built successfully, and all attempts have ended in a system that
was either way too simple or is gameable by rational players.
In lightning we have a trivial solution to this - your wallet vendor/LSP is already extracting a fee
from you for every HTLC routed through it, it has you captive and can set the fee (largely)
arbitrarily (up to you paying on-chain fees to switch LSPs). They can happily tell you their view of
the network ~live and you should generally accept it. Its by no means perfect, and there's plenty of
games they could play on, eg, your privacy, but its pretty damned good.
If we care a ton about the risks here, we could have a few altruistic nodes that release similar
info and users can median-filter the data in one way or another to reduce risk.
I just do not buy that this is a difficult problem for the "end user" part of the network. For
larger nodes its (obviously, and trivially) not a problem either, which leaves the "middle nodes"
stranded without good data but without an LSP they want to use for data. I believe that isn't a
large enough cohort to change the whole network around for, and them asking a few altruistic (let's
say, developer?) nodes for scoring data seems more than sufficient.
I also object to the wording of penalizing nodes that haven't been as
reliable in the past. It's not penalizing them if, based on our local
information, we decide to route over other nodes for a bit. Our goal is
optimize the payment process, chosing the best possible routes, not
making a judgement on the honesty or reliability of a node. When talking
about penalizing we see node operators starting to play stupid games to
avoid that perceived penalty, when in reality they should do their best
to route as many payments successfully as possible (the negative fees
for direct peers "exhausting" a balanced flow is one such example of
premature optimization in that direction imho).
Yes! Very much yes! I hate this line of thinking.
So I guess what I'm saying is that we need to get away from this
patchwork mode of building the protocol, and have a much clearer model
for a) what we want to achieve, b) how much untrustworthy information we
want to rely on, and c) how we protect (and possibly prove security)
against manipulation by rational players. For the last question we at
least have one nice feature (for now), namely that the identities are
semi-permanent, and so white-washing attacks at least are not free.
And after all this rambling, let's get back to the topic at hand: I
don't think enshrining the differences of availability in the protocol,
thus creating two classes of nodes, is a desirable
feature. Communicating up-front that I intend to be reliable does
nothing, and penalizing after the fact isn't worth much due to the
repeat interactions issue. It'd be even worse if now we had to rely on a
third party to aggregate and track the reliability, in order to get
enough repeat interactions to build a good model of their liquidity,
since we're now back in the hearsay world, and the third party can feed
us wrong information to maximize their profits.
Regards,
Christian
Matt Corallo <lf-li...@mattcorallo.com> writes:
Hi Joost,
I’m not sure I agree that lightning is “capital efficient” (or even close to
it), but more generally I don’t see why this needs a signal.
If nodes start aggressively preferring routes through nodes that reliably route
payments (which I believe lnd already does, in effect, to some large extent),
they should do so by measurement, not signaling.
In practice, many channels on the network are “high availability” today, but
only in one direction (I.e. they aren’t regularly spliced/rebalanced and are
regularly unbalanced). A node strongly preferring a high payment success rate
*should* prefer such a channel, but in your scheme would not.
This ignores the myriad of “at what threshold do you signal HA” issues, which
likely make such a signal DOA, anyway.
Finally, I’m very dismayed at this direction in thinking on how ln should work
- nodes should be measuring the network and routing over paths that it thinks
are reliable for what it wants, *robustly over an unreliable network*. We
should absolutely not be expecting the lightning network to be built out of
high reliability nodes, that creates strong centralization pressure. To truly
meet a “high availability” threshold, realistically, you’d need to be able to
JIT 0conf splice-in, which would drive lightning to actually being a credit
network.
With reasonable volume, lightning today is very reliable and relatively fast,
with few retries required. I don’t think we need to change anything to fix it.
:)
Matt
On Feb 13, 2023, at 06:46, Joost Jager <joost.ja...@gmail.com> wrote:
Hi,
For a long time I've held the expectation that eventually payers on the
lightning network will become very strict about node performance. That they
will require a routing node to operate flawlessly or else apply a hefty penalty
such as completely avoiding the node for an extended period of time - multiple
weeks. The consequence of this is that routing nodes would need to manage their
liquidity meticulously because every failure potentially has a large impact on
future routing revenue.
I think movement in this direction is important to guarantee competitiveness
with centralised payment systems and their (at least theoretical) ability to
process a payment in the blink of an eye. A lightning wallet trying multiple
paths to find one that works doesn't help with this.
A common argument against strict penalisation is that it would lead to less
efficient use of capital. Routing nodes would need to maintain pools of
liquidity to guarantee successes all the time. My opinion on this is that
lightning is already enormously capital efficient at scale and that it is worth
sacrificing a slight part of that efficiency to also achieve the lowest
possible latency.
This brings me to the actual subject of this post. Assuming strict penalisation
is good, it may still not be ideal to flip the switch from one day to the
other. Routing nodes may not offer the required level of service yet, causing
senders to end up with no nodes to choose from.
One option is to gradually increase the strength of the penalties, so that
routing nodes are given time to adapt to the new standards. This does require
everyone to move along and leaves no space for cheap routing nodes with less
leeway in terms of liquidity.
Therefore I am proposing another way to go about it: extend the
`channel_update` field `channel_flags` with a new bit that the sender can use
to signal `highly_available`.
It's then up to payers to decide how to interpret this flag. One way could be
to prefer `highly_available` channels during pathfinding. But if the routing
node then returns a failure, a much stronger than normal penalty will be
applied. For routing nodes this creates an opportunity to attract more traffic
by marking some channels as `highly_available`, but it also comes with the
responsibility to deliver.
Without shadow channels, it is impossible to guarantee liquidity up to the
channel capacity. It might make sense for senders to only assume high
availability for amounts up to `htlc_maximum_msat`.
A variation on this scheme that requires no extension of `channel_update` is to
signal availability implicitly through routing fees. So the more expensive a
channel is, the stronger the penalty that is applied on failure will be. It
seems less ideal though, because it could disincentivize cheap but reliable
channels on high traffic links.
The effort required to implement some form of a `highly_available` flag seem
limited and it may help to get payment success rates up. Interested to hear
your thoughts.
Joost
_______________________________________________
Lightning-dev mailing list
Lightning-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/lightning-dev
_______________________________________________
Lightning-dev mailing list
Lightning-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/lightning-dev
_______________________________________________
Lightning-dev mailing list
Lightning-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/lightning-dev
