Hi Suhas,

Thanks for your attention and feedback!

> Transaction A is both low-fee and non-standard to some nodes on the network...
> ...Whenever a transaction T that spends A is relayed, new nodes will send INV(PKGINFO1, T) to all package-relay peers...
> ...because of transaction malleability, and to avoid being blinded to a transaction unnecessarily, these nodes will likely still send getdata(PKGINFO1, T) to every node that announces T...

Yes, we'd request pkginfo unless we already had the transaction in our mempool. The pkginfo step is intended to prevent nodes from ever downloading a transaction more than once; I was going for a benchmark of "packages are announced once per p2p connection, transaction data downloaded once per node".

In this scenario, both A and T's wtxids would be sent once per p2p connection and transaction data downloaded once per node. If T has other unconfirmed parents, the low-fee ones will only be announced once (in pkginfo) per link. If it has high-fee parents, they will indeed be announced more than once per link (once individually, then again in pkginfo).

More precisely: if a package contains any transactions which are non-standard to one peer and standard to another, the package transactions (parents, not child) that pass the fee filter on their own will be announced twice instead of once.

> I think a good design goal would be to not waste bandwidth in non-adversarial situations. In this case, there would be bandwidth waste from downloading duplicate data from all your peers, just because the announcement doesn't commit to the set of parent wtxids that we'd get from the peer (and so we are unable to determine that all our peers would be telling us the same thing, just based on the announcement).

Each transaction is only downloaded once per node here, and each package announced/pkginfo sent once per link. I definitely understand that this doesn't pass a benchmark of "every transaction is announced at most once per link," but it's still on the magnitude of 32-byte hashes. Adding a commitment to parents in the announcements is an extra hash per link in all cases - my question is whether it's worth it? We'd also need to write new inv/getdata message types for package relay, though that's probably a weaker argument.

> it won't always be the case that a v1 package relay node will be able to validate that a set of package transactions is fully sorted topologically, because there may be (non-parent) ancestors that are missing from the package and the best a peer can validate is topology within the package -- this means that a peer can validly (under this BIP) relay transaction packages out of the true topological sort (if all ancestors were included).

Good point. Since v1 packages don't necessarily include the full ancestor set, we wouldn't be able to verify that two parents are in the right order if they have an indirect dependency, e.g. parent 1 spends a tx ("grandparent") which spends parent 2. Note that the grandparent couldn't possibly be in the mempool unless parent 2 is. We'd eventually get everything submitted as long as we received the grandparent, and then know whether the package was topologically sorted. But I think you're right that this could be a "nice to have" instead of a protocol requirement.

> Could you explain again what the benefit of including the blockhash is? It seems like it is just so that a node could prioritize transaction relay from peers with the same chain tip to maximize the likelihood of transaction acceptance, but in the common case this seems like a pretty negligible concern...

The blockhash is necessary in order to disambiguate between a malformed package and difference in chain tip. If a parent is missing from a package, it's possible it was mined in a recent block that we haven't seen yet. Validating using a UTXO set, all we see is "missing inputs" when we try to validate the child; we wouldn't know if our peer had sent us a malformed package or if we were behind.

I'm hoping some of these clarifications are helpful to post publicly, but I know I haven't fully addressed all the concerns you've brought up. Will continue to think about this.

Best,
Gloria

On Wed, Jun 8, 2022 at 4:59 PM Suhas Daftuar via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> wrote:
Hi,

Thanks again for your work on this!

One question I have is about potential bandwidth waste in the case of nodes running with different policy rules.  Here's my understanding of a scenario I think could happen:

1) Transaction A is both low-fee and non-standard to some nodes on the network.
2) Whenever a transaction T that spends A is relayed, new nodes will send INV(PKGINFO1, T) to all package-relay peers.
3) Nodes on the network that have implemented package relay, but do not accept A, will send getdata(PKGINFO1, T) and learn all of T's unconfirmed parents (~32 bytes * number of parents(T)).
4) Such nodes will reject T.  But because of transaction malleability, and to avoid being blinded to a transaction unnecessarily, these nodes will likely still send getdata(PKGINFO1, T) to every node that announces T, in case someone has a transaction that includes an alternate set of parent transactions that would pass policy checks.

Is that understanding correct?  I think a good design goal would be to not waste bandwidth in non-adversarial situations.  In this case, there would be bandwidth waste from downloading duplicate data from all your peers, just because the announcement doesn't commit to the set of parent wtxids that we'd get from the peer (and so we are unable to determine that all our peers would be telling us the same thing, just based on the announcement).

Some ways to mitigate this might be to: (a) include a hash (maybe even just a 20-byte hash -- is that enough security?) of the package wtxids (in some canonical ordering) along with the wtxid of the child in the initial announcement; (b) limit the use of v1 packages to transactions with very few parents (I don't know if this is reasonable for the use cases we have in mind).

Another point I wanted to bring up is about the rules around v1 package validation generally, and the use of a blockhash in transaction relay specifically.  My first observation is that it won't always be the case that a v1 package relay node will be able to validate that a set of package transactions is fully sorted topologically, because there may be (non-parent) ancestors that are missing from the package and the best a peer can validate is topology within the package -- this means that a peer can validly (under this BIP) relay transaction packages out of the true topological sort (if all ancestors were included).

This makes me wonder how useful this topological rule is.  I suppose there is some value in preventing completely broken implementations from staying connected and so there is no harm in having the rule, but perhaps it would be helpful to add that nodes SHOULD order transactions based on topological sort in the complete transaction graph, so that if missing-from-package ancestors are already known by a peer (which is the expected case when using v1 package relay on transactions that have more than one generation of unconfirmed ancestor) then the remaining transactions are already properly ordered, and this is helpful even if unenforceable in general.  

The other observation I wanted to make was that having transaction relay gated on whether two nodes agree on chain tip seems like an overly restrictive criteria.  I think an important design principle is that we want to minimize disruption from network splits -- if there are competing blocks found in a small window of time, it's likely that the utxo set is not materially different on the two chains (assuming miners are selecting from roughly the same sets of transactions when this happens, which is typical).  Having transaction relay bifurcate on the two network halves would seem to exacerbate the difference between the two sides of the split -- users ought to be agnostic about how benign splits are resolved and would likely want their transactions to relay across the whole network.

Additionally, use of a chain tip might impose a larger burden than is necessary on software that would seek to participate in transaction relay without implementing headers sync/validation.  I don't know what software exists on the network, but I imagine there are a lot of scripts out there for transaction submission to the public p2p network, and in thinking about modifying such a script to utilize package relay it seems like an unnecessary added burden to first learn a node's tip before trying to relay a transaction.

Could you explain again what the benefit of including the blockhash is?  It seems like it is just so that a node could prioritize transaction relay from peers with the same chain tip to maximize the likelihood of transaction acceptance, but in the common case this seems like a pretty negligible concern, and in the case of a chain fork that persists for many minutes it seems better to me that we not partition the network into package-relay regimes and just risk a little extra bandwidth in one direction or the other.  If we solve the problem I brought up at the beginning (of de-duplicating package data across peers with a package-wtxid-commitment in the announcement), I think this is just some wasted pkginfo bandwidth on a single-link, and not across links (as we could cache validation failure for a package-hash to avoid re-requesting duplicate pkginfo1 messages).

Best,
Suhas


On Tue, Jun 7, 2022 at 1:57 PM Gloria Zhao via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> wrote:
Hi Eric, aj, all,

Sorry for the delayed response. @aj I'm including some paraphrased points from our offline discussion (thanks).

> Other idea: what if you encode the parent txs as a short hash of the wtxid (something like bip152 short ids? perhaps seeded per peer so collisions will be different per peer?) and include that in the inv announcement? Would that work to avoid a round trip almost all of the time, while still giving you enough info to save bw by deduping parents?

> As I suggested earlier, a package is fundamentally a compact block (or
> block) announcement without the header. Compact block (BIP152) announcement
> is already well-defined and widely implemented...

> Let us not reinvent the wheel and/or introduce accidental complexity. I see
> no reason why packaging is not simply BIP152 without the 'header' field, an
> updated protocol version, and the following sort of changes to names

Interestingly, "why not use BIP 152 shortids to save bandwidth?" is by far the most common suggestion I hear (including offline feedback). Here's a full explanation:

BIP 152 shortens transaction hashes (32 bytes) to shortids (6 bytes) to save a significant amount of network bandwidth, which is extremely important in block relay. However, this comes at the expense of computational complexity. There is no way to directly calculate a transaction hash from a shortid; upon receipt of a compact block, a node is expected to calculate the shortids of every unconfirmed transaction it knows about to find the matches (BIP 152: [1], Bitcoin Core: [2]). This is expensive but appropriate for block relay, since the block must have a valid Proof of Work and new blocks only come every ~10 minutes. On the other hand, if we require nodes to calculate shortids for every transaction in their mempools every time they receive a package, we are creating a DoS vector. Unconfirmed transactions don't need PoW and, to have a live transaction relay network, we should expect nodes to handle transactions at a high-ish rate (i.e. at least 1000s of times more transactions than blocks). We can't pre-calculate or cache shortids for mempool transactions, since the SipHash key depends on the block hash and a per-connection salt.

Additionally, shortid calculation is not designed to prevent intentional individual collisions. If we were to use these shortids to deduplicate transactions we've supposedly already seen, we may have a censorship vector. Again, these tradeoffs make sense for compact block relay (see shortid section in BIP 152 [3]), but not package relay.

TLDR: DoSy if we calculate shortids on every package and censorship vector if we use shortids for deduplication.

> Given this message there is no reason  
> to send a (potentially bogus) fee rate with every package. It can only be  
> validated by obtaining the full set of txs, and the only recourse is  
> dropping (etc.) the peer, as is the case with single txs.

Yeah, I agree with this. Combined with the previous discussion with aj (i.e. we can't accurately communicate the incentive compatibility of a package without sending the full graph, and this whole dance is to avoid downloading a few low-fee transactions in uncommon edge cases), I've realized I should remove the fee + weight information from pkginfo. Yay for less complexity!

Also, this might be pedantic, but I said something incorrect earlier and would like to correct myself:

>> In theory, yes, but maybe it was announced earlier (while our node was down?) or had dropped from our mempool or similar, either way we don't have those txs yet.  

I said "It's fine if they have Erlay, since a sender would know in advance that B is missing and announce it as a package." But this isn't true since we're only using reconciliation in place of flooding to announce transactions as they arrive, not for rebroadcast, and we're not doing full mempool set reconciliation. In any case, making sure a node receives the transactions announced when it was offline is not something we guarantee, not an intended use case for package relay, and not worsened by this.

Thanks for your feedback!

Best,
Gloria

On Thu, May 26, 2022 at 3:59 AM <eric@voskuil.org> wrote:
Given that packages have no header, the package requires identity in a
BIP152 scheme. For example 'header' and 'blockhash' fields can be replaced
with a Merkle root (e.g. "identity" field) for the package, uniquely
identifying the partially-ordered set of txs. And use of 'getdata' (to
obtain a package by hash) can be eliminated (not a use case).

e

> -----Original Message-----
> From: eric@voskuil.org <eric@voskuil.org>
> Sent: Wednesday, May 25, 2022 1:52 PM
> To: 'Anthony Towns' <aj@erisian.com.au>; 'Bitcoin Protocol Discussion'
> <bitcoin-dev@lists.linuxfoundation.org>; 'Gloria Zhao'
> <gloriajzhao@gmail.com>
> Subject: RE: [bitcoin-dev] Package Relay Proposal
>
> > From: bitcoin-dev <bitcoin-dev-bounces@lists.linuxfoundation.org> On
> Behalf
> > Of Anthony Towns via bitcoin-dev
> > Sent: Wednesday, May 25, 2022 11:56 AM
>
> > So the other thing is what happens if the peer announcing packages to us
> is
> > dishonest?
> >
> > They announce pkg X, say X has parents A B C and the fee rate is
garbage.
> But
> > actually X has parent D and the fee rate is excellent. Do we request the
> > package from another peer, or every peer, to double check? Otherwise
> we're
> > allowing the first peer we ask about a package to censor that tx from
us?
> >
> > I think the fix for that is just to provide the fee and weight when
> announcing
> > the package rather than only being asked for its info? Then if one peer
> makes
> > it sound like a good deal you ask for the parent txids from them,
dedupe,
> > request, and verify they were honest about the parents.
>
> Single tx broadcasts do not carry an advertised fee rate, however the'
> feefilter' message (BIP133) provides this distinction. This should be
> interpreted as applicable to packages. Given this message there is no
reason
> to send a (potentially bogus) fee rate with every package. It can only be
> validated by obtaining the full set of txs, and the only recourse is
> dropping (etc.) the peer, as is the case with single txs. Relying on the
> existing message is simpler, more consistent, and more efficient.
>
> > >> Is it plausible to add the graph in?
> >
> > Likewise, I think you'd have to have the graph info from many nodes if
> you're
> > going to make decisions based on it and don't want hostile peers to be
> able to
> > trick you into ignoring txs.
> >
> > Other idea: what if you encode the parent txs as a short hash of the
wtxid
> > (something like bip152 short ids? perhaps seeded per peer so collisions
> will
> > be different per peer?) and include that in the inv announcement? Would
> > that work to avoid a round trip almost all of the time, while still
giving
> you
> > enough info to save bw by deduping parents?
>
> As I suggested earlier, a package is fundamentally a compact block (or
> block) announcement without the header. Compact block (BIP152)
> announcement
> is already well-defined and widely implemented. A node should never be
> required to retain an orphan, and BIP152 ensures this is not required.
>
> Once a validated set of txs within the package has been obtained with
> sufficient fee, a fee-optimal node would accept the largest subgraph of
the
> package that conforms to fee constraints and drop any peer that provides a
> package for which the full graph does not.
>
> Let us not reinvent the wheel and/or introduce accidental complexity. I
see
> no reason why packaging is not simply BIP152 without the 'header' field,
an
> updated protocol version, and the following sort of changes to names:
>
> sendpkg
> MSG_CMPCT_PKG
> cmpctpkg
> getpkgtxn
> pkgtxn
>
> > > For a maximum 25 transactions,
> > >23*24/2 = 276, seems like 36 bytes for a child-with-parents package.
> >
> > If you're doing short ids that's maybe 25*4B=100B already, then the
above
> is
> > up to 36% overhead, I guess. Might be worth thinking more about, but
> maybe
> > more interesting with ancestors than just parents.
> >
> > >Also side note, since there are no size/count params,
>
> Size is restricted in the same manner as block and transaction broadcasts,
> by consensus. If the fee rate is sufficient there would be no reason to
> preclude any valid size up to what can be mined in one block (packaging
> across blocks is not economically rational under the assumption that one
> miner cannot expect to mine multiple blocks in a row). Count is
incorporated
> into BIP152 as 'shortids_length'.
>
> > > wondering if we
> > >should just have "version" in "sendpackages" be a bit field instead of
> > >sending a message for each version. 32 versions should be enough right?
>
> Adding versioning to individual protocols is just a reflection of the
> insufficiency of the initial protocol versioning design, and that of the
> various ad-hoc changes to it (including yet another approach in this
> proposal) that have been introduced to compensate for it, though I'll
> address this in an independent post at some point.
>
> Best,
> e
>
> > Maybe but a couple of messages per connection doesn't really seem worth
> > arguing about?
> >
> > Cheers,
> > aj
> >
> >
> > --
> > Sent from my phone.
> > _______________________________________________
> > bitcoin-dev mailing list
> > bitcoin-dev@lists.linuxfoundation.org
> > https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev


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