(To everyone):
I should have made it much clearer that version 1 is only supposed to solve 1 of the 2 use cases. I was a lot more focused on the fee-bumping use case, since it’s more important. Orphan-fetching was added to the motivation section last-minute because John Newbery mentioned to me “hey you could deal with orphans really easily with this.” Of course, child-with-unconfirmed-parents packages aren’t very useful for orphan-fetching since non-parent ancestors are quite common.

Maybe a version 2 package for orphan-fetching could look like this:

“pckginfo2” message contains a tx with all of its ancestors

“MSG_PCKG2” inv type refers to a “pckginfo2” for a tx. You don’t send inv(MSG_PCKG2), but a node can request getdata(MSG_PCKG2) for a transaction they want the ancestors for, provided they sent sendpackages(version=2) ahead of time. It seems to me that orphan-fetching only ever needs to be receiver-initiated.

Protocol flow would look like this:
https://user-images.githubusercontent.com/25183001/168891185-1630f583-de47-4937-86b1-2652cf8852f2.png

We don’t have a policy for dealing with anything more than a child with its direct parents, but I also don’t think anybody is relying on fee-bumping more than 2 generations, so the validation logic here could probably just submit them all individually. Maybe they can request a pckginfo1 if they see something that’s too-low-fee, and/or use the child-with-unconfirmed-parents logic opportunistically.

Thanks aj for the feedback! Responding:

> The "PCKG" abbreviation threw me for a loop; isn't the usual
> abbreviation "PKG" ?

Oh I didn't know that. I could change it if people feel strongly.

> Does it make sense for these to be configurable, rather than implied
> by the version?
> … would it be better to either just not do sendpackages
> at all if you're limiting ancestors in the mempool incompatibly

Effectively: if you’re setting your ancestor/descendant limits lower than the default, you can’t do package relay. I wonder if this might be controversial, since it adds pressure to adhere to Bitcoin Core’s current mempool policy? I would be happy to do it this way, though - makes things easier to implement.

> > 5. If 'fRelay==false' in a peer's version message, the node must not
> >    send "sendpackages" to them. If a "sendpackages" message is
> > received by a peer after sending `fRelay==false` in their version
> > message, the sender should be disconnected.

> Seems better to just say "if you set fRelay=false in your version
> message, you must not send sendpackages"? You already won't do packages
> with the peer if they don't also announce sendpackages.

I guess, theoretically, if you allow bloom filters with this peer, it’s plausible they’re saying “fRelay=false, I’ll send you a bloom filter later, and I’ll also want to talk about packages.”
I don’t know if that’s a use case we want to support - my gut reaction is no.

> Maybe: "You must not send sendpackages unless you also send wtxidrelay" ?

Do you mean if we get a verack, and the peer sent “sendpackages” but not “wtxidrelay,” we should disconnect them?

> As I understand it, the two cases for the protocol flow are "I received
> an orphan, and I'd like its ancestors please" which seems simple enough,
> and "here's a child you may be interested in, even though you possibly
> weren't interested in the parents of that child".

(Btw, please see my notes at the top of this email about separating those two use cases. sorry for the confusion).

> I think the logic for the latter is: […]

I have it as: we send a PCKG INV when this transaction’s feerate is above the fee filter, but one or more of its parents don’t. I don’t think using ancestor feerate is better.
See this counterexample: https://raw.githubusercontent.com/glozow/bitcoin-notes/master/mempool_garden/abc_1parent_2kids.png
A (0fee) has 2 kids, B (3sat/vB) and C (20sat/vB), everything’s the same vsize. Let’s say the fee filter is 3sat/vB.
If we do it based on ancestor feerate, we won’t send B. But B is actually fine; C is paying for A.

> Are "getpckgtxns" / "pcktxns" really limited to packages, or are they
> just a general way to request a batch of transactions?

> Maybe call those messages "getbatchtxns" and "batchtxns" and allow them to
> be used more generally, potentially in ways unrelated to packages/cpfp?

Indeed, it’s a general way to request a batch of transactions. I’ll highlight that it is “all or nothing,” i.e. if the sender is missing any of them, they’ll just send a notfound.
The idea here was to avoid downloading any transactions that can’t be validated right away. With packages, this makes sense, because there are dependency relationships. But if you’re requesting multiple unrelated transactions, for example, it’s unnecessary. You might end up with even more transaction data that’s just sitting around waiting to be validated.

> The "only be sent if both peers agreed to do package relay" rule could
> simply be dropped, I think.

Wouldn’t we need some way of saying “hey I support batchtxns?” Otherwise you would have to guess by sending a request and waiting to see if it’s ignored?

> Shouldn't the sender only be sending package announcements when they know
> the recipient will be interested in the package, based on their feefilter?

I think there are cases where the sender doesn’t necessarily know.
Consider this example: https://raw.githubusercontent.com/glozow/bitcoin-notes/master/mempool_garden/rich_parent_bad_cpfp.png
D (5sat/vB) has 2 parents, A (0sat/vB) and B (20sat/vB). All same size. Feefilter is 3sat/vB.
If the receiver already has B, they’ll know they can just reject the package already based on the pckginfo.
But the sender doesn’t really know that. The sender just knows A is below feerate and D is above. D is above the fee filter, and its ancestor feerate is above the fee filter.

> CAmount in consensus/amount.h is a int64_t
> The maximum block weight is 4M…

 Oops yes. I think we just usually use int64_t for vsizes afaik. Agree that it should be 8 bytes for fee, and 4 bytes is enough for vsize.

> I guess tx relay is low priority enough that it wouldn't be worth tagging
> some peers as "high bandwidth" and having them immediately announce the
> PCKGINFO1 message, and skip the INV/GETDATA step?

I had the same idea as well, but seemed unnecessary. It would reduce the number of round trips, but I don’t think an extra round trip is that big of a deal for transaction relay. Block relay, yes of course, but I don’t think we care that much if it takes an extra second to send a transaction?

Best,
Gloria

On Tue, May 17, 2022 at 8:35 PM Anthony Towns <aj@erisian.com.au> wrote:
On Tue, May 17, 2022 at 12:01:04PM -0400, Gloria Zhao via bitcoin-dev wrote:
> ====New Messages====
> Three new protocol messages are added for use in any version of
> package relay. Additionally, each version of package relay must define
> its own inv type and "pckginfo" message version, referred to in this
> document as "MSG_PCKG" and "pckginfo" respectively. See
> BIP-v1-packages for a concrete example.

The "PCKG" abbreviation threw me for a loop; isn't the usual
abbreviation "PKG" ?

> =====sendpackages=====
> |version || uint32_t || 4 || Denotes a package version supported by the
> node.
> |max_count || uint32_t || 4 ||Specifies the maximum number of transactions
> per package this node is
> willing to accept.
> |max_weight || uint32_t || 4 ||Specifies the maximum total weight per
> package this node is willing
> to accept.

Does it make sense for these to be configurable, rather than implied
by the version?

I presume the idea is to cope with people specifying different values for
-limitancestorcount or -limitancestorsize, but if people are regularly
relaying packages around, it seems like it becomes hard to have those
values really be configurable while being compatible with that?

I guess I'm asking: would it be better to either just not do sendpackages
at all if you're limiting ancestors in the mempool incompatibly; or
alternatively, would it be better to do the package relay, then reject
the particular package if it turns out too big, and log that you've
dropped it so that the node operator has some way of realising "whoops,
I'm not relaying packages properly because of how I configured my node"?

> 5. If 'fRelay==false' in a peer's version message, the node must not
>    send "sendpackages" to them. If a "sendpackages" message is
> received by a peer after sending `fRelay==false` in their version
> message, the sender should be disconnected.

Seems better to just say "if you set fRelay=false in your version
message, you must not send sendpackages"? You already won't do packages
with the peer if they don't also announce sendpackages.

> 7. If both peers send "wtxidrelay" and "sendpackages" with the same
>    version, the peers should announce, request, and send package
> information to each other.

Maybe: "You must not send sendpackages unless you also send wtxidrelay" ?


As I understand it, the two cases for the protocol flow are "I received
an orphan, and I'd like its ancestors please" which seems simple enough,
and "here's a child you may be interested in, even though you possibly
weren't interested in the parents of that child". I think the logic for
the latter is:

 * if tx C's fee rate is less than the peer's feefilter, skip it
   (will maybe treat it as a parent in some package later though)
 * if tx C's ancestor fee rate is less than the peer's feefilter, skip
   it?
 * look at the lowest ancestor fee rate for any of C's in-mempool
   parents
 * if that is higher than the peer's fee filter, send a normal INV
 * if it's lower than the peer's fee filter, send a PCKG INV

Are "getpckgtxns" / "pcktxns" really limited to packages, or are they
just a general way to request a batch of transactions? Particularly in
the case of requesting the parents of an orphan tx you already have,
it seems hard for the node receiving getpckgtxns to validate that the
txs are related in some way; but also it doesn't seem very necessary?

Maybe call those messages "getbatchtxns" and "batchtxns" and allow them to
be used more generally, potentially in ways unrelated to packages/cpfp?
The "only be sent if both peers agreed to do package relay" rule could
simply be dropped, I think.

> 4. The reciever uses the package information to decide how to request
>    the transactions. For example, if the receiver already has some of
> the transactions in their mempool, they only request the missing ones.
> They could also decide not to request the package at all based on the
> fee information provided.

Shouldn't the sender only be sending package announcements when they know
the recipient will be interested in the package, based on their feefilter?

> =====pckginfo1=====
> {|
> |  Field Name  ||  Type  ||  Size  ||   Purpose
> |-
> |blockhash || uint256 || 32 || The chain tip at which this package is
> defined.
> |-
> |pckg_fee||CAmount||4|| The sum total fees paid by all transactions in the
> package.

CAmount in consensus/amount.h is a int64_t so shouldn't this be 8
bytes? If you limit a package to 101kvB, an int32_t is enough to cover
any package with a fee rate of about 212 BTC/block or lower, though.

> |pckg_weight||int64_t||8|| The sum total weight of all transactions in the
> package.

The maximum block weight is 4M, and the default -limitancestorsize
presumably implies a max package weight of 404k; seems odd to provide
a uint64_t rather than an int32_t here, which easily allows either of
those values?

> 2. ''Only 1 child with unconfirmed parents.'' The package must consist
>    of one transaction and its unconfirmed parents. There must not be
> any other transactions in the package. Other dependency relationships
> may exist within the package (e.g. one parent may spend the output of
> another parent) provided that topological order is respected.

I think this means that some of the parents could also have unconfirmed
parents, but they won't be included in the package, and must be requested
via the recipient-initiated approach?

> 5. ''Total fees and weight.'' The 'total_fee' and 'total_weight'
>    fields must accurately represent the sum total of all transactions'
>    fees and weights as defined in BIP141, respectively.

Presumably this excludes any unconfirmed grandparents and earlier
ancestors since they aren't part of the package, in this approach? Doesn't
that make this both harder to calculate (assuming we already have
ancestor summaries) and less useful, in the case where those ancestors
have a lower fee rate?

> ''Q: Can "getpckgtxns" and "pckgtxns" messages contain only one
> transaction?''
> Yes.

This would be normal if you're requesting a single missing parent for
an orphan you've received, I think?

I'm slightly surprised the process is:

 ->  INV PCKG1 C
  <- GETDATA PCKG1 C
 ->  PCKGINFO1 blockhash A B C fee weight

rather than announcing the package fee info in the first message.
But if the sender is already applying the feefilter to the package before
announcing it, it probably doesn't matter, and means you're only getting
a 32B INV from every peer, rather than a 32*(n+2) PCKGINFO1 message from
every peer.

I guess tx relay is low priority enough that it wouldn't be worth tagging
some peers as "high bandwidth" and having them immediately announce the
PCKGINFO1 message, and skip the INV/GETDATA step?

Cheers,
aj