tl;dr: By connecting to every Bitcoin Core v24 node I could, and measuring
transaction invs, I determined that at this moment about 17% of all Bitcoin
Core v24 nodes listening on IPv4 are running with full-rbf enabled and
successfully propagating full-rbf replacements.

Procedure:

0) Modify MAX_ADDNODE_CONNECTIONS to 5000 and recompile.
1) Run ./bitcoind -mempoolfullrbf=0 -debug=inv -debug=mempool -debug=mempoolrej
2) Manually addnode every IPv4 address of a node matching 'Satoshi:24' and
   *not* advertising the full-rbf service bit in my DNS seed's 'dnsseed.dump'
   file. This happened to be 692 IPv4 addresses.
3) Wait for connection counts to stabilize. I managed to connect to ~500 nodes
   out of the 692 I tried connecting too.
4) Wait for one of my OpenTimestamps calendars to perform a full-rbf
   replacement¹. They wait a significant amount of time (60s) between
   transactions and blocks to ensure good propagation, and a true full-rbf
   replacement.
5) Wait 2 minutes to ensure complete propagation of the replacement transaction.
6) Run grep <wtxid> ~/.bitcoin/debug.log | grep 'got inv' | wc -l to count the
   number of invs. (I obtained the wtid from another node running full-rbf)
7) Repeat steps 4 to 6 three more times to verify counts are stable.


Discussion:

This data shows substantial adoption of the mempoolfullrbf=1 option among IPv4
listening nodes, above and beyond people choosing to run Bitcoin Knots or
another full-rbf peering fork of Bitcoin Core. This data is also an
underestimate: I'm only measuring successful propagation. Nodes which have
full-rbf enabled - but do not have any full-rbf peers - are not counted by this
measurement. Thus the true number of full-rbf nodes will be even higher than
these stats indicate.

Since v24 nodes are currently only ~5% of all listening nodes, the probability²
of a non-listening node having a full-rbf peer in their outgoing 8 connections
is still low, ~8%. However, if this 17% was maintained as all nodes eventually
upgrade to v24, the probability of a full-rbf peer in the outgoing 8 would be
quite high, ~80%.


Future Work:

How are full-rbf nodes distributed among the IPv4 address space? Bitcoin
Core, by default, groups IPv4 addresses into /16 buckets, and does not connect
to more than 1 outgoing node per bucket. The true probability of connecting to
a full-rbf peer may be changed by this distribution.

How are full-rbf nodes distributed among other connection types? At the moment
bitnodes.io reports that a majority of listening nodes are listening on .onion
addresses. Due to the difficulty of connecting to very large numbers of Tor
nodes at once, and a lack of a convenient source of onion addresses to try, I
did not attempt to measure full-rbf adoption among onion nodes. IIUC a number
of pre-built "node in a box" solutions such as the Start9 Labs Embassy are
currently only able to listen via Tor.

How are full-rbf nodes distributed among non-listening nodes? A potential
strategy to measure this could be to measure inv's on a listening node with a
large number of incoming peers. Anecdotally, I have been told by a number of
people that they're running mempoolfullrbf=1 on non-listening nodes.


References:

1) https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2022-November/021143.html
2) https://stacker.news/items/98441

-- 
https://petertodd.org 'peter'[:-1]@petertodd.org