OVERVIEW

To improve block propagation, add a new block message that doesn't include
transactions the peer is known to have. The message must never require an
additional round trip due to any transactions the peer doesn't have, but
should
be compatible with peers sometimes forgetting transactions they have known.

APPROACH

For peers advertising support for squashed blocks: a node tracks what txes
it
knows each peer has seen (inv received, tx sent, tx appeared in competing
block
known to peer). Nodes push block contents as txes-not-already-known +
txids-known.

A node should be able to forget invs it has seen without invalidating what
peers
know about its known txes. To allow for this, a node assembles a bloom
filter of
a set of txes it is going to forget, and sends it to peers. The node can
erase
the txes as soon as no blocks requested before the filter was pushed are in
flight (relying on the assumption that messages can be expected to be
processed
in order).

When a node receives a forgotten-filter, it ORs it into its
forgotten-filter for
that peer. Any transactions matching the forgotten-filter are always
included in
full with a block. If the filter is getting full, the node can just clear it
along with peer.setTxKnown.

COSTS

Bloom filtering:
Since the bloom filter is likely to grow slowly and can be dropped when it
is
becoming full, a cheap set of hash functions and element size can be used to
keep overhead more restricted than the bloom filtering done for spv. It's
important for testing txes against the filter to be fast so that it doesn't
delay pushing the block more than the squashing helps.
Nodes currently forget txes rarely, so the bloom filters would only need to
be
used at all under conditions that are not currently common -- but I think
they're important to include to allow for different node behavior in this
regard
in the future.

Tracking txes known to peers:
A multimap of txid->peerId would obviate the current setCurrentlyKnown, and
would not take much more space since each additional peer adds about 1
peerId
per txid (setCurrentlyKnown keeps a uint256 per peer per txid, although it
tracks somewhat fewer txid per node).

Potential vulnerabilities:
- Since the bloom filters will have lower maximum overhead than the current
SPV
  filters and can be dropped at will, this shouldn't enable any resource
  exhaustion attacks that aren't already possible.
- A squashed block with bogus or missing data would be easily detected not
to
  produce the correct merkle root for its BlockHeader.

BENEFITS

Assuming a fairly typical 500 tx block with transaction sizes averaging 300b
(both on the low side), for a 150kb block:

% pruned | block size reduction | relative size reduction
-------- | -------------------- | -----------------------
100      | 134 kB               | 89%
50       | 67 kB                | 45%
25       | 33.5 kB              | 17%

I've been doing some logging, and when my node pushes a block to a peer it
seems
to typically know that a peer has seen most of the txes in the block. Even
in
the case of a small block with only 25% known-known transactions, total
network
bandwidth saved is greater than the bloom filters transmitted unless a node
is
forgetting transactions so rapidly that it pushes new maximum-size
forget-filters every block.

So this is a net gain even in total bandwidth usage, but most importantly
it's
an improvement in block propagation rate and in how block propagation rate
scales with additional transactions.

IMPLEMENTATION QUESTIONS

How should block squashing capability be advertised -- new service bit?

Bloom filters:
- How fast to test against could a suitable bloom filter be made?
- How much memory would each filter need to take, at maximum?
- Can the inputs all being 32 byte hashes be used to optimize filter hash
  calculations?

ROADMAP

If there's support for this proposal, I can begin working on the specific
implementation details, such as the bloom filters, message format, and
capability advertisment, and draft a BIP once I have a concrete proposal for
what those would look like and a corresponding precise cost/benefit
analysis.

--kaz