MIME-Version: 1.0
From: "James O'Beirne" <james.obeirne@gmail.com>
Date: Tue, 2 Apr 2019 16:43:11 -0400
Message-ID: <CAPfvXf+JS6ZhXUieWVxiaNa4uhhWwafCk3odMKy5F_yi=XwngA@mail.gmail.com>
To: bitcoin-dev@lists.linuxfoundation.org
Content-Type: multipart/alternative; boundary="0000000000001d72890585923051"
Subject: [bitcoin-dev] assumeutxo and UTXO snapshots
Precedence: list

--0000000000001d72890585923051
Content-Type: text/plain; charset="UTF-8"

Hi,

I'd like to discuss assumeutxo, which is an appealing and simple
optimization in the spirit of assumevalid[0].

# Motivation

To start a fully validating bitcoin client from scratch, that client
currently
needs to perform an initial block download. To the surprise of no one, IBD
takes a linear amount time based on the length of the chain's history. For
clients running on modest hardware under limited bandwidth constraints,
say a mobile device, completing IBD takes a considerable amount of time
and thus poses serious usability challenges.

As a result, having fully validating clients run on such hardware is rare
and
basically unrealistic. Clients with even moderate resource constraints
are encouraged to rely on the SPV trust model. Though we have promising
improvements to existing SPV modes pending deployment[1], it's worth
thinking about a mechanism that would allow such clients to use trust
models closer to full validation.

The subject of this mail is a proposal for a complementary alternative to
SPV
modes, and which is in the spirit of an existing default, `assumevalid`. It
may
help modest clients transact under a security model that closely resembles
full validation within minutes instead of hours or days.

# assumeutxo

The basic idea is to allow nodes to initialize using a serialized version
of the
UTXO set rendered by another node at some predetermined height. The
initializing node syncs the headers chain from the network, then obtains and
loads one of these UTXO snapshots (i.e. a serialized version of the UTXO set
bundled with the block header indicating its "base" and some other
metadata).

Based upon the snapshot, the node is able to quickly reconstruct its
chainstate,
and compares a hash of the resulting UTXO set to a preordained hash
hard-coded
in the software a la assumevalid. This all takes ~23 minutes, not
accounting for
download of the 3.2GB snapshot[2].

The node then syncs to the network tip and afterwards begins a simultaneous
background validation (i.e., a conventional IBD) up to the base height of
the
snapshot in order to achieve full validation. Crucially, even while the
background validation is happening the node can validate incoming blocks and
transact with the benefit of the full (assumed-valid) UTXO set.

Snapshots could be obtained from multiple separate peers in the same manner
as
block download, but I haven't put much thought into this. In concept it
doesn't
matter too much where the snapshots come from since their validity is
determined via content hash.

# Security

Obviously there are some security implications due consideration. While this
proposal is in the spirit of assumevalid, practical attacks may become
easier.
Under assumevalid, a user can be tricked into transacting under a false
history
if an attacker convinces them to start bitcoind with a malicious
`-assumevalid`
parameter, sybils their node, and then feeds them a bogus chain encompassing
all of the hard-coded checkpoints[3].

The same attack is made easier in assumeutxo because, unlike in assumevalid,
the attacker need not construct a valid PoW chain to get the victim's node
into
a false state; they simply need to get the user to accept a bad
`-assumeutxo`
parameter and then supply them an easily made UTXO snapshot containing,
say, a
false coin assignment.

For this reason, I recommend that if we were to implement assumeutxo, we not
allow its specification via commandline argument[4].

Beyond this risk, I can't think of material differences in security
relative to
assumevalid, though I appeal to the list for help with this.

# More fully validating clients

A particularly exciting use-case for assumeutxo is the possibility of mobile
devices functioning as fully validating nodes with access to the complete
UTXO
set (as an alternative to SPV models). The total resource burden needed to
start a node
from scratch based on a snapshot is, at time of writing, a ~(3.2GB
+ blocks_to_tip * 4MB) download and a few minutes of processing time, which
sounds
manageable for many mobile devices currently in use.

A mobile user could initialize an assumed-valid bitcoin node within an hour,
transact immediately, and complete a pruned full validation of their
assumed-valid chain over the next few days, perhaps only doing the
background
IBD when their device has access to suitable high-bandwidth connections.

If we end up implementing an accumulator-based UTXO scaling design[5][6]
down
the road, it's easy to imagine an analogous process that would allow very
fast
startup using an accumulator of a few kilobytes in lieu of a multi-GB
snapshot.

---

I've created a related issue at our Github repository here:
  https://github.com/bitcoin/bitcoin/issues/15605

and have submitted a draft implementation of snapshot usage via RPC here:
  https://github.com/bitcoin/bitcoin/pull/15606

I'd like to discuss here whether this is a good fit for Bitcoin
conceptually. Concrete
plans for deployment steps should be discussed in the Github issue, and
after all
that my implementation may be reviewed as a sketch of the specific software
changes necessary.

Regards,
James


[0]:
https://bitcoincore.org/en/2017/03/08/release-0.14.0/#assumed-valid-blocks
[1]: https://github.com/bitcoin/bips/blob/master/bip-0157.mediawiki
[2]: as tested at height 569895, on a 12 core Intel Xeon Silver 4116 CPU @
2.10GHz
[3]:
https://github.com/bitcoin/bitcoin/blob/84d0fdc/src/chainparams.cpp#L145-L161
[4]: Marco Falke is due credit for this point
[5]: utreexo: https://www.youtube.com/watch?v=edRun-6ubCc
[6]: Boneh, Bunz, Fisch on accumulators: https://eprint.iacr.org/2018/1188

--0000000000001d72890585923051
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div dir=3D"ltr"><div>Hi,<br></div><div><br></div><div>I&#=
39;d like to discuss assumeutxo, which is an appealing and simple=C2=A0</di=
v><div>optimization in the spirit of assumevalid[0].</div><div><br></div><d=
iv># Motivation</div><div><br></div><div>To start a fully validating bitcoi=
n client from scratch, that client currently</div><div>needs to perform an =
initial block download. To the surprise of no one, IBD=C2=A0</div><div>take=
s a linear amount time based on the length of the chain&#39;s history. For=
=C2=A0</div><div>clients running on modest hardware under limited bandwidth=
 constraints,=C2=A0</div><div>say a mobile device, completing IBD takes a c=
onsiderable amount of time=C2=A0</div><div>and thus poses serious usability=
 challenges.</div><div><br></div><div>As a result, having fully validating =
clients run on such hardware is rare and</div><div>basically unrealistic. C=
lients with even moderate resource constraints</div><div>are encouraged to =
rely on the SPV trust model. Though we have promising</div><div>improvement=
s to existing SPV modes pending deployment[1], it&#39;s worth</div><div>thi=
nking about a mechanism that would allow such clients to use trust</div><di=
v>models closer to full validation.</div><div><br></div><div>The subject of=
 this mail is a proposal for a complementary alternative to SPV</div><div>m=
odes, and which is in the spirit of an existing default, `assumevalid`. It =
may</div><div>help modest clients transact under a security model that clos=
ely resembles</div><div>full validation within minutes instead of hours or =
days.</div><div><br></div><div># assumeutxo</div><div><br></div><div>The ba=
sic idea is to allow nodes to initialize using a serialized version of the<=
/div><div>UTXO set rendered by another node at some predetermined height. T=
he</div><div>initializing node syncs the headers chain from the network, th=
en obtains and</div><div>loads one of these UTXO snapshots (i.e. a serializ=
ed version of the UTXO set</div><div>bundled with the block header indicati=
ng its &quot;base&quot; and some other metadata).</div><div><br></div><div>=
Based upon the snapshot, the node is able to quickly reconstruct its chains=
tate,</div><div>and compares a hash of the resulting UTXO set to a preordai=
ned hash hard-coded</div><div>in the software a la assumevalid. This all ta=
kes ~23 minutes, not accounting for</div><div>download of the 3.2GB snapsho=
t[2].=C2=A0</div><div><br></div><div>The node then syncs to the network tip=
 and afterwards begins a simultaneous</div><div>background validation (i.e.=
, a conventional IBD) up to the base height of the</div><div>snapshot in or=
der to achieve full validation. Crucially, even while the</div><div>backgro=
und validation is happening the node can validate incoming blocks and</div>=
<div>transact with the benefit of the full (assumed-valid) UTXO set.</div><=
div><br></div><div>Snapshots could be obtained from multiple separate peers=
 in the same manner as</div><div>block download, but I haven&#39;t put much=
 thought into this. In concept it doesn&#39;t</div><div>matter too much whe=
re the snapshots come from since their validity is</div><div>determined via=
 content hash.</div><div><br></div><div># Security</div><div><br></div><div=
>Obviously there are some security implications due consideration. While th=
is</div><div>proposal is in the spirit of assumevalid, practical attacks ma=
y become easier.</div><div>Under assumevalid, a user can be tricked into tr=
ansacting under a false history</div><div>if an attacker convinces them to =
start bitcoind with a malicious `-assumevalid`</div><div>parameter, sybils =
their node, and then feeds them a bogus chain encompassing</div><div>all of=
 the hard-coded checkpoints[3].=C2=A0</div><div><br></div><div>The same att=
ack is made easier in assumeutxo because, unlike in assumevalid,</div><div>=
the attacker need not construct a valid PoW chain to get the victim&#39;s n=
ode into</div><div>a false state; they simply need to get the user to accep=
t a bad `-assumeutxo`</div><div>parameter and then supply them an easily ma=
de UTXO snapshot containing, say, a</div><div>false coin assignment.</div><=
div><br></div><div>For this reason, I recommend that if we were to implemen=
t assumeutxo, we not</div><div>allow its specification via commandline argu=
ment[4].</div><div><br></div><div>Beyond this risk, I can&#39;t think of ma=
terial differences in security relative to</div><div>assumevalid, though I =
appeal to the list for help with this.</div><div><br></div><div># More full=
y validating clients</div><div><br></div><div>A particularly exciting use-c=
ase for assumeutxo is the possibility of mobile</div><div>devices functioni=
ng as fully validating nodes with access to the complete UTXO</div><div>set=
 (as an alternative to SPV models). The total resource burden needed to sta=
rt a node</div><div>from scratch based on a snapshot is, at time of writing=
, a ~(3.2GB</div><div>+ blocks_to_tip * 4MB) download and a few minutes of =
processing time, which sounds</div><div>manageable for many mobile devices =
currently in use.</div><div>=C2=A0=C2=A0</div><div>A mobile user could init=
ialize an assumed-valid bitcoin node within an hour,</div><div>transact imm=
ediately, and complete a pruned full validation of their</div><div>assumed-=
valid chain over the next few days, perhaps only doing the background</div>=
<div>IBD when their device has access to suitable high-bandwidth connection=
s.</div><div><br></div><div>If we end up implementing an accumulator-based =
UTXO scaling design[5][6] down</div><div>the road, it&#39;s easy to imagine=
 an analogous process that would allow very fast</div><div>startup using an=
 accumulator of a few kilobytes in lieu of a multi-GB snapshot.</div><div><=
br></div><div>---</div><div><br></div><div>I&#39;ve created a related issue=
 at our Github repository here:</div><div>=C2=A0 <a href=3D"https://github.=
com/bitcoin/bitcoin/issues/15605">https://github.com/bitcoin/bitcoin/issues=
/15605</a></div><div><br></div><div>and have submitted a draft implementati=
on of snapshot usage via RPC here:</div><div>=C2=A0 <a href=3D"https://gith=
ub.com/bitcoin/bitcoin/pull/15606">https://github.com/bitcoin/bitcoin/pull/=
15606</a></div><div><br></div><div>I&#39;d like to discuss here whether thi=
s is a good fit for Bitcoin conceptually. Concrete</div><div>plans for depl=
oyment steps should be discussed in the Github issue, and after all=C2=A0</=
div><div>that my implementation may be reviewed as a sketch of the specific=
 software</div><div>changes necessary.</div><div><br></div><div>Regards,</d=
iv><div>James</div><div><br></div><div><br></div><div>[0]: <a href=3D"https=
://bitcoincore.org/en/2017/03/08/release-0.14.0/#assumed-valid-blocks">http=
s://bitcoincore.org/en/2017/03/08/release-0.14.0/#assumed-valid-blocks</a><=
/div><div>[1]: <a href=3D"https://github.com/bitcoin/bips/blob/master/bip-0=
157.mediawiki">https://github.com/bitcoin/bips/blob/master/bip-0157.mediawi=
ki</a></div><div>[2]: as tested at height 569895, on a 12 core Intel Xeon S=
ilver 4116 CPU @ 2.10GHz</div><div>[3]: <a href=3D"https://github.com/bitco=
in/bitcoin/blob/84d0fdc/src/chainparams.cpp#L145-L161">https://github.com/b=
itcoin/bitcoin/blob/84d0fdc/src/chainparams.cpp#L145-L161</a></div><div>[4]=
: Marco Falke is due credit for this point</div><div>[5]: utreexo: <a href=
=3D"https://www.youtube.com/watch?v=3DedRun-6ubCc">https://www.youtube.com/=
watch?v=3DedRun-6ubCc</a></div><div>[6]: Boneh, Bunz, Fisch on accumulators=
: <a href=3D"https://eprint.iacr.org/2018/1188">https://eprint.iacr.org/201=
8/1188</a></div><div><br></div></div></div>

--0000000000001d72890585923051--