Date: Sun, 12 Jan 2020 00:42:07 +1000
From: Anthony Towns <aj@erisian.com.au>
To: Bitcoin Protocol Discussion <bitcoin-dev@lists.linuxfoundation.org>
Message-ID: <20200111144207.5xzspeptstspsbsf@erisian.com.au>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CABm2gDq3dAFmRkH2J7d7PcN0A6-F+ZOT22YsDpiORDARmpvu9g@mail.gmail.com>
 <4a132f8a-22e3-8e31-e338-bed9ef46d2ef@mattcorallo.com>
User-Agent: NeoMutt/20170113 (1.7.2)
Subject: Re: [bitcoin-dev] Modern Soft Fork Activation
Precedence: list

On Fri, Jan 10, 2020 at 09:30:09PM +0000, Matt Corallo via bitcoin-dev wrote:
> 1) a standard BIP 9 deployment with a one-year time horizon for
> activation with 95% miner readiness,
> 2) in the case that no activation occurs within a year, a six month
> quieting period during which the community can analyze and discussion
> the reasons for no activation and,
> 3) in the case that it makes sense, a simple command-line/bitcoin.conf
> parameter which was supported since the original deployment release
> would enable users to opt into a BIP 8 deployment with a 24-month
> time-horizon for flag-day activation (as well as a new Bitcoin Core
> release enabling the flag universally).

FWIW etc, but my perspective on this is that the way we want consensus
changes in Bitcoin to work is:

 - decentralised: we want everyone to be able to participate, in
   designing/promoting/reviewing changes, without decision making
   power getting centralised amongst one group or another

 - technical: we want changes to be judged on their objective technical
   merits; politics and animal spirits and the like are fine, especially
   for working out what to prioritise, but they shouldn't be part of the
   final yes/no decision on consensus changes

 - improvements: changes might not make everyone better off, but we
   don't want changes to screw anyone over either -- pareto
   improvements in economics, "first, do no harm", etc. (if we get this
   right, there's no need to make compromises and bundle multiple
   flawed proposals so that everyone's an equal mix of happy and
   miserable)

In particular, we don't want to misalign skills and responsibilities: it's
fine for developers to judge if a proposal has bugs or technical problems,
but we don't want want developers to have to decide if a proposal is
"sufficiently popular" or "economically sound" and the like, for instance.
Likewise we don't want to have miners or pool operators have to take
responsibility for managing the whole economy, rather than just keeping
their systems running.

So the way I hope this will work out is:

 - investors, industry, people in general work out priorities for what's
   valuable to work on; this is an economic/policy/subjective question,
   that everyone can participate in, and everyone can act on --
   either directly if they're developers who can work on proposals and
   implementations directly, or indirectly by persuading or paying other
   people to work on whatever's important

 - developers work on proposals, designing and implementing them to make
   (some subset of) bitcoin users better off, and to not make anyone worse
   off.

 - if someone discovers a good technical reason why a proposal does make
   people worse off, we don't try to keep pushing the proposal over the
   top of objections, but go back to the drawing board and try to fix
   the problems

 - once we've done as much development as we can, including setting up
   experimental testnet/signet style deployments for testing, we setup a
   deployment. the idea at this point is to make sure the live network
   upgrade works, and to retain the ability to abort if last minute
   problems come up. no doubt some review and testing will be left until
   the last minute and only done here, but *ideally* the focus should be
   on catching errors *well before* this point.

 - as a result, the activation strategy mostly needs to be about ensuring
   that the Bitcoin network stays in consensus, rather than checking
   popularity or voting -- the yes/no decisions should have mostly been
   made earlier already. so we have two strategies for locking in the
   upgrade: either 95%+ of hashpower signals that they've upgraded to
   software that will enforce the changes forever more, _or_ after a
   year of trying to deploy, we fail to find any technical problems,
   and then allow an additional 2.5 years to ensure all node software is
   upgraded to enforce the new rules before locking them in.

The strategy behind the last point is that we need to establish that
there's consensus amongst all of Bitcoin before we commit to a flag day,
and if we've found some method to establish consensus on that, then we're
done -- we've already got consensus, we don't need to put a blockchain
protocol on top of that and signal that we've got consensus. (Activating
via hashpower still needs signalling, because we need to coordinate on
*when* sufficient hashpower has upgraded)

This approach is obviously compatible with BIP-148 or BIP-91 style
forced-signalling UASFs if some upgrade does need to be done urgently
despite miner opposition; the forced signalling just needs to occur during
the BIP-9 or BIP-8 phases, and no during the "quiet period". Probably the
first period of BIP-8 after the quiet period would make the most sense.

But without that, this approach seems very friendly for miners: even
if they don't upgrade, they won't mine invalid blocks (unless the rules
activate and someone else deliberately mines an invalid block and they
build on top of it), and if a change is incompatible with, say 10%
of hashpower, it won't be forced on them for 3.5 years, by which point
it's probably a good bet that everyone's upgrading to a new generation
of mining hardware anyway. But even that's a backstop, because if a
change *is* incompatible with existing mining hardware, that's an easily
describable technical problem that should mean we go back to the drawing
board and fix it, not deploy the change despite the problems. [0]

On Fri, Jan 10, 2020 at 11:21:51PM +0100, Jorge Timón via bitcoin-dev wrote:
> Regarding bip8-like activation, luke-jr suggested that [..] a
> consensus rule could require the blocks to signal for activation in
> the last activation window.

FWIW, that had been my (strong) preference too, but I think I'm now
convinced it's not needed/useful.

> I see 2 main advantages for this:
> 1) Outdated nodes can implement warnings (like in bip9) and they can
> see those warnings even if it's activated in the last activation
> window.

The 3.5 year window from BIP-9-starttime to BIP-8-flagday means you'd
have to be using *very* out of date software to need to autodetect
unknown upgrades. If an upgrade starts on 2021-01-01 say, it'd be
supported by 0.21.x, 0.22.0, and 0.23.0 (with bip8 as an opt-in) and
0.24.0, 0.25.0, 0.26.0, 0.27.0, and 0.28.0 (with bip8 as always on)
before flag day activation on 2024-06-01.

0.21.x to 0.23.x could warn if they see potential early BIP-8 activation
via versionbits, and also warn if the flag day date is seen saying "flag
day activation may have happened, please check external sources and
consider upgrading your node software".

So you'd need to be running 0.20.x, released 4 years prior to the
activation to be outdated enough to not get warnings, I think.

> 2) It is easier for users to actively resist a given change they
> oppose. Instead of requiring signaling, their nodes can be set to
> ignore chains that activate it. This will result in a fork, but if
> different groups of users want different things, this is arguably the
> best behaviour: a "clean" split.

If you're knowingly doing a deliberate minority chain split, you'll
almost certainly change the PoW function, and trivially get a clean
split as a result of doing that.

But I think what we want is to move away from consensus decisions being a
"who has the most money/power/hashpower/nodes/reddit-accounts" contest
to being a question of "have we dealt with every reasonable technical
objection?" -- I think that's better for decentralisation in that anyone
can stop a bad proposal without having to be rich/powerful/persuasive,
and better for encouraging constructive contributions. 

The other side to this is that if it's just a matter of resolving
technical problems, then it's also straightforward for a small but skilled
group to get a consensus change through even if the vast majority doesn't
think it's a priority -- they just need to make a good proposal, make
sure it doesn't make people worse off, work through all the objections
people find, and be willing to wait for it to go through reviews and
upgrade steps which may take extra time if other people don't think it's
a high priority. But those are all just technical challenges, that can
be resolved with skill and patience, whoever you might be. So to me,
that's a win for decentralisation as well.

> I assume many people won't like this, but I really think we should
> consider how users should ideally resist an unwanted change, even if
> the proponents had the best intentions in mind, there may be
> legitimate reasons to resist it that they may not have considered.

For me, the focus there is on Matt's first point: "avoid activating
[or merging, or even proposing] in the face of significant, reasonable,
and directed objection". If you want to stop a change you should have to
do nothing more than describe the problems with it; and if there aren't
problems with it, you shouldn't be trying to stop the change.

(A benefit to having the BIP-8 settings defined simultaneously with
the initial activation attempt is that it means that if the core
devs/maintainers go crazy with power and try to force/prevent the BIP-8
activation despite clear community consensus going the other way, then
it will be easy to for the client, and set the parameter correctly --
literally just a matter of changing a value in chainparams.cpp, unlike the
difficulties of changing the blocksize from 1MB to 2MB. Other variations
of this overall approach have the same benefit)

Cheers,
aj (very grateful to Greg and Matt for explaining a lot of thing
    about this approach and helping resolve my concerns with it)

[0] Trigger warning, PTSD over the 2015-2017 blocksize wars...

    The segwit timeline was something like this:

     2015-05 - blocksize debate begins on bitcoin-dev
     2015-08 - bitcoin xt with bip101 hardfork released
     2015-09 - scaling bitcoin phase 1
     2015-12 - segwit proposal at scaling bitcoin phase 2
     2016-01 - segwit testnet launched
     2016-02 - bitcoin classic with bip109 hardfork released
     2016-04 - first release (0.12.1) with a bip9 deployment (csv)
     2016-06 - segwit merged
     2016-07 - csv activated
     2016-10 - first release (0.13.1) with segwit activation params
     2016-11 - segwit activation starttime
     2017-02 - UASF first proposed
     2017-03 - antpool to swith to bitcoin unlimited
     2017-04 - covert ASICBoost vs segwit conflict described
     2017-05 - NY segwit2x agreement, btc1 with bip102 hardfork started
     2017-05 - BIP-91 proposed
     2017-06 - UAHF proposal from bitmain that became BCH
     2017-07 - BIP-91 lockin
     2017-08 - BIP-148 activation
     2017-08 - BCH chainsplit
     2017-08 - segwit lockin and activation
     2017-11 - 2x fork called off; btc1 nodes stall; 2x chain stillborn
     2018-02 - first release (0.16.0) with segwit wallet support

    (That's about 33 months in total, compared to the 24 months we've
    already spent since taproot was first described in Jan 2018, or the
    42 months before flag-day activation in Matt's proposal)

    I don't think that timeline is a good example of how things should
    work, and would call out a few mistakes in particular:

     * too much urgency to increase the blocksize resulting in rushed
       decision making, especially for the hardfork implementations, but
       also for segwit

     * alternative clients attempted to activate forks without
       resolving technical problems (eventually resulting in the btc1
       client stalling prior to the expected hard fork block, eg)

     * a lot of emphasis was on numbers (market share, hashpower, etc)
       rather than technical merits, resulting in a lot of false
       signalling an political meaneuvering

     * the incompatibility between ASICBoost and segwit wasn't noticed
       prior to activation, and wasn't fixed when it was noticed
       (certainly you can justify this as a tit-for-tat response to the
       other errors having been made in bad faith, or as not being a real
       problem because everyone claimed that they weren't doing covert
       ASICBoost, but considered on its own I think the incompatibility
       should have been resolved)

     * the UASF approach had significant potential technical problems
       (potential for long reorgs, p2p network splits) that weren't
       resolved by the time it became active. happily, this was mitigated
       by hashpower enforcement of BIP-148 rules via BIP-91. neither
       BIP-148 or BIP-91 gained enough consensus to be supported in
       bitcoin core though

    I don't personally think we need to fix every problem we had with
    segwit's process -- it eventually mostly worked out okay, after all --
    but I think Matt's approach has a good chance of fixing a lot of
    them, while still leaving us flexibility to deal with whatever new
    problems we come up with in their place.