Date: Sun, 30 Oct 2022 11:02:43 +1000
From: Anthony Towns <aj@erisian.com.au>
To: "David A. Harding" <dave@dtrt.org>,
 Bitcoin Protocol Discussion <bitcoin-dev@lists.linuxfoundation.org>
Message-ID: <Y13NM4dyuD6ktvlf@erisian.com.au>
References: <Y1nIKjQC3DkiSGyw@erisian.com.au>
 <194063b733e539e8e24cfd83fa879ed0@dtrt.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <194063b733e539e8e24cfd83fa879ed0@dtrt.org>
Subject: Re: [bitcoin-dev] On mempool policy consistency
Precedence: list

On Fri, Oct 28, 2022 at 09:45:09PM -1000, David A. Harding via bitcoin-dev wrote:
> I think this might be understating the problem.  A 95% chance of having
> an outbound peer accept your tx conversely implies 1 in 20 payments will
> fail to propagate on their initial broadcast.

Whether that's terrible or not depends on how easy it is to retry (and how
likely the retry is to succeed) after a failure -- if a TCP packet fails,
it just gets automatically resent, and if that succeeds, there's a little
lag, but your connection is still usable. I think it's *conceivable* that
a 5% failure rate could be detectable and automatically rectified. Not
that I have a good idea how you'd actually do that, in a way that's
efficient/private/decentralised...

> Some napkin math: there are about 250,000 transactions a day; if
> we round that up to 100 million a year and assume we only want one
> transaction per year to fail to initially propagate on a network where
> 30% of nodes have adopted a more permissive policy, lightweight clients
> will need to connect to over 50 randomly selected nodes.[1]  

A target failure probability of 1-in-1e8 means:

 * with 8 connections, you need 90% of the network to support your txs
 * with 12 connections, you need ~79%
 * with 24 connections (eg everyone running a long-lived node is
   listening, so long lived nodes make 12 outbound and receive about
   ~12 inbound; shortlived nodes just do 24 outbound), you need ~54%

So with that success target, and no preferential peering, you need
a majority of listening nodes to support your tx's features in most
reasonable scenarios, I think.

> For a more
> permissive policy only adopted by 10% of nodes, the lightweight client
> needs to connect to almost 150 nodes.

I get 175 connections needed for that scenario; or 153 with a target
failure rate of 1-in-10-million.

Cheers,
aj