[Monotone-devel] Re: dumb servers

[graydon hoare]

Zbynek Winkler wrote:

> I like the option to have a "mirror" of the my database somewhere on the 
> server accessible to anyone at any time and when the server does not 
> have to be smart it extends my choice of servers.

yeah, I understand this. but we've been experiencing some difficulty 
with the temporal log-based model of communication, so I'm considering 
other alternatives. if such alternatives can be made to work on "dumb" 
servers, all the better; but I'm more concerned with figuring out which 
model of communication is convergent, robust, self-stabilizing, etc. 
than I am insisting that the server be a pre-existing bit of internet 
infrastructure. if it requires a custom server, I'm willing to pay that 
price, if it gets us big robustness improvements.

> you would have removed the these things? Suppose you have two people 
> working on a project that are rarely online at the same time... Do you 
> propose to sync the two databases each of them have on the workstations?

they would sync via a server or peer which *is* online all the time. 
that's sort of the nice thing about hashtree synchronization: it's 
really just set union, so if I sync with a server then you sync with the 
same server, you have all my changes. if I sync with the server again 
later, I have all your changes. it would replace probably 10 different 
commands in monotone with just one: "monotone sync <url>". things which 
reduce the amount of code so dramatically are very attractive.

> While thinking more about this - depot.cgi could function as a subset of 
> monotone knowing only how to sync databases. Is this what you meant?

I haven't made any sort of personal decision about whether to try to 
layer such synchronization on top of HTTP or otherwise. it's a really 
different model than log-replay; the only thing I am pretty sure of is 
that SMTP and NNTP won't fit. I'm going to experiment with using a "raw" 
file-descriptor based synchronization protocol between a couple 
experimental peer programs. if I can get that working, and get a good 
feeling for it being a robust algorithm, I'll look into whether we can 
wedge it into HTTP+CGI (or FTP/SFTP), or need to use a custom server.

I'm sorry if this all sounds like a disruptive surprise, but I've never 
*really* been satisfied with the networking system in monotone -- it's 
efficient, but a bit too fragile -- and I'd prefer to run some 
experiments to see if I can make it better. I promise I won't make it 
worse; if this stuff doesn't work in experimental runs, I won't touch 
any of the existing code.

-graydon