Re: [Monotone-devel] [RFC] Monotone NETSYNC Hook Extension & Abstraction Layer

[Markus Schiltknecht]

Hi,

Thomas Keller wrote:
> > http://www.sqlite.org/cvstrac/wiki?p=UndoRedo might be a good starting point.
>
> Doesn't this violate the ACID properties of a normal transaction

Uh huh? The above link describes a temporary table "holding the 
information needed to undo/redo changes to the database". That certainly 
does not pose any problems regarding ACID properties.

I'm somewhat concerned about the doubled throughput to disk. As with a 
growing amount of revisions in the database, the disk quickly becomes 
the bottleneck during netsync. (Not quite sure, subjective impression, 
not hard measure).

I'm thinking of the transaction multiplexing as a kind of a hack around 
sqlite's lack of fine grained locking (which would allow multiple 
concurrent transactions). Instead of adding yet another hack (manual 
redo log, sort of) I'd rather add support for a real RDBMS [1]. For the 
server side part of the netsync story, that would make things simpler 
and faster.

BTW: is netsync protocol documented somewhere? If not, I'm happy to 
create such a document, as I'm trying to figure out the details anyway.

Regards

Markus

[1]: Whenever I'm saying this I'm clearly referring to Postgres :-)