[Monotone-devel] netsync server timeout errors

[Ron Palmer]

Hi all,

My company is looking for new revision control software to manage 
schematic and layout (binary) files used in IC design.  Monotone is one 
of the packages we've been looking at.

First, I think we've found a timeout issue when a slow client is 
performing an initial pull of a database.  We noticed that pulling 
monotone from off.net (the suggested test from the INSTALL file) would 
succeed from some of our machines (Opteron RHEL 3.0) and fail from 
others (Ultra2/400 Solaris 5.8).  The failure message was

> monotone: read from fd 8 (peer off.net) failed, disconnecting

Using --debug on the Solaris machines did not show any obvious problems 
or hangs.  A snippet of the end of the debug log follows.  We added the 
date lines to stderr with a pipe to check for hangs or slow debug outputs.

> Tue Nov  2 16:18:26 PST 2004
> monotone: db.execute("INSERT OR REPLACE INTO merkle_nodes VALUES ('mcert', 
> 'net.venge.monoto ...")
> Tue Nov  2 16:18:26 PST 2004
> monotone: updating subtree setting to 605972f5ec0d824f758811f7e81cc72eb717dd75 
> at slot 0xa of node 0e, level 1
> Tue Nov  2 16:18:26 PST 2004
> monotone: db.execute("INSERT OR REPLACE INTO merkle_nodes VALUES ('mcert', 
> 'net.venge.monoto ...")
> Tue Nov  2 16:18:26 PST 2004
> monotone: updating subtree setting to 622c98ad93a8f60796602e9ed580521b85b81027 
> at slot 0xe of node , level 0
> Tue Nov  2 16:18:26 PST 2004
> monotone: -- finished top level insert --
> Tue Nov  2 16:18:26 PST 2004
> monotone: db.execute("INSERT OR REPLACE INTO merkle_nodes VALUES ('mcert', 
> 'net.venge.monoto ...")
> Tue Nov  2 16:18:26 PST 2004
> monotone: read from fd 8 (peer off.net) failed, disconnecting


Serving the monotone database we successfully pulled from off.net on one 
of the Opteron machines, we tried the same test internally.  The pull 
failed with the same error type (different peer, of course) and we 
noticed the server had closed the client connection with a no response / 
timeout message while the client was still working on the last hunk of 
data transferred.

Is there an option for a keepalive packet from the client that I've 
missed?  If not, can one be added?  It wouldn't need to repeat quickly, 
and probably should be tied to the server timeout.

We also had a question about the public/private keypairs.  We are 
planning for each design to be a separate database used by a group of 
engineers.  These engineers could be working on several different 
projects at once.  Is there a way to keep one keypair for each user 
across multiple databases?

Thanks,

Ron Palmer