I think I know what's happening. With delayed commits each database
waits ~1 sec before fully committing to disk. So each database is
considered "open" until that commit happens. So this looks like normal
behavior to me.
We could try to tell a database to fully commit when that happens,
then close it, but it can take a long time to complete a fsync and a
client could reopen the database in that time, etc. There are no
perfect solutions here.
-Damien
On Nov 25, 2009, at 2:55 PM, Adam Kocoloski wrote:
On Nov 25, 2009, at 2:23 PM, Paul Davis wrote:
#!/usr/bin/perl
use FindBin qw($Bin);
use lib ("$Bin/lib","$Bin/cpan-lib");
use Net::CouchDB;
my $host=shift;
my $secs=shift;
my $couch=Net::CouchDB->new($host);
my @docs=map{ { '_id' => $_, 'lang' => 'erlang' } } (1...500);
foreach (1...200) {
my $dbh=$couch->create_db("event-$_");
print "Created database $_\n";
$dbh->insert(@docs);
sleep($secs);
}
IANAPM, but if $dbh is holding an open connection you could very well
trigger this quite easily. Can you try replacing the sleep($secs)
with
something like $dbh->close()? Any easy way to check this is to watch
`netstat -tap tcp` and see if the number of sockets on either machine
is growing monotonically.
HTH,
Paul Davis
Paul, that was my first thought too, but isn't the DB only
considered "active" for the lifetime of a request? E.g. it doesn't
matter if the connections are kept open or not, once
couch_db:close(Db) is called the reference counter gets decremented
and couch_server can drop it from the LRU cache. At least that's my
reading of the code. Best,
Adam
