On 3/19/2007 4:28 AM, Andrew Sullivan wrote:
> On Sat, Mar 17, 2007 at 01:06:54PM -0400, Steve Singer wrote:
>> 
>> If your DDL has errors then you won't find out about it 
>> until slon tries to execute it on the other node (slonik won't give you an 
>> error but making changes just on your subscribers can have the potential 
>> for all sorts of trouble if your not careful)
> 
> I think this is the reason it didn't work as expected -- I think
> someone raised the objection about what sort of mess this could make.
> That doesn't mean it's not useful.
> 
> This piece of functionality seems to me to be exactly the sort of area
> where using savepoints on the remote note might be a good idea, but
> I'm worried about the potential for really nasty locks.  I think in
> the medium term, we probably need a complete proposal on how to
> improve EXECUTE SCRIPT's safety, or some sort of comprehensive
> discussion about the various locks that are being taken and the like.
> I'm slightly worried that some of the design in that area is happening
> a little _ad hoc_, and I think a more complete idea of what the
> trade-offs are, and which ones we can countenance, would be really
> nice to have.

You can always roll the dice with a left or a right twist ... so 
sometimes you want the thing to halt on error, sometimes you just want 
it to skip errors. The "halt on error" implies some sort of "until" 
thing ... yet to be defined. Anyhow, this means that EXECUTE SCRIPT 
basically needs to perform the script itself in a subtransaction and 
that another option (parameter) should tell if the outer transaction 
(confirming the event) shall commit regardless or not.

This brings back the "cancel event" discussion, we had on IRC a while 
back. I don't really like "cancel event" by itself because the event 
might have been successfully executed on some nodes prior to being 
canceled, so this opens another can of worms. But I would certainly 
agree to the usefulness of some "cancel event X on node Y on error" 
feature. Not that this would be easy to implement, since it somehow 
implies that the node can see the future of its own event queue. Or, it 
would require another Slony table "sl_events_to_ignorei_on_error".


Jan

-- 
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#================================================== JanWieck at Yahoo.com #