Slight, maybe problem.

[Tristan Wibberley]

Scott James Remnant wrote:
> On Mon, 2006-11-13 at 21:13 +0100, Luka Renko wrote:

>> ... for "and" 
>> conditions, we need to keep the state of event and evaluate the complete 
>> condition on every change of state (or better event). In your example, newly 
>> registered job (apache) would evaluate conditon of network-up that would be 
>> cached in upstart. 
>>
> The obvious problem here is that upstart would have to cache whether an
> event has happened or not.  And while this is suitable for "from
> network-up until network-down" stanzas, this would not be suitable for
> "on control-alt-delete"
> 
> Also there's a majorly obvious problem...
> 
> while true; do
> 	initctl emit `uuidgen`
> done
> 
> upstart will consume all available memory.

Why not have jobs that have "and" conditions check when they come up 
whether all their conditions are met (via atomic mechanism provided by 
those jobs it "depends" upon) - and die (probably without emitting 
anything) if not? So they get triggered once for each event and only 
stay triggered (telling upstart to emit its "up" event) on the last one.

Jobs should then be written to cope with jobs they depend on 
disappearing. There's no reason for this to be high overhead and a job 
could stay in a semi-alive state to avoid exec-thrashing (with upstart 
providing a unix domain socket based signalling mechanism for such 
clever jobs.

-- 
Tristan Wibberley - "it" is the preferred noun of an abated intellect.

These opinions are my own, and do not reflect those of my employer.