Re: [Ganglia-developers] g3 xml

[matt m. <ma...@cs...>]

> Would gmetad be looking only at the age and expiration values?

gmetad would not unpdate any of the timestamp information from the data 
sources since the data is once removed and the "freshness" of the data is 
defined elsewhere.  gmetad would however wrap each of the data sources 
inside of a "ou" element which would be timestamped.  those timestamps 
would be passed upstream to other interested apps/daemons who.. in turn.. 
would do nothing to them.  each domain is responsible for there own time.

> Chewing through all those human-friendly dates doesn't sound like a fun 
> time.  Plus, RRD (and many other data storage methods) likes 
> second-based time/date stamps.

parsing human readable text isn't efficient.  we would use the age of data 
for rrdtool (e.g. age="5" .. send rrdtool N-5 as a timestamp).  i do know 
that 8601 is fairly easy to parse back to binary .. of course not quite as 
fast as atoi() as we've been using.

> Am I to understand there's a differentiation being made between host 
> attributes (things which don't change until maybe after a reboot, such 
> as total memory or kernel version) and host metrics (which can change at 
> any time) ?

there is no differentiation.  in 2.x we did have a differentiation and 
don't think it is the way we should go.

we had
<HOST NAME="foo" IP="1.2.3.4" LOCATION="..." ..>

which could just as easily been written as

<host name="foo">
  <metric name="ip" value="1.2.3.4"/>
  <metric name="location" value="2,3,4"/>
</host>

i guess i don't see a reason to put certain host metrics as host 
attributes and others as host metrics.  we need to unify them.  

> Does "expires" refer to the time after which a metric should be marked 
> as stale/out-of-date, or culled entirely from the tree?  

the "step" attribute is the time threshold.  if the "age" > "step", we 
have a clue that data was lost somewhere in transport.  

the "expires" attribute is the old Dmax attribute.  when "age" = "expires" 
the data is deleted from the xml/internal data structure all together.

> Should those two values be marked explicitly for each element or
> determined programmatically (one being half the value of the other?) ?  
> I'm assuming that the host values would be inherited by its metric
> units...

the modules will register there metric and define the "step" and "expires" 
intervals (if they exist).  non-registered metrics will require a more 
explicit wire format that contains those values.

> I think that's all the brainpower (such as it is) I can spare for g3 at 
> this time...

i really appreciate the brain dump.  i was getting worried.. i haven't 
heard from any of the developers.  it was good to hear from you.

so here's a more relaxed question: what was you favorite atari game?  i
just bought a ps2 and starting thinking about it.. and my last game
console was an atari.  i've been playing with atari emulators (stella) and
playing the old atari games.  it's odd.. even though the video/sound is
horrible.. the games are still fun.  i forgot how many hours i wasted
challenging my cousin to a game of "asteroids".  my brother and i would
play "boxing" and "combat" all the time.  my single player favorite 
was "yars' revenge" and "missile command".

-matt