yet another pending latency question

[Peter Warren]

I'm trying to understand why app engine spawns new instances of our
app given our settings and a miniscule load of 7 requests from one
client.

Our application settings are:
min idle instances: 3
max idle instances: 3
min pending latency: 15 secs
max pending latency: automatic

We run a client-based web app that initially makes a series of
requests to our app engine instance(s). In the case I'm concerned with
there are 7 requests that all take well under 15 seconds. According to
the app engine instance logs the times are: 36ms, 207ms, 52ms, 160ms,
624ms, 1405ms, 3232ms.

There are no other requests being executed against that app; all the
response times are *well* under 15 seconds, and yet app engine spawns
3 new app instances. Why?

In fact the app engine logs show 6 requests being serviced, then 1 new
app instance spawned, then the 7th request serviced, then another 2
app instances spawned.

Our app gets very little traffic, but when a new client app is loaded
in a browser somewhere between 3 & 7 requests are sent to app engine.
I'm trying to keep app engine from launching new instances of our app
since the 3 resident instances should be more than capable of handling
our traffic volume, and spawning new instances kills the response
time.

Thanks for any tips!

Peter

[Peter Warren]

Additionally I just changed the max pending latency setting to 15 secs
and still see app engine spawning new instances.

Given a min latency of 15 secs and a max latency of 15 secs, I would
think that app engine would definitely not spawn a new instance until
a request had been in the queue for 15 secs.

In my latest test I see 5 requests taking under 500 ms, then app
engine spawns a new instance, then the final 2 requests are serviced,
taking 19.5 secs each!

[Jeff Schnitzer]

You probably aren't going to like this tip but:

I find that any deviation from Auto/Auto produces a horrible UX. I
haven't done a great deal of experimentation but all preliminary
results have been bad - lots of unnecessary instance starts, many of
which block user-facing requests. All on a nearly idle application.
At this point I have taped over those sliders like you would a
lightswitch: DO NOT TOUCH.

The behavior of the scheduler is opaque, so those sliders act as
unpredictable inputs to a chinese puzzle whose only perceptible
outputs are tortured screams from your website visitors. It's hard to
imagine a way to win at this game.

Some official transparency as to the actual behavior of the scheduler
would help out a lot.

Jeff

> --
> You received this message because you are subscribed to the Google Groups "Google App Engine" group.
> To post to this group, send email to google-a...@googlegroups.com.
> To unsubscribe from this group, send email to google-appengi...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/google-appengine?hl=en.
>

[Peter Warren]

Setting latency to auto/auto does seem to help. New instances are
still spawned sporadically, but much less frequently. Thanks for the
tip!

[Thomas]

We're facing exactly the same problems: appengine scheduler was, and unfortunately, still is a pain. Google should really get this fixed because it is a major problem for almost every app running there.

Thanks for your suggestion, we will give it a try. 

> To post to this group, send email to google-appengine@googlegroups.com.
> To unsubscribe from this group, send email to google-appengine+unsubscribe@googlegroups.com.

[Cesium]

I'm going to do this right now as well.

David

PS: You, sir, consistently make good points, well stated.

(And make green G2 fly out my nose.)

> To post to this group, send email to google-appengine@googlegroups.com.
> To unsubscribe from this group, send email to google-appengine+unsubscribe@googlegroups.com.