Some of these threads are getting a bit long ("tl;dr"), but following Jon's
interesting explanations of the "15 minute minimum" issue that has had many
of us worried, let me try a brief explanation of how I see it working (I'm a
user, not a Google employee, so take this with a pinch of salt until
confirmed/refuted).
- From when an instance starts, to when it dies, it is at any time in one
of 3 states
- ACTIVE: actively serving a request (ie from receiving an HTTP request
until it finishes serving the HTTP response)
- IDLE: inactive, not serving a request but still notionally allocated to
a machine and in memory, will NOT be killed
- FREE-IDLE: the process itself is in precisely the same state as "idle",
but is NOT billed, and is eligible to be killed at any time
- You are billed for total instance time of all instances in states
"active" and "idle", but you are NOT billed for any time spent in "free
idle" state
- Only "free idle" instances are truly killed (other than failure, load
balancing etc)
- The "free idle" state is how GAE attempts to optimise scheduling and
fault tolerance without thrashing etc without costing you anything
The question now is how an instance starts, ends, and changes from one state
to another.
- An instance starts when either
- it is needed to serve a request (there are no IDLE or FREE-IDLE
instances available subject to max-pending-latency) - starts in ACTIVE state
- or you have specified an "always-on" count - instances will start in
IDLE state until "always-on" count is reached
- with this regards to "always-on", I don't know if what you specify
is a number of machines you in "ACTIVE or IDLE" state, or "ACTIVE or IDLE or
FREE-IDLE" state
- Any instance becomes ACTIVE when the scheduler allocates it a request to
serve
- An ACTIVE instance becomes IDLE when it finishes serving a request
- (with concurrency, this is actually "when it is serving 0 requests")
- An IDLE instance becomes FREE-IDLE when either
- it has been IDLE for 15 consecutive minutes
- or the number of IDLE instances exceeds your max-idle-instances count
and this instance is chosen to become FREE-IDLE
- or something within GAE (fault tolerance, load, outages) needs to
otherwise kill your IDLE instance
- that is, they can switch any IDLE instance to FREE-IDLE and then
kill it at any time subject to their own discretion
Now looked at in this way, you can see that "billed instance time" is not
the same as "actual instance time" (what you see on your dashboard doesn't
currently distinguish between IDLE and FREE-IDLE) but that "billed instance
time" is essentially "time spent active serving a single request, including
waiting for synchronous API calls to complete" plus "time spent in IDLE
state". This is not the same "CPU cycles consumed", but more like "share of
the capacity" (like paying for time in an internet cafe).
But the big point is that the FREE-IDLE state is completely unbilled, and
you have some control over when an instance enters this state - it's not
true to say "all instances will be billed for a minimum of 15 minutes once
started" as we all first heard, but you should instead see that "up to
max-idle instances number of instances in the IDLE state will remain so for
15 minutes without activity before entering FREE-IDLE" state.
So what's the point of IDLE state for you?
Well, if you know you have a big startup time (hello Java) or you use
in-process-memory state caching (rather than, say, memcache) then IDLE
instances are your best friend, and you'll want to set a max-idle-instances
sufficient to handle the rate of traffic increases for bursts, but if your
startup time is minimal (hello python-with-webapp and handlers that only use
memcache/datastore to preserve state) then IDLE instances deliver minimal
benefit, and you'll want to set IDLE instances to a low number, maybe 1 or
even 0 if GAE agree to allow such a setting.
This mental model is what's put the smile back on my face ... here's hoping
Jon and Greg don't come back to tell me I'm completely wrong...
Cheers
--
Tim
--
You received this message because you are subscribed to the Google Groups
"Google App Engine" group.
To view this discussion on the web visit
https://groups.google.com/d/msg/google-appengine/-/78I4vMk1Al8J.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to
[email protected].
For more options, visit this group at
http://groups.google.com/group/google-appengine?hl=en.
