Boat Load of Questions (which might turn in to feature requests)
Drake
Even with threading I don’t seem to be able to get an F8 backend to do more than 1.9 CPU Seconds per second. Which seems to be the same number I get on an F4 Backend.
Am I doing something wrong? Or is the Python Gil not capable of using all the CPU on this system? Or is an F8 CPU Second twice as much as an F4 and this is as expected. If the latter why is the app only 15% faster on an F8 (I know you can’t answer that with out the code)
If the gil is limited case why not allow backends to have more than one request at a time? If the Gil is not limited still why can I only have 1 request per backend?
Why can’t I configure how many requests Frontends handle? F1’s serving 8 requests is silly. F4’s handling 8 requests is often equally silly in the opposite direction.
Why can’t I configure how many requests a backend can handle? Or have a back end act like a front end and handle 8?
Why can’t I specify a specific version of an App in Apps For domains? It would be very useful to have User.Somedomian.com be a version of an app and Affiliates.somedomian.com be an app and they would share a database and nothing else.
Public backends can run longer than 60s but only handle one request. But they 503 if the request before them in line takes more than 10s to complete. Why is this not configurable? Especially since dynamic Public frontends seem to only spin up a new instances if there are 8 requests in the queue. This means you get a lot of 503’s
Defer task buckets don’t have a mechanism for manual token addition. I’d like this. It would let me read the queue and decide how many backends to spin up, and when tasks completed in 1 to 500s I would put a token in the bucket. (that’s not a question so much as a Does this make sense?)
Memcache gets slow sometimes. Really slow. Slower than datastore. Is there a “memcache health status” that we can hit to identify when we should stop using it for a while? MasterSlave had a Readonly we could detect and change behavior. I can’t find any such thing for HR.
Requests through edge cache get edge cached. Why do outgoing requests not do the same? There seems to be a lot of people “scraping” content via Appengine. Why are those requests not “edge” cached so that a. it would be harder to build appengine DOS attacks. B. fetches would take less time.
Because I am threaded and use 200% more CPU seconds than Realtime seconds I suspect I’m a bad neighbor to share an instance with. Is this true? Am I clobbering the neighbor?
Do we always have to go through Apps For Domains? How much money do I have to send you to never have to talk to their tech support again? Similar question Naked Domians. I want one. Building a URL Shortener that has to have www. On the front is breaking my heart.
Brandon Wirtz |
| |||
| ||||
Marcel Manz
In my use case I would be highly interested in reading the current add to taskqueue latency. So when app engine has a hiccup I don't have to wait for enqueuing tasks but rather could process them directly.
Certainly I could measure those latencies myself and store the information in memcache for subsequent requests, but I would prefer if those already GAE measured latencies could be read via an API, such as an extended capabilities API?
https://developers.google.com/appengine/docs/java/javadoc/com/google/appengine/api/capabilities/CapabilityState
Barry Hunter
>
> Certainly I could measure those latencies myself and store the information
> in memcache for subsequent requests,
operating out of multiple data-centres.
So the latencies you see accessing APIs are only really going to apply
to your little corner. The (taskqueue) server you happen could be
seeing very different access patterns to another one.
In fact different instances of your app, might even be experiencing
different latencies, because they may be seperated into different
parts. Although it makes sense for AppEngine to keep your instances
clustered so can for example use a local memcache server.
So rather than memcache, use instance storage to store API health data?
--
Its one of the reasons you see so many people complaining about the
AppEngine Status site - it can be showing 'all fine' when someones
particular app is massivly degraded.
AppStatus would have to make sure it has instances running in every
corner of AppEngine datacenters, and possible report
'best/worse/average' or something.
Marcel Manz
As for the task queue latency, my app so far always could see very much the same delays as reported in the system status. I hardly believe my app all the time lives in the same part of the system as the probe that is providing measurement data to the system status. So how come, this matches so nicely (at least for tasks add) ?
Kyle Finley
Why can’t I configure how many requests Frontends handle? F1’s serving 8 requests is silly. F4’s handling 8 requests is often equally silly in the opposite direction.
Do we always have to go through Apps For Domains? How much money do I have to send you to never have to talk to their tech support again? Similar question Naked Domians. I want one. Building a URL Shortener that has to have www. On the front is breaking my heart.
Kristopher Giesing
On Sunday, July 29, 2012 12:43:31 PM UTC-7, Brandon Wirtz wrote:
Why can’t I configure how many requests a backend can handle? Or have a back end act like a front end and handle 8?