Hey,
We've decided to go to Europython this year. We're obviously going to give a talk about the exciting things we're doing with LAX and GoogleAppEngine. We're on wednesday at midday in the alfa room, check out the schedule here. Since we think it's important that these events take place, we're also chipping in and sponsoring the event.
We hope to see you there. Drop us a note if you want to meet up.
Here at Logilab we find Munin pretty useful. We monitor a lots of machines and a lot of services with it, and it usually gives us pretty useful indicators over time that guide us through to optimizations.
One of the reasons we adopted this technology is it's modular approach with the plugin architecture. And when we realized we could write plugins in python, we knew we'd like it. After years of using it, we're now actually writing plugins for it. Optimizing zope and zeo servers is not an easy task so we're developping plugins to be able to see the difference between before and after changing things.
You check out the project here, and download it from the ftp.
After one month of internship at logilab, I'm pleased to announce the 0.12.1 release of apycot.
for more information read the apycot 0.12.1 release note
You can also check the new sample configuration.
Pierre-Yves David
Here is a piece of code I've written which I thought may be useful to some other people...
You'll find here a simple python module to use with the Google AppEngine SDK to monkey patch the datastore API in order to get an idea of the calls performed by your application.
To instrument of the datastore, put at the top level of your handler file
import instrdatastore
Note that it's important to have this before any other import in your application or in the google package to avoid that some modules will use the unpatched version of datastore functions (and hence calls to those functions wouldn't be considered).
Then add at the end of your handler function
instrdatastore.print_info()
The handler file should look like this:
"""my handler file with datastore instrumenting activated""" import instrdatastore # ... other initialization code # main function so this handler module is cached def main(): from wsgiref.handlers import CGIHandler from ginco.wsgi.handler import ErudiWSGIApplication application = ErudiWSGIApplication(config, vreg=vreg) CGIHandler().run(application) instrdatastore.print_info() if __name__ == "__main__": main()
Now you should see in your logs the number of Get/Put/Delete/Query which has been done during request processing
2008-06-23 06:59:12 - (root) WARNING: datastore access information
2008-06-23 06:59:12 - (root) WARNING: nb Get: 2
2008-06-23 06:59:12 - (root) WARNING: arguments (args, kwargs):
((datastore_types.Key.from_path('EGroup', u'key_users', _app=u'winecellar'),), {})
((datastore_types.Key.from_path('EUser', u'key_test@example.com', _app=u'winecellar'),), {})
2008-06-23 06:59:12 - (root) WARNING: nb Query: 1
2008-06-23 06:59:12 - (root) WARNING: arguments (args, kwargs):
(({'for_user =': None}, 'EProperty'), {})
2008-06-23 06:59:58 - (root) WARNING: nb Put: 1
2008-06-23 06:59:58 - (root) WARNING: arguments (args, kwargs):
(({u'login': None, u'last_usage_time': 1214204398.2022741, u'data': ""},), {})
I'll probably extend this as the time goes. Also notice you may encounter some problems with the automatic reloading feature of the dev app server when instrumentation is activated, in which case you should simply restart the web server.
Bienvenue à SciLab version 5.0 dans le monde du logiciel libre. SciLab 5.0, plateforme open source de calcul scientifique sous licence CeCill, est une alternative crédible et maintenant reconnue comme telle à Matlab. Pour assurer le développement pérenne de Scilab, le consortium Scilab rejoint DIGITEO, parc de recherche d'envergure mondiale dans le domaine des sciences et technologies de l'information en Île-de-France.
Previous documentation was merged into a LAX Book now featuring step-by-step screenshots to get up and running faster.
http://lax.logilab.org/lax-book
Don't we all like screenshots...
Update: LAX is now included in the CubicWeb semantic web framework.
At Google IO, a large part of the Tools track was dedicated to AppEngine. Brett Slatkin gave a talk titled Building scalable Web Applications with Google AppEngine which focused on optimizing the server part of web apps. As other presenters demonstrated it, like Steve Souders in his talk Even Faster Websites, optimizing the browser part of webapps is not to be neglected either.
First of all, I must confess I am used to repeat that "early optimisation is the root of all evil" and "delay commitment until the last responsible time". But reading about AppEngine and listening to the Google IO talks, it appears that the tools we have today ask for human intervention to reach web-scale performance, even when "we" stands for "Google".
In order for web-scale applications to handle the kind of load they are facing, they must be designed and implemented carefully. As carefully as any application was designed before the exponential growth of PC computation power let us move away from low-level implementation details and made some inefficiencies acceptable as long as the time spent developing was short enough.
It all depends on the parameters of your cost function, but for web-scale applications, it seems like we have not enough computer-time and can not trade it for human-time.
To get a better idea of the work constraints, one should know that a disk seek is about 10ms, which means there will be a maximum of 100 accesses per second. On the other hand, if we need consistent data as opposed to transactional data (the latter implying that data is fetched each time it is asked for), data can be read from disk once then cached. Following reads are done from memory at a rate of about 4GB/sec, which means 4000 accesses per second if entities are around 1MB in size. Result of this back of the envelope approximation is 40 reads equals one write.
It follows that, although the actual time depends on the size and shape of data, writes are very expensive compared to reads and both are better done in batches to optimise disk access.
The AppEngine Datastore was designed with this constraints in mind. Entities are sets of property name/value pairs. Each entity may have a parent. An entity without a parent is the root of a hierarchy called an entity group.
Entities of the same group are stored on disk close to each other, but two distinct entity groups may be stored on different computers. Read access to entities of the same group is thus faster than read access to entities of different groups.
Write access is serialized per entity group. As opposed to a traditionnal RDBMS that provides row locking, the datastore only provides entity group locking. Writes to the a single entity group will always happen in sequence, even though changes concern different entities.
There is no limit to the number of entity groups or to the number of entities per group, but because of the locking strategy, large entity groups will cause high contention and a lot of failed transactions. Since writes are expensive, not thinking about write throughput is a very bad idea when designing an AppEngine application if one want it to scale.
On the other hand, the parallel nature of the datastore make it scale wide and there is no limit to the number of entity groups that can be written to in parallel, nor to the number of reads that can be done in parallel.
To understand this design in details, you will have to read about GFS, BigTable and other technologies developed by Google to implement large-scale clustering.
Counters are a good example to address when discussing write throughput, because the datastore locking strategy makes writing to global data very expensive.
Let us assume that we want to display on the main page of a wiki application the total number of comments posted.
A global counter would serialize all its updates. If 100 users were to add comments at the same time, some of them would have to wait several seconds for their action to complete: one write for the comment, one write for the counter, at most 100 writes per second for the counter and a lot of time lost due to failed transaction that need to be restarted.
The solution to make the counter scale is to partition it among all entity groups then sum these partial counters when the global value is needed.
Since chances are low that a given user will write more than one comment at a time, comment entities for a user can be grouped together and a partial counter can be added to the same entity group. Creating a new comment and increasing the partial counter will be done in the same batch.
When a new request for the main page is received, the counter total is looked up in the cache. If it is not found, all partial counters are fetched and summed up, then the cache is refreshed with a short timeout, for example one minute.
During the next minute, the counter will be "consistent", read no too far-off, and served extremely fast from the cache.
To sum things up, when implementing applications on top of AppEngine with web-scale usage as a goal, everything that can be done to save time should be considered. Including the following:
As a conclusion, the interface AppEngine is exhibiting today requires to optimize early, but I would bet that in the years to come, new languages and domain-specific compilers or database engines will take part of that burden off the hands of the developers.
Did not Yahoo and Google start developping PigLatin and Sawzall to make it easier to write parallel data-processing programs ? The same could happen with describe a data-model in a high-level language and get a tool to optimize it for write contention and web-scale application.
LAX (Logilab App engine eXtension) is a full-featured web application framework running on Google AppEngine developed by Logilab.
Several of us went to San Francisco last week to attend Google IO. As usual with conferences, meeting people was more interesting than listening to most talks. The AppEngine Fireside Chat was a Q&A session that lasted about an hour. Here is what I learned from this session and various chats with AppEngineers.
My conclusion is twofold:
This is why Logilab published its own framework running on App Engine as free software and is providing support for it: Logilab Appengine eXtension.
La version 0.3.0 de LAX est sortie aujourd'hui voir : http://lax.logilab.org/
Il suffit de 10 petites minutes pour avoir une application qui tourne, suivez le guide :
Mise à jour: LAX est maintenant inclus dans CubicWeb.
LAX version 0.3.0 was released today, see http://lax.logilab.org/
Get a new application running in ten minutes with the install guide and the tutorial:
Enjoy!
Update: LAX is now included in the CubicWeb semantic web framework.