App Engine Bulk Loader Performance

问题:

I am using the App Engine Bulk loader (Python Runtime) to bulk upload entities to the data store. The data that i am uploading is stored in a proprietary format, so i have implemented by own connector (registerd it in bulkload_config.py) to convert it to the intermediate python dictionary.

import google.appengine.ext.bulkload import connector_interface
class MyCustomConnector(connector_interface.ConnectorInterface):
   ....
   #Overridden method
   def generate_import_record(self, filename, bulkload_state=None):
      ....
      yeild my_custom_dict

To convert this neutral python dictionary to a datastore Entity, i use a custom post import function that i have defined in my YAML.

def feature_post_import(input_dict, entity_instance, bulkload_state):
    ....
    return [all_entities_to_put]

Note: I am not using entity_instance, bulkload_state in my feature_post_import function. I am just creating new data store entities (based on my input_dict), and returning them.

Now, everything works great. However, the process of bulk loading data seems to take way too much time. For e.g. a GB (~ 1,000,000 entities) of data takes ~ 20 hours. How can I improve the performance of the bulk load process. Am i missing something?

Some of the parameters that i use with appcfg.py are (10 threads with a batch size of 10 entities per thread).

Linked a Google App Engine Python group post: http://groups.google.com/group/google-appengine-python/browse_thread/thread/4c8def071a86c840

Update: To test the performance of the Bulk Load process, I loaded entities of a 'Test' Kind. Even though this entity has a very simple FloatProperty, it still took me the same amount of time to bulk load those entities.

I am still going to try to vary the bulk loader parameters, rps_limit, bandwidth_limit and http_limit, to see if i can get any more throughput.

回答1:

There is parameter called rps_limit that determines the number of entities to upload per second. This was the major bottleneck. The default value for this is 20.

Also increase the bandwidth_limit to something reasonable.

I increased rps_limit to 500 and everything improved. I achieved 5.5 - 6 seconds per 1000 entities which is a major improvement from 50 seconds per 1000 entities.