Issue with async Python SDK for Cosmos DB (SQL)

Question

Hi all,

I am using azure.cosmos.aio.CosmosClient from azure-cosmos=4.3.1 in my project.

I created a singleton client for better performances:

• instantiate CosmosClient once with retry policy
• call get_database_client once (I use one database in project)
• for every request, obtain new container proxy, use it for a query (mostly read) and close it

My retry policy looks like this:

COSMOS_RETRY_POLICY = {    
    "request_timeout": 2000,    
    "retry_read": 10,   
    "retry_connect": 10,    
    "retry_total": 10,    
    "retry_status": 10
}

So the issue occurs when number of requests is around 60 per minute (around 30 failures in half hour).

For some requests, I receive status 500 with aiohttp.client_exceptions.ServerTimeoutError as response.

So my questions are:

1. What is the maximum number of simultaneous requests which Cosmos DB can handle?
2. Is there a way to configure connection pool for client?
3. How should retry policy look like for higher loads (10 requests/second)?
4. Would introducing dedicated gateway help?

Answer 1

httpx can not work after all, I am not able to handle requests with enable_cross_partition_query set to true

Answer 2

Hello Geetha,

After further investigation, I figured that issue is caused by aiohttp

Traceback (most recent call last):
  
File "/usr/local/lib/python3.10/site-packages/aiohttp/connector.py", line 980, in _wrap_create_connection
    return await self._loop.create_connection(*args, **kwargs)  # type: ignore[return-value]  # noqa
  
File "/usr/local/lib/python3.10/asyncio/base_events.py", line 1103, in create_connection
    transport, protocol = await self._create_connection_transport(
  
File "/usr/local/lib/python3.10/asyncio/base_events.py", line 1133, in _create_connection_transport
    await waiter
asyncio.exceptions.CancelledError
During handling of the above exception, another exception occurred:

Any idea on how to fix this?

More info about server:

• Containers are not big, the biggest one contains around 50k items
• Provisioned throughput is 2000 RU/s (manual), and we use around 40% of that in peak