MetricsQueryClient giving different results with different timespans

Question

I'm using the Python MetricsQueryClient to list out how many tokens were used on certain days via the APIM policy "azure-openai-emit-token-metric". The problem is that when I call the query_resource() function with "timespan" set for the entire month of October, I get different results for token count usage for today's date than when I set the "timespan" to just the last 48 hours. For example, when setting the timespan to be from 10/20/2024 to 10/22/2024, I see 34 prompt tokens for today's date. But if I set the timespan to be 10/1/24 to 11/1/24, I see 0 prompt tokens for today's date.

 

Is this a known issue? It is documented somewhere? 

Answer 1

Hi Ben,

This is likely due to a delay in Azure’s metric aggregation for recent data. When querying for a smaller, more recent window, the data is near real-time. But when you query for a longer period, such as an entire month, today's data may not have fully propagated or been processed yet, so it doesn’t appear in the results. To work around this, it’s a good idea to split the queries: use a smaller time range to get recent data and a larger range for older data. That way, you can ensure you're capturing today's usage accurately.