Detailed description
This is a Breaking change when using window functions inside partition operator sub-query scope.
Window functions require a serialized/ordered input.
There is a bug in Azure Data Explorer logic that allows using window functions inside the partition operator when the ordering is performed outside the partition query scope like this:
T
| where Timestamp > ago(1d)
| order by Timestamp asc
| partition by key
(
extend next_value = next(value)
...
)
This kind of query currently works unintentionally and sometimes the output is not correct.
We’re planning to make a change that will fail this kind of queries by throwing a semantic error indicating that the input to the window function is not serialized.
Required change
Fix all relevant queries by pushing the ordering inside the partition subquery like shown below.
T
| where Timestamp > ago(1d)
| partition by key
(
order by Timestamp asc
| extend next_value = next(value)
...
)
Schedule & plan
Support for previous window function pattern will be blocked - ETA: November 30, 2022
