Azure Stream Analytics: Specified cast is not valid - azure

Question

I'm having a hard time tracking down a casting error in Azure Stream Analytics. The input data is coming from an Azure IoT Hub. Here's my query code:
-- Create average data from raw data
WITH
AverageSensorData AS
(
SELECT
[Node],
[SensorType],
udf.round(AVG([Value]), 2) AS [Value],
MAX(TRY_CAST([Timestamp] AS DateTime)) AS [Timestamp]
FROM [SensorData]
WHERE TRY_CAST([Timestamp] AS DateTime) IS NOT NULL
GROUP BY
[Node],
[SensorType],
TumblingWindow(minute, 2)
)
-- Insert average data into PowerBI-Output
SELECT
[Node],
[SensorType]
[Value],
[Timestamp],
DATETIMEFROMPARTS(
DATEPART(yy, [Timestamp]),
DATEPART(mm, [Timestamp]),
DATEPART(dd, [Timestamp]),
0, 0, 0, 0) AS [Date]
INTO [SensorDataAveragePowerBI]
FROM [AverageSensorData]
While this runs fine most of the time (at least for a couple of hundreds or thousands of input entities) it will eventually fail. After having turned on Diagnostic logs I was able to find the following eror message in the corresponding execution log (in reality it was in JSON format, I cleaned it up a little for readability):
Message: Runtime exception occurred while processing events, - Specified cast is not valid. OutputSourceAlias: averagesensordata; Type: SqlRuntimeError, Correlation ID: 49938626-c1a3-4f19-b18d-ee2c5a5332f9
And here's some JSON input that probably caused the error:
[
{
"Date": "2017-04-27T00:00:00.0000000",
"EventEnqueuedUtcTime": "2017-04-27T07:53:52.3900000Z",
"EventProcessedUtcTime": "2017-04-27T07:53:50.6877268Z",
"IoTHub": {
/* Some more data that is not being used */
},
"Node": "IoT_Lab",
"PartitionId": 0,
"SensorType": "temperature",
"Timestamp": "2017-04-27T09:53:50.0000000",
"Value": 21.06
},
{
"Date": "2017-04-27T00:00:00.0000000",
"EventEnqueuedUtcTime": "2017-04-27T07:53:53.6300000Z",
"EventProcessedUtcTime": "2017-04-27T07:53:52.0157515Z",
"IoTHub": {
/* Some more data that is not being used */
},
"Node": "IT_Services",
"PartitionId": 2,
"SensorType": "temperature",
"Timestamp": "2017-04-27T09:53:52.0000000",
"Value": 27.0
}
]
The first entry was the last one to go through, so the second one might have been the one breaking everything. I'm not sure, though, and do not see any suspicious values here. If I upload this as test data within the Azure portal then no errors are being raised.
The query above explicitely uses casting for the [Timestamp] column. But since I'm using TRY_CAST I wouldn't expect any casting errors:
Returns a value cast to the specified data type if the cast succeeds; otherwise, returns null.
As I said, the error only appears once in a while (sometimes after 20 minutes and sometimes after a couple of hours) and cannot be reproduced explicitely. Does anyone have an idea about where the error originates or if there is a chance of getting more detailed error information?
Thanks a lot in advance.
UPDATE: Here's the source of the udf.round function:
function main(value, decimalPlaces) {
if (typeof(value) === 'number'){
var decimalMultiplier = 1;
if (decimalPlaces){
decimalMultiplier = Math.pow(10, decimalPlaces);
}
return Math.round(value * decimalMultiplier) / decimalMultiplier
}
return value;
}
Unfortunately, it's been a while since I wrote this, so I'm not a hundred percent sure why I wrote exactly this code. One thing I do remember, though, is, that I all the analyzed messages always contained a valid number in the respective field. I still think that there's a good chance of this function being responsible for my problem.