Group and sum double counts

we have topics and partitions and sizes for each pair, I want the total size of the topic across all partitions
I thought, this would simply be group by topic and sum the value

time						topic	partition	value
2023-09-04T15:10:00.000Z	A		0			2
2023-09-04T15:10:00.000Z	B		0			3
2023-09-04T15:10:00.000Z	B		1			4

2023-09-04T15:15:00.000Z	A		0			2
2023-09-04T15:15:00.000Z	B		0			3
2023-09-04T15:15:00.000Z	B		1			5

desired output:

time						topic	value
2023-09-04T15:10:00.000Z	A		2
2023-09-04T15:10:00.000Z	B		7

2023-09-04T15:15:00.000Z	A		2
2023-09-04T15:15:00.000Z	B		8

the following query:

|> filter(fn: (r) => r.topic == "A")
|> aggregateWindow(every: 5m, fn: mean, createEmpty: false)
|> group(columns:["_time", "topic"])
|> sort(columns: ["_time"], desc: true)
|> aggregateWindow(every: 5m, fn: sum, createEmpty: false)

the sum result depends on the time window I choose so it’s obviously double (or more) counting
If I replace the sum in the last line by an count I see results > 1 for every topic