Hi. In other indicators like ratio, latency, series are summed but not for native latency, so for each instance (kubernetes pod) slo is calculated separately. Following record rule is generated:
histogram_fraction(0, 100, rate(http_server_duration{service_name="my-service"}[12h]))
bur following is expected:
histogram_fraction(0, 100, sum(rate(http_server_duration{service_name="my-service"}[12h]))).

Also grouping seems does not have any affect to the output.