Aggregations involving time-series should only process documents where values for metric fields are not null. This is in line with PromQL that implicitly selects time-series containing values and tracks the ones with no recent data as stale.

A workaround is to explicitly filter out null values, e.g.

TS my-metrics
 | WHERE cpu_usage IS NOT NULL
 | STATS max(avg_over_time(cpu_usage)) BY TBUCKET(1 hour), hostname

Still, adding WHERE <metric> IS NOT NULL for all metrics in all queries is fairly verbose and redundant. It should be implicitly applied within the TS command scope.