Downsampling fails if source index contains fields of type passthrough and these fields are configured with time_series_dimension to true. There are two problems that occur in this case:

• The target index that the downsample api tries to creates is with a flattened like mapping (fields with dots) and so the type of the object field is never specified (1).
• The downsample api detects all fields with time_series_dimension set to true as dimension fields. This isn't true as passthrough fields are just container fields.

1: This is a field mapping that the downsample tries to create:

 "metrics.process.memory.usage": { "type": "aggregate_metric_double", "metrics": [ "max", "min", "value_count", "sum" ], "default_metric": "max", "time_series_metric": "gauge" }, 

The metrics. is an object field here and but in the mapping that this fields get merged with it is a passthrough object field.