The current implementation of csv time series source loads the full time series at initialization. This can be very resource consuming, if the time series is long or you'd like to load a lot of time series at the same time.

The targeted implementation shall use a window buffer with configurable look ahead interval. This means, only chunks of the time series are read in and held in storage at the same time. However, this has the drawback, that this source only supports linear querying resp. no possibility to jump back in time.