A one-unit repairable system operating under three types of weathers, normal, abnormal 1 and abnormal 2, is analysed. Weather changes from one type to another at different exponential rates. Failure time distributions of the system from all types of weathers are assumed to be exponential whereas repair time distributions are different and general. In order to account for the effects of weather changes during a repair, along with keeping the model computationally tractable, the repair time distributions are chosen as general Erlang-3 distributions which allow good approximation of a broad class of repair times. The system is analysed as a Markov process to find mean time to system failure (MTSF) and absorption probabilities, as a closed queueing network to obtain steady-state availability and as a M/G/I queueing model to evaluate mean repair time. The results are expressed in computationally tractable form. Some numerical computations have been done which help us to draw some significant conclusions.