Different strategies for implementing ensemble-based data assimilation systems are discussed. Ensemble filters like ensemble Kalman filters and particle filters can be implemented so that they are nearly independent from the model into which they assimilate observations. This allows to develop implementations that clearly separate the data assimilation algorithm from the numerical model. For coupling the model with a data assimilation software one possibility is to use disk files to exchange the model state information between model and ensemble data assimilation methods. This offline coupling does not require changes in the model code, except for a possible component to simulate model error during the ensemble integration. However, using disk files can be inefficient, in particular when the time for the model integrations is not significantly larger than the time to restart the model for each ensemble member and to read and write the ensemble state information with the data assimilation program. In contrast, an online coupling strategy can be computational much more efficient. In this coupling strategy, subroutine calls for the data assimilation are directly inserted into the source code of an existing numerical model and augment the numerical model to become a data assimilative model. This strategy avoids model restarts as well as excessive writing of ensemble information into disk files. To allow for ensemble integrations, one of the subroutines modifies the parallelization of the model or adds one, if a model is not already parallelized. Then, the data assimilation can be performed efficiently using parallel computers. As the required modifications to the model code are very limited, this strategy allows one to quickly extent a model to a data assimilation system. In particular, the numerics of a model do not need to be changed and the model itself does not need to be a subroutine. The online coupling shows an excellent computational scalability on supercomputers and is well suited for high-dimensional numerical models. Further, a clear separation of the model and data assimilation components allows to continue the development of both components separately. Thus, new data assimilation methods can be easily added to the data assimilation system. Using the example of the parallel data assimilation framework [PDAF, http://pdaf.awi.de] and the ocean model NEMO, it is demonstrated how the online coupling can be achieved with minimal changes to the numerical model.