CMB data analysis is in general done through two main steps : map-making of the time data streams and power spectrum extraction from the maps. The latter basically consists in the separation between the variance of the CMB and that of the noise in the map. Noise must therefore be deeply understood so that the estimation of CMB variance (the power spectrum) is unbiased. I present in this article general techniques to make maps from time streams and to extract the power spectrum from them. We will see that exact, maximum likelihood solutions are in general too slow and hard to deal with to be used in modern experiments such as Archeops and should be replaced by approximate, iterative or Monte-Carlo approaches that lead to similar precision.