This paper addresses the problem of streaming packetized media over a lossy packet network, with sender-driven (re)transmission using acknowledgement feedback. The different transmission scenarios associated to a group of interdependent media data units are abstracted in terms of a finite alphabet of policies, for each single data unit. A rate-distortion optimized markovian framework is proposed, which supports the use of comprehensive feedback information. Contrarily to previous works in rate-distortion optimized streaming, whose transmission policies definitions do not take into account the feedback expected for other data units, our framework considers all the acknowledgment packets in defining the streaming policy of a single data unit. More specifically, the notion of master and slave data unit is introduced, to define dependent streaming policies between media packets; the policy adopted to transmit a slave data unit becomes dependent on the acknowledgments received about its masters. One of the main contributions of our work is to propose a methodology that limits the space of dependent policies for the RD optimized streaming strategy. A number of rules are formulated to select a set of relevant master/slave relationships, defined as the dependencies that are likely to bring RD performance gain in the streaming system. These rules provide a limited complexity solution to the rate-distortion optimized streaming problem, with comprehensive use of feedback information. Based on extensive simulations, we conclude that (i) the proposed set of relevant dependent policies achieves close to optimal performance, while being computationally tractable, and (ii) the benefit of dependent policies is driven by the relative sizes and importance of interdependent data units. Our simulations demonstrate that dependent streaming policies can perform significantly better than independent streaming strategies, especially for cases where some media data units bring a relatively large gain in distortion, in comparison with other data units they depend on for correct decoding. We observe however that the benefit becomes marginal when the gain in distortion per unit of rate decreases along the media decoding dependency path. Since such a trend characterizes most conventional scalable coders, the implementation of dependent policies can reasonably be ruled out in these specific cases.