The traditional information theoretic approach to studying feedback is to consider ideal instantaneous high-rate feedback of the channel outputs to the encoder. This was acceptable in classical work because the results were negative: Shannon pointed out that even perfect feedback often does not improve capacity and in the context of symmetric DMCs, Dobrushin showed that it does not improve the fixed block-coding error exponents either in the interesting high rate regime. However, we have recently shown that perfect feedback *does* allow great improvements in the asymptotic tradeoff between delay and probability of error, even for symmetric channels at high rate. Since gains are claimed with ideal instantaneous feedback, it is natural to wonder whether these improvements remain if the feedback is unreliable or otherwise limited. Here, we consider packet erasure channels on both the forward and feedback links. First, we consider the feedback channel as a given, and show how to optimally balance forward and feedback error correction in the suitable information-theoretic limit. This shows that at high enough rate, the perfect-feedback performance is asymptotically attainable despite having only unreliable feedback! Second, we consider the more physical zero-sum case where the allocation of bandwidth to the feedback channel comes at the direct expense of the forward channel. It turns out that even here, feedback is worthwhile since dramatically lower asymptotic delays are possible by appropriately balancing forward and feedback error correction.