We examine the impact of non-Gaussian photometry errors on photometric redshift performance. We find that they greatly increase the scatter, but this can be mitigated to some extent by incorporating the correct noise model into the photometric redshift estimation process. However, the remaining scatter is still equivalent to that of a much shallower survey with Gaussian photometry errors. We also estimate the impact of non-Gaussian errors on the spectroscopic sample size required to verify the photometric redshift rms scatter to a given precision. Even with Gaussian {\it photometry} errors, photometric redshift errors are sufficiently non-Gaussian to require an order of magnitude larger sample than simple Gaussian statistics would indicate. The requirements increase from this baseline if non-Gaussian photometry errors are included. Again the impact can be mitigated by incorporating the correct noise model, but only to the equivalent of a survey with much larger Gaussian photometry errors. However, these requirements may well be overestimates because they are based on a need to know the rms, which is particularly sensitive to tails. Other parametrizations of the distribution may require smaller samples.