We have examined historical distributions of sea surface salinity (SSS) observations in a data set consisting of a combination of the World Ocean Database 1998 (WOD98) and a thermosalinograph and bucket salinity database collected from volunteer observing ships. It is well known that SSS in much of the world's ocean is measured infrequently or not at all. We find that 27% of one-degree squares in the world ocean (open and coastal, excluding the Arctic Ocean) had no observations of SSS in the historical database, and 70% had 10 or fewer. Systematic sampling of SSS (more than 10,000 observations per year globally) did not start until after 1960. Most SSS observations in the WOD98 are concentrated in the North Sea and coast of northern Europe, the east and west coasts of North America, and around Japan. About 28% of SSS measurements are in coastal waters. We plotted frequency histograms of SSS for some selected well-sampled one-degree squares in the North Atlantic and tropical Pacific. We found most frequency histograms to be non-Gaussian. The main departure from normal distribution is due to anomalous low-salinity measurements creating a negative skewness. This conclusion is verified as a global phenomenon by examining statistics of mean-median SSS difference within one-degree squares. This quantity is found to be predominantly negative over the global ocean. These anomalous low-salinity values may be due to rainfall events, but there are other plausible physical mechanisms, like frontal movement and eddy activity. There were also areas where the distributions were bimodal due to the presence of meandering fronts with little cross-frontal mixing. Examples are shown where the non-Gaussian nature of the distributions in the areas examined is both a short-term and a long-term phenomenon. That is, the distributions are skewed on a nearly instantaneous (~1 month) basis and averaged over long time periods (1 + years). This has important implications for climatologies because the differences between mean and modal SSS, for the analyzed one-degree squares, is of order 0.1. Furthermore, the implication for validation studies for remote sensing missions is that the studies must make enough measurements of SSS to determine the extent to which the probability density is not Gaussian.