We provide some of the first rigorous evidence on performance spillovers and social network in the workplace. The data we use are rather extraordinary – weekly data for rejection rates (proportion of defective output) for all weavers in a firm during a 12 months (April 2003-March 2004) period, more than 10,000 observations. Our fixed effect estimates first point to significant spillovers of performance from high-ability weavers to low-ability weavers. On the other hand, we find no evidence for performance spillovers from low-ability to high-ability weavers. The findings are consistent with the knowledge sharing hypothesis that low-ability workers learn from high-ability workers but not vice versa. Second, by exploiting the well-documented fact that an exogenously-formed sharp divide between urban workers and rural migrant workers exists in firms in Chinese cities, we find that performance spillovers/knowledge sharing take place only within the confines of social network. Specifically rural low-ability weavers are found to improve their performance as their high-ability teammates (who are also rural migrants) improve their performance while they do not benefit from performance improvement of their high-ability teammates who are urban residents. Such heterogeneous performance interdependence of workers within the same team suggests that our evidence for performance spillovers is less likely to be a result of team specific demand shocks that generate spurious performance interdependence of all team members.