This paper presents an in-depth study of YouTube video service delivery. We have designed a tool that crawls YouTube videos in order to precisely evaluate the quality of experience (QoE) as perceived by the user. We enrich the main QoE metric, the number of video stalls, with many network measurements and use multiple DNS servers to understand the
main factors that impact QoS and QoE. This tool has been used in multiple configurations: first, to understand the main delivery policies of YouTube videos, then to understand the impact of the ISP on these policies and finally, to compare the US and Europe YouTube policies. Our main results are that: (i) geographical proximity does not matter inside Europe or the US, but link cost and ISP-dependent policies do; (ii) usual QoS metrics (RTT) have no impact on QoE (video stall); (iii) QoE is not impacted nowadays (with good access
networks) by access capacity but by peering agreements between ISPs and CDNs, and by server load. We also indicate a network monitoring metric that can be used by ISPs to roughly evaluate the QoE of HTTP video streaming of a large set of clients at a reduced computational cost.