Systems and methods for performing near real-time merging of distributed data streams are described. For example, streams of ad impressions, ad clicks, and conversions are sorted by user id into virtual buckets. The buckets of data are distributed across multiple servers, so that each server can process their respective buckets of data independently. Each server uses synchronization logic to determine a running delay distribution of the data streams. Based on the delay distributions, merge processing of the streams is appropriately delayed to ensure that the ad impression and ad click stream information needed for correlating with the conversion stream information is likely to be available in real time.