facebook Mystery Machine 2017-06-05 cloud Data Flowconsistent sampling 1local log --> Scribe --> Hive --> UberTrace --> UI Log Schema1234request idhost idhost-local timestampunique event label(event name, task name) Timestamp Normalize 不考虑local clock skew 假设client/server间的RTT是对称的 123456789101112131415Client Server 1|------------>| | |--+1.1 | | | logic | |<-+1.2 2|<------------|1.2 - 1.1 = 0.12 - 1 = 1.0RTT = (1.0 - 0.1)/2 = 0.45clock(1.1) = 1 + 0.45 = 1.45clock(1.2) = 1.45 + 0.1 = 1.55RTT是个经验值,根据大量的trace后稳定下来: 使用最小值 Newer Kafka vs Kinesis vs Redis Older storm acker