《API Performance Monitoring》.pdfVIP

API Performance Monitoring George Schlossnagle, Message Systems If we have data, let’s look at the data. If all we have are opinions, lets go with mine. -Jim Barksdale About Message Systems §  World’s largest provider of email messaging infrastructure §  Traditionally delivered as on-premise software, moved into the cloud recently -  ~ 30B message per day on-premise, ~250M messages per day in the cloud §  Platform fully accessible through multiple developer-focused APIs Some of our Customers Our Agenda §  General goals for monitoring §  Active monitoring §  Passive monitoring §  Some other topics A note about tools §  This is a talk about monitoring strategy and general techniques, not about specific tools §  Most of what we use is built around Circonus with a large custom code base of agents and infrastructure §  Plenty of alternative options: -  Nagios -  Graphite -  Reconnoiter §  Depending on your platform, investing some time in R or NumPY may be useful General Qualities We Want in Monitoring §  Real Time Analysis §  Trending/Analytics §  Anomaly Detection §  Alerting §  Ease of Deployment and Provisioning The Difference Between Monitoring and Alerting §  Monitoring -  Collecting telemetry data that may be interesting -  Good for analytics/trending -  Good for contextualizing issues -  Good for understanding typical runtime behavior -  Can drive and inform alerting strategy §  Alerting -  Should only be used for actionable events -  Beware of ‘boy who cried wolf’ -  Thresholds should be a compromise between reality and your ideals Two Types of Monitoring §  Active Monitoring -  Trigge