Three Pillars, No Answers: Helping Platform Teams Solve Real Observability Problems
0 likes173 views
The document critiques conventional approaches to observability in microservices, highlighting flaws in the reliance on metrics, logs, and distributed tracing. It provides a new scorecard for evaluating observability that includes detection and refinement activities focused on service level indicators (SLIs) and their consumer perceptions. Ultimately, it emphasizes the need for high specificity, fidelity, and freshness in observability data while proposing a context-rich and real-time approach to refining performance measurements.
Three Pillars, No Answers: Helping Platform Teams Solve Real Observability Problems
- 1. Three Pillars, No Answers: Helping Platform Teams Solve Real Observability Problems Austin Parker, Principal Developer Advocate at Lightstep
- 2. Who Am I? Austin Parker Principal Developer Advocate @austinlparker austin@lightstep.com✉
- 4. The Conventional Wisdom ● Observing microservices is hard ● Google and Facebook solved this (right???) ● They used Metrics, Logging, and Distributed Tracing… ● So we should, too.
- 5. The Three Pillars of Observability - Metrics - Logging - Distributed Tracing
- 6. Metrics
- 7. Logging
- 8. Tracing
- 10. Fatal Flaws
- 11. A word nobody knew in 2015… Dimensions (aka “tags”) can explain variance in timeseries data (aka “metrics”) …… but cardinality
- 12. Logging Data Volume: a reality check transaction rate x all microservices x cost of net+storage x weeks of retention ----------------------- way too much $$$$
- 13. The Life of Transaction Data: Dapper Stage Overhead affects… Retained Instrumentation Executed App 100.00% Buffered within app process App 000.10% Flushed out of process App 000.10% Centralized regionally Regional network + storage 000.10% Centralized globally WAN + storage 000.01%
- 14. Fatal Flaws: A Review Logs Metrics Dist. Traces TCO scales gracefully – ✓ ✓ Accounts for all data (i.e., unsampled) ✓ ✓ – Immune to cardinality ✓ – ✓
- 16. Data vs UI
- 17. Data vs UI
- 19. Metrics, Logs, and Traces are Just Data, … not a feature or use case.
- 20. Part 2: A New Scorecard for Observability
- 21. Mental Model: Goals and Activities ● Goals: how our services perform in the eyes of their consumers ● Activities: what we (as operators) actually do to further our goals
- 22. Quick Vocab Refresher: SLIs “SLI” = “Service Level Indicator” TL;DR: An SLI is an indicator of health that a service’s consumers would care about. … not an indicator of its inner workings
- 23. Observability: 2 Fundamental Goals Gradually improving an SLI Rapidly restoring an SLI Reminder: “SLI” = “Service Level Indicator” NOW!!!! days, weeks, months…
- 24. Observability: 2 Fundamental Activities 1. Detection: measuring SLIs precisely 2. Refinement: reducing the search space for plausible explanations
- 25. An interlude about stats frequency
- 26. Scorecard: Detection 1. Specificity: - Cost of cardinality ($ per tag value) - Stack support (mobile/web platforms, managed services, “black-box OSS infra” like Kafka/Cassandra) 2. Fidelity: - Correct stats!!! (global p95, p99) - High stats frequency (stats sampling frequency, in seconds) 3. Freshness (lag from real-time, in seconds)
- 27. Why “Refinement”? # of things your users actually care about # of microservices # of failure modes Must reduce the search space!
- 28. The Refinement Process Discover Variance Explain Variance Deploy Fix
- 30. Scorecard: Refinement Identifying Variance: - Cardinality ($ per tag value) - Robust stats (histograms (see prev slide)) - Retention horizons for plausible queries (time duration) Explaining variance: - Correct stats!!! (global p95, p99) - “Suppress the messengers” of microservice failures
- 31. Wrapping Up...
- 32. (first, a hint at my perspective)
- 33. A fun game! (“Observability Whack-a-Mole”) Design your own observability system: ❏ High-throughput ❏ High-cardinality ❏ Lengthy retention window ❏ Unsampled Choose three
- 34. The Life of Trace Data: Dapper Stage Overhead affects… Retained Instrumentation Executed App 100.00% Buffered within app process App 000.10% Flushed out of process App 000.10% Centralized regionally Regional network + storage 000.10% Centralized globally WAN + storage 000.01%
- 35. The Life of Trace Data: Dapper Other Approaches Stage Overhead affects… Retained Instrumentation Executed App 100.00% Buffered within app process App 100.00% Flushed out of process App 100.00% Centralized regionally Regional network + storage 100.00% Centralized globally WAN + storage “fancy”
- 36. An Observability Scorecard Detection - Specificity: cardinality cost, stack coverage - Fidelity: correct stats, high stats frequency - Freshness: ≤ 5 seconds Refinement - Identifying variance: cardinality cost, correct stats, hi-fi histograms, retention horizons - “Suppress the messengers”
- 37. LightStep: Observability with context Automatic deployment and regression detection System and service diagrams Real-time and historical root cause analysis Correlations Custom alerting Easy Setup with no vendor lock-in No cardinality limitations, really
- 39. Extra Slides
- 42. Ideal Refinement: Real-time Must be able to test and eliminate hypotheses quickly Actual data must be ≤10s fresh UI / API latency must be very low
- 44. Ideal Refinement: Context-Rich We can’t expect humans to know what’s normal