DEVOURED
Beyond Line Charts: Why Some Diversity in Telemetry Visualization Is Long Overdue (42 minute video)
Observability experts argue that relying on line charts often distorts reality and fails to answer critical operational questions.
What: Yao Yue, a former Twitter performance engineer, advocates for adopting scatter plots, CDFs, and bucketed data to better analyze capacity and latency distributions.
Why it matters: Line charts hide high-frequency noise and encourage misleading extrapolation, often masking true system behaviors when systems are stressed.
Takeaway: Stop defaulting to time-series line charts for everything; try using raw dot plots or CDFs when analyzing latency distributions or fleet-sizing questions.
Deep dive
- Line charts visually interpolate data that does not exist, creating misleading 'extrapolated pixels'.
- Counters are best represented by rates (derivative), not raw cumulative lines.
- Gauges (spot measurements) should avoid line connections entirely to prevent false interpolation.
- Histograms (latency) are essential because P99.9 can be orders of magnitude higher than averages.
- Operational questions (e.g., 'is my fleet right-sized?') are often timeless and should be solved by grouping data by parameters rather than time.
- CDF plots are superior for visualizing regressions between different software versions.
Decoder
- CDF: Cumulative Distribution Function, a plot showing the probability that a value will be less than or equal to a given point, useful for comparing latency across distributions.
- SLO: Service Level Objective, a measurable target for system performance, such as '99% of requests return under 100ms'.
- Quantile: A point in a distribution that divides the data into intervals (e.g., P99 is the 99th percentile).
Original article
Full article content is not available for inline reading.
Explore all 59 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi