Cross-signals correlation — the place metrics, logs, and traces work collectively in live performance to supply a full view of your system’s well being — is usually cited because the “holy grail” of observability. Nevertheless, given the elemental variations of their knowledge fashions, these alerts normally stay in separate, remoted backends. Pivoting between sign varieties may be laborious, with no pure pointers or hyperlinks between your completely different observability programs.
Hint exemplars present cross-signals correlation between your metrics and your traces, permitting you to determine and zoom in on particular person customers who skilled irregular software efficiency. Storing hint info with metric knowledge helps you to rapidly determine the traces related to a sudden change in metric values; you do not have to manually cross-reference hint info and metric knowledge by utilizing timestamps to determine what had occurred within the software when the metric knowledge was recorded.
To make it even simpler to get began with this cross-signals story, we’re excited to announce that Managed Service for Prometheus now natively helps Prometheus exemplars!
Get a beginning-to-end view of excessive latency person journeys
As Google’s SRE ebook discusses in its part on monitoring distributed programs, it’s far more helpful to measure tail latency as a substitute of common latency. Latency is usually very unbalanced, because the SRE ebook explains:
“In the event you run an online service with a median latency of 100 ms at 1,000 requests per second, 1% of requests would possibly simply take 5 seconds. In case your customers rely on a number of such internet companies to render their web page, the 99th percentile [p99] of 1 backend can simply grow to be the median response of your frontend.”
By utilizing a histogram (a.okay.a., a distribution) of latencies as a substitute of a median latency metric, you’ll be able to see these high-latency occasions and take motion earlier than the p99.9 (99.ninth percentile) latency turns into the p99, p90, or worse.
Exemplars present the lacking hyperlink between noticing an latency challenge with metrics and performing root trigger evaluation with traces. While you add hint exemplars to your histograms, you’ll be able to pivot from a chart exhibiting a distribution of latencies into an instance hint that generated p99.9 latency. You’ll be able to then examine the hint to see what calls took probably the most time, permitting you to determine and resolve creeping latency points earlier than they have an effect on extra of your customers.