Services and latency
Every service your traces report, with its traffic, errors and latency percentiles.
The service list #
A service appears once your traces report it. The list shows the numbers you normally judge a service by, for the selected window.
- Requests per second — the entry spans in the window.
- Error rate — the share of requests that failed.
- Latency percentiles — 50th, 95th and 99th.
- Apdex — a single satisfaction score.
- Environment, language and team — where it runs.
- Last deploy and last seen — how current the service is.
What counts as a request #
A request is an entry span: a server or consumer span, or a span with no parent. Counting entry spans rather than every span is what makes the number mean requests rather than internal work.
What counts as an error #
A span is an error when its status says so, or when a server span returned a 5xx status. An outbound client call returning a 5xx is that dependency's problem, not your service's, so it is not counted here.
Percentiles #
Percentiles come from your data rather than from a sample, so a p99 that is driven by a handful of slow requests reflects them. Read the p95 and p99 together: a wide gap is a tail worth investigating.
Apdex #
Apdex scores a service from 0 to 1. A request is satisfying at or under 500 ms by default, tolerating up to four times that, and frustrating beyond it.
Apdex summarises, percentiles explain
Apdex is a good health signal to sort by. Percentiles are what tell you whether a low score came from a broad slowdown or a long tail.
Service detail #
Open a service for its endpoints, its versions, and the resources it depends on.
- Endpoints — the routes it serves, ranked by traffic.
- Versions — the deployed versions, with activity per version.
- Version comparison — put two versions side by side to see a change's effect.
- Dependencies — the services and datastores it calls.
Comparing versions #
Comparing two versions of a service over the same window is the direct way to answer whether a release helped. Look at error rate first, then latency.
A service whose traffic has stopped but whose last deploy is recent is usually a deployment problem rather than a code one. Check the version list.
Onward #
A number that needs explaining takes you to the requests behind it: Searching traces filters to this service and opens one. For who depends on whom, Service map. For what is failing, Error tracking.