Problem
The overview shows error rate climbing. The viewer now needs the specific service, the specific endpoint, and eventually the specific request, and at every step they must not lose the time range, the filters, or their place.
Solution
Make every aggregate a doorway. A bar, a row, a cell, a node: clicking it opens the narrower view of the same thing, already scoped.
Shneiderman’s mantra names this as one of the seven tasks and puts it in sequence: overview first, zoom and filter, then details on demand. Drill-down is the transition between those stages, and it is judged almost entirely on how much context survives the jump.
Grafana’s data links show what “carrying context” concretely requires. A link can interpolate __url_time_range for the dashboard’s current window, __from and __to, __series.name, __field.name and __field.labels.<LABEL>, and the value under the cursor as __value.raw, __value.numeric, __value.text or __value.time. That list is essentially an inventory of what the viewer had in their head at the moment they clicked. A drill-down that drops any of it makes them rebuild it by hand on arrival.
The most common failure is silent scope loss: clicking a spike on a chart filtered to one cluster lands on a page showing all clusters, at the default time range, and the number the viewer was chasing is no longer there. They usually assume they misread the first chart.
The second failure is the dead end. Somewhere down the chain there is a level with nothing below it, and if that level doesn’t say so, viewers keep clicking and start distrusting the whole path.
Use when
There is a genuine hierarchy—fleet to host, service to endpoint, issue to event—and the overview is a summary of things that individually exist somewhere.
Don’t use when
There is no level below. A click that reloads the same information in a modal teaches viewers that clicking does nothing. Also avoid it when the answer is better reached by filtering in place; jumping pages costs orientation that cross-filtering doesn’t.
Trade-offs
Every drill step is a context switch, and deep paths lose people. Back is the most-used control in this pattern and the most often broken, especially when the drill target is a different product. Links encode assumptions about the destination’s URL structure, so they rot quietly when the other page changes. And drill-downs create an implicit hierarchy that may not be the only sensible one: a path from service to host cannot answer a question that starts at the host.
Checklist
- Does the time range survive the jump?
- Do the active filters and template variables survive it?
- Does the value or series the viewer clicked arrive as scope on the other side?
- Is the target obviously the same thing, narrower—or does it look like a different page?
- Is it visually clear what is clickable before the cursor is over it?
- Does browser back return to the exact prior state, including scroll?
- What happens at the bottom of the hierarchy, and does the last level say it is the last?
- Does the destination handle an empty result, and does it say why?
- Does it open in place or a new tab, and is that consistent across the dashboard?
- Who owns the destination’s URL contract, and what happens when it changes?
Compare
Grafana makes the context explicit as interpolated variables, which is powerful and means a drill-down is only as good as the person who assembled the link string. Sentry builds the hierarchy into its object model—issue to event to trace to span—so drilling is navigation between things rather than between dashboards, and back always means something. Datadog ties panels to a scope inherited from the page, so a click carries tags rather than a hand-written URL, which rots less and constrains where you can go. Honeycomb largely replaces the pattern: rather than navigating to a narrower page you add a group-by or a filter to the query you already have, so there is no jump to lose context across.
Related
Overview then detail is the structure this pattern moves through. Cross-filter is the alternative that narrows without navigating. Detail on demand is the in-place version for a single item. Trace waterfall is often the last stop on the path. Ranked list is the most common origin, because a rank position is a promise that the item exists.