Thursday, 4 p.m.: the head of operations asks whether the archive database will still have room at the end of the quarter. You have a free-space curve, but no answer. A Davis forecast in a dashboard exists for exactly this moment: it turns a past curve into a numeric estimate, with a margin of error, before you order storage or decide to do nothing.
This article starts from the decision, not from the feature. When does a forecast really help? How do you attach it to a tile? And how do you avoid trusting a line that has no value? The advice applies to Dynatrace SaaS with Grail and the Dashboards app. I have not checked Dynatrace Managed, so I promise nothing there.
A word on method: the facts come from the Dashboards documentation and the Davis for Workflows page, read on 10 November 2026. span.observer has not tested these settings on a tenant, so anything missing from those pages is marked « unverified ».
When a Davis forecast changes a decision
A forecast is only worth something if it leads to a dated action. Three situations justify one:
- A resource running out: disk, quota, connection count. The question is « in how many days? ».
- A load that keeps climbing: traffic, message queues. The question is « do we add capacity before the next campaign? ».
- A budget drifting: ingested log volume, monthly cost. The question is « where will we be at month end? ».
Conversely, an erratic curve with no trend or rhythm has nothing to predict. In that case, a threshold or anomaly detection is better, and the documentation offers both in the same tile section: auto-adaptive threshold, seasonal baseline or static threshold.
A post on the official Dynatrace blog, Better dashboarding with Dynatrace Davis AI, presents the idea from the product side. Here I stick to the walkthrough and the operational pitfalls.
Step 1: write a query worth predicting
The forecast is computed on the tile result. A poor query therefore gives a poor forecast. Start with a regular series over a window long enough to contain one full cycle of your activity.
Here is a free disk space tile per host and disk, over seven days, with one point per hour:
Two choices matter. Grouping gives one series per disk, so you can spot the failing disk instead of averaging the whole fleet. And the 50-series limit keeps the chart readable, because the documentation says the analyzer accepts up to 1000 series, set through the Limit option.
Step 2: turn on the Forecast analyzer for the tile
In the Dashboards app, the documented procedure is short:
- Run the tile query.
- Expand the Davis AI section of the options panel.
- In the Analyzers list, pick Forecast.
- Switch the toggle at the top from « AI data analysis is not active » to « AI data analysis is active »: analyzers are off by default.
- Choose the Davis AI analysis visualization.
Two settings drive the forecast. Data points to predict sets how many future points you get; the page warns that more points mean a less reliable forecast and a longer run. Forecast offset ignores the last N points and forecasts them instead, which is handy for testing the model against what really happened.
On permissions, you need davis:analyzers:read to see the configuration and davis:analyzers:execute to run the analyzer. If a colleague sees the tile without the forecast, check their IAM policies first.
Step 3: read the forecast with suspicion
The Davis AI analysis visualization can show the chart, the table or both (the Visible sections option). Keep both while you form an opinion.
The Grail and Davis AI capacity post says the model is chosen automatically from the characteristics of the series, and that the result is probabilistic, with a predicted value and bounds. So remember that a forecast is a range, not a single date. Decide on the unfavourable bound.
This is the test I apply before telling anyone about a forecast:
- Backtest: set Forecast offset to predict points you already know, then compare.
- Sensible horizon: do not go beyond a third of the history window. That is a personal rule of thumb, not a vendor recommendation.
- Past events: a recent migration or clean-up breaks the trend; exclude the period or note it on the tile.
Step 4: get out of the tile with a workflow
A dashboard wakes nobody up. If the decision must be automatic, the Davis for Workflows documentation describes the Generic Forecast Analysis analyzer, used in an « Analyze data » action. Its example predicts disk capacity from avg(dt.host.disk.free) over seven days in 120 bins, with 100 points to predict.
The key point is the safeguard: the example only uses a series if analysisStatus is OK and forecastQualityAssessment is VALID. Adopt that habit. Without it, you will alert on forecasts the engine itself considers doubtful. The page also warns that the output volume can be very large, so filter before acting.
One useful caveat: that page does not mention dashboards and routes capacity setup through Notebooks. I therefore cannot say that a workflow can be triggered from a tile: unverified.
Best practices
- One tile, one decision: write the expected action in the tile title (« order storage before… »).
- Limit the series; a forecast over 1000 curves cannot be read.
- Keep anomaly detection for series without a trend, and forecasting for those that have one.
- Review the forecast weekly at first: a drifting model shows up quickly.
- Give the tile an owner and a runbook, as you would for an SLO dashboard for management.
Frequently asked questions
Does a Davis forecast work on every query?
It applies to a numeric time series. A tile that returns a list of rows is a poor candidate: turn it into a series first with timeseries or makeTimeseries.
Do I need to write DQL to get a forecast?
You need a query that produces the series, but the analysis itself is set through the interface. If DQL is new to you, see Davis CoPilot and query generation, and always review the generated query.
How many points can I predict?
The Dashboards page gives no ceiling; it only says that more points make the forecast less reliable. No numeric limit is verified here.
Conclusion
Start from a capacity question, write a regular query, turn on the Forecast analyzer, then test it on the past before showing it to anyone. To learn how to build the tiles themselves, go through the SaaS dashboard examples in DQL.
span.observer is an independent community blog, not affiliated with Dynatrace LLC. Trademarks belong to their owners. Tools change quickly: check the options in the official documentation.
Originally published on span.observer. French version: https://span.observer/fr/previsions-davis-tableaux-de-bord-dynatrace/.



Top comments (0)