Course: Dynatrace SaaS · Module 5/8 — Dynatrace Davis problems and root cause · Duration: ~60 min · Level: intermediate
Objectives:
- Explain how Davis turns isolated anomalies into a single problem with a root cause.
- Read a problem in the Problems app and find the same data in Grail with DQL.
- Create a custom anomaly detection on a time series.
- Choose the right notification channel for a team.
Prerequisites: modules 1 to 4 of the course (platform, OneAgent, DQL, Notebooks and Dashboards) and a trial environment with at least one monitored application.
This training module covers Dynatrace Davis problems in Dynatrace SaaS, in other words how the platform spots an anomaly, links it to the other symptoms, points to a probable root cause and alerts the right people. So far, you have learnt to collect data and query it. Now it is about letting the platform work for you overnight, then knowing how to reread its reasoning in the morning.
Davis is Dynatrace’s causal AI engine; in the vendor’s recent messaging, its capabilities are grouped under the name Dynatrace Intelligence. The principle does not change: it starts from the topology (Smartscape) to explain anomalies, instead of correlating alerts by timestamp alone.
How does Davis go from an anomaly to a problem?
The cycle has four stages, which are worth keeping in mind when you read a problem:
| Stage | What happens | What you see |
|---|---|---|
| 1. Baseline | Davis learns the normal behaviour of each entity (response time, error rate, traffic, resources). | Nothing: this is the silent learning of the first few days. |
| 2. Event | A significant deviation or a crossed threshold produces a Davis event attached to an entity. | An event in the entity view and in dt.davis.events. |
| 3. Correlation | Events linked by the topology (a service, its database, its host, its pod) are grouped together. | A single problem instead of ten alerts. |
| 4. Root cause | Davis walks up the dependencies to identify the entity most likely to be the origin. | A root cause section and the list of affected entities. |
Remember the difference in vocabulary: an event is an observed fact about an entity; a problem is a group of events that Davis considers related, with a lifecycle (open, then closed) and an impact.
Baselines and thresholds
For services and applications, Davis relies first on automatic baselines: you have nothing to configure to be alerted to a sudden rise in error rate. For infrastructure, default thresholds exist (CPU saturation, disk almost full, and so on). You can adjust the sensitivity of these detections in the anomaly detection settings, per environment or for a narrower scope.
Dynatrace Davis problems: reading a problem in the Problems app
The Problems app is the entry point. For each problem, use the same four-question reading:
- What? The title and category summarise the main symptom.
- Who is affected? The affected entities and, where relevant, the number of users.
- Why? The proposed root cause, with the events that support it (deployment, configuration change, saturation).
- Since when? The timeline: the first event is often more telling than the most recent one.
From the problem, you can jump to the distributed traces, the logs for the period and the entity’s metrics.
Querying problems with DQL
Davis problems and events are stored in Grail. So you can analyse them like any other data, for example for a weekly review:
fetch dt.davis.problems, from:now()-7d
| summarize total = count(), by:{event.category}
| sort total desc
To list problems that are still open:
fetch dt.davis.problems, from:now()-24h
| filter event.status == "ACTIVE"
| fields timestamp, display_id, event.name, event.category
| sort timestamp desc
Field names may change: before building a dashboard, run a simple fetch dt.davis.problems | limit 5 and look at the fields actually present in your environment. The same applies to dt.davis.events.
Creating your own anomaly detections
Built-in detections do not cover your business indicators. In the Anomaly Detection app, you can define a custom detection from a DQL query that returns a time series, then choose a method:
- static threshold: simple and predictable, suited to a contractual limit;
- auto-adaptive threshold: the platform learns the recent normal range;
- seasonal baseline: takes cycles (day, week) into account, useful for highly variable traffic.
An example of a time series you can use as a source, the number of failed requests per minute across services:
timeseries failures = sum(dt.service.request.failure_count), interval:1m
Check the metric name in your environment (the query editor offers autocompletion); you can also start from a metric extracted from logs, as you will see in module 6. A custom detection produces Davis events, which can then open a problem.
Notifying the right people
In Dynatrace SaaS, two approaches coexist. Classic notification integrations, configured in the settings, send problems to an email address, a webhook or an incident management tool. The modern approach goes through Workflows: a trigger on a Davis problem, filters (category, entities, tags), then one or more actions. This second mode is more flexible and is covered in detail in module 7.
Whatever the channel, three rules prevent alert fatigue: filter by scope (a team only receives its own services), wait a few minutes before notifying a low-severity problem, and include the link to the problem in every message. Teams coming from Managed will recognise the spirit of the alerting profiles described in alerting profiles and notification integrations in Managed, with different tools.
Hands-on exercise
Step 1 — Observe a problem
In your trial environment, open the Problems app. If no problem is open, generate a load of errors on a demo application (for example by repeatedly calling a non-existent URL) for about ten minutes. Expected result: an error-category problem appears, with the service concerned as the affected entity.
Step 2 — Answer the four questions
Open the problem and note in a Notebook: the symptom, the affected entities, the proposed root cause and the first event. Expected result: a five-line analysis sheet.
Step 3 — Find the problem in Grail
In the same Notebook, run the query on dt.davis.problems filtered on the last 24 hours. Expected result: a row whose identifier matches the one shown in the app.
Step 4 — Create a custom detection
In Anomaly Detection, create a static-threshold detection on a time series of your choice, with a deliberately low threshold. Expected result: a Davis event appears after a few minutes; then delete the detection so as not to clutter the environment.
Quiz
- What is the difference between an event and a problem?
A. None, they are synonyms · B. A problem groups related events, with a root cause · C. An event is always more serious · D. A problem only concerns infrastructure - What does Davis mainly rely on to correlate anomalies?
A. The alphabetical order of entities · B. The Smartscape topology · C. The number of logged-in users · D. Tags only - Which method best suits highly seasonal traffic?
A. Static threshold · B. Seasonal baseline · C. No detection · D. A manual alert - Where are problems stored so they can be analysed in DQL?
A. In a local file · B. In Grail · C. In the Cluster Management Console · D. They are not stored
Answers
- B — the problem aggregates events that Davis considers related and assigns them a root cause.
- B — dependencies between entities make it possible to explain rather than simply correlate in time.
- B — it learns daily or weekly cycles and avoids alerts during off-peak hours.
- B — Davis problems and events are Grail data, queryable with
fetch.
Further reading
- Dynatrace SaaS course overview
- Log ingestion: detecting silence, coverage gaps and excess volume
- The Dynatrace plugin for Cursor: investigating production without leaving the editor
- Official Dynatrace documentation (Davis, Problems and Anomaly Detection sections)
← Module 4: Notebooks and Dashboards · Course overview · Module 6: Logs and OpenPipeline →
Frequently asked questions
Does Davis need to be configured to work?
No. Baselines and correlation are active as soon as data arrives. Configuration is for fine-tuning: sensitivity, custom detections, notifications.
Is the proposed root cause always right?
It is the most likely hypothesis given the topology and the available events. It is reliable when coverage is good; an uninstrumented entity, on the other hand, can escape the analysis.
Can you close a problem manually?
A problem closes automatically when the events that make it up end. It is better to fix the cause or adjust the detection than to try to hide it.
With Davis, an alert becomes an explanation: a problem, a probable cause, an impact. In the next module, you will learn to give Davis more material to work with by structuring your logs with OpenPipeline.
span.observer is an independent community blog, not affiliated with Dynatrace LLC. This course does not replace the official Dynatrace University training and certifications. Check features against the official documentation for your version.
2 comments