Unlocking Faster Database Troubleshooting: Grafana Assistant AI-Powered Diagnostics

When your database slows down, pinpointing the root cause can feel like searching for a needle in a haystack. Grafana Cloud Database Observability already provides deep visibility into SQL queries with RED metrics, execution samples, wait event breakdowns, and visual explain plans. But raw data alone doesn't always tell you what to do next. Now, the new Grafana Assistant integration brings AI-powered analysis directly into your observability workflow, turning metrics into actionable insights without requiring you to copy-paste SQL or manually assemble context. Below, we answer common questions about how this integration transforms database troubleshooting.

What is the Grafana Assistant integration for Database Observability and how does it help?

The Grafana Assistant is an AI-driven feature embedded within Grafana Cloud Database Observability. It helps you move beyond simply seeing that a query's latency spiked, to understanding why it happened and what to do about it. Instead of generic chat prompts, the Assistant uses purpose-built analysis actions crafted by database engineers. It automatically runs queries against your real Prometheus and Loki data sources within the same time window you're investigating. This means it already knows your table schemas, indexes, and execution plans. The result is a specific, data-backed health assessment that highlights issues like wasted row scans, intermittent performance problems, or mysterious wait events. With the Assistant, you get guided diagnoses rather than raw numbers.

How does the Assistant differ from using a separate AI tool with copied SQL?

Traditional approaches involve copying SQL queries into standalone AI tools, which lack context about your database schema, indexes, or the exact time range of the problem. The Grafana Assistant eliminates these gaps. It works directly within Grafana Cloud, querying your actual Prometheus (for metrics) and Loki (for logs) data sources. It automatically loads the real table schemas, indexes, and execution plans for the query you're investigating. This ensures the analysis is based on live data, not a static snippet. Additionally, your query text and schema metadata are used only for the current analysis and are never stored or used for model training, addressing privacy concerns. The Assistant's responses are specific, actionable, and grounded in your environment—not generic AI guesses.

What are the guided AI buttons and what common issues do they address?

The integration includes pre-built AI buttons that offer a guided experience for tackling common database problems. These buttons appear on tabs like the query overview and provide one-click access to purpose-built prompts. The main prompts focus on diagnosing slow or degraded queries, as well as getting recommendations for structural changes. For example, a button labeled Why is this query slow? triggers an analysis that examines duration spikes, row examination ratios, and wait event contributions. Another button might suggest index improvements or schema adjustments. These are not generic prompts but carefully designed analyses that reflect how expert database engineers approach troubleshooting. You can still use the free‐form chat box, but the buttons streamline the path from observation to solution.

Can you walk through an example of diagnosing a slow query using the Assistant?

Imagine you identify a query in the overview where P99 latency is spiking and error rates are climbing. Instead of manually sifting through metrics, you click the Why is this query slow? button. The Assistant immediately queries Prometheus and Loki for your selected time window and synthesizes a health assessment. It reports that 50 times more rows are examined than returned, meaning most work is wasted on filtering. It notes that P99 latency is 12 times the median, suggesting an intermittent problem rather than a constant bottleneck. CPU time looks healthy, but wait events consume 40% of execution time. These findings are presented in plain language, so you know exactly which area to investigate—filter efficiency, concurrency patterns, or wait events.

How does the Assistant handle obscure wait event names like innodb mutex?

Database wait events often have cryptic names such as wait/synch/mutex/innodb or io/table/sql/handler. To a human, these don't immediately convey the problem. The Assistant leverages its engineered knowledge to interpret these events in context. It reads the actual wait event data from your database's diagnostic tables and correlates them with your query's execution plan and time-series metrics. For instance, when encountering wait/synch/mutex/innodb, the Assistant explains that this indicates internal InnoDB mutex contention, often caused by high concurrency on hot rows or inefficient locking. It then provides specific advice, such as checking for long-running transactions or considering row-level lock optimization. The Assistant translates technical jargon into actionable guidance.

What data sources does the Assistant query and how is privacy handled?

The Grafana Assistant queries your existing Prometheus and Loki data sources—the same ones you already use for monitoring. It pulls metrics (like RED metrics) and logs (including execution samples and wait events) from the exact time window you are viewing. This means the analysis is always based on current, real-world data. Regarding privacy: the Assistant uses your query text and schema metadata only for the current analysis session. This data is not stored persistently, nor is it used for model training. The integration operates entirely within your Grafana Cloud environment, ensuring sensitive information remains under your control. All AI processing respects the same data residency and security policies as the rest of Grafana Cloud.

How does the Assistant provide a health assessment by combining Prometheus and Loki data?

The Assistant doesn't treat metrics and logs as separate silos. When you click a guided prompt, it executes queries against Prometheus for time-series data (e.g., duration spikes, CPU usage, row examination rates) and against Loki for detailed event logs (e.g., wait event breakdowns, schema metadata). It then synthesizes this information into a unified health assessment. For example, it might correlate a high P99 latency from Prometheus with a specific wait event pattern from Loki. The output is a concise report that tells you not only what is happening but why, focusing on the most impactful factors. This cross-data-source analysis is what makes the Assistant's diagnoses more comprehensive than manual inspection of each tool separately.