When your database slows down, identifying the root cause can feel like searching for a needle in a haystack. Grafana Cloud's Database Observability already surfaces key metrics like RED (Rate, Errors, Duration), execution samples, wait event breakdowns, table schemas, and visual explain plans. But visibility alone isn't enough—you need actionable guidance. The new Grafana Assistant integration bridges that gap by combining AI with live observability data. Instead of juggling separate tools, you get contextualized, real-time analysis directly within the investigation workflow. Below, we answer common questions about how this assistant transforms database troubleshooting.
What is Grafana Assistant for Database Observability and how does it work?
The Grafana Assistant is an AI-powered feature integrated into Grafana Cloud's Database Observability. It helps you diagnose performance issues by running queries against your actual Prometheus and Loki data sources—within the exact time window you're investigating. Unlike generic AI chatbots that rely on pasted SQL, this assistant automatically loads your real table schemas, indexes, and execution plans. It synthesizes data from multiple sources to generate a health assessment. For example, if a query's P99 latency spikes, the assistant examines metrics like rows examined vs. returned, CPU time, and wait events to pinpoint why. Each analysis is purpose-built by database engineers, offering specific advice rather than generic responses. You can trigger pre-built prompts or type free-form questions, but every answer is grounded in your database's live state.
How does the assistant help when you find a slow query?
Suppose you spot a query in the overview with spiking duration and climbing error rates. Clicking into it shows detailed time-series performance data, but the root cause isn't obvious—bad join, lock contention, or a table scan that only now causes issues. With the assistant, you simply click a pre-built prompt button labeled "Why is this query slow?" The assistant immediately queries Prometheus and Loki for the selected time window. It synthesizes the data: perhaps duration spikes because rows examined are 50 times rows returned (most work wasted on filtering). It might note that P99 is 12x the median (intermittent issue) and CPU is healthy, but wait events consume 40% of execution time. This step-by-step breakdown saves you from manually correlating metrics, directly pointing to the bottleneck.
What makes this assistant different from pasting SQL into a generic AI tool?
The key difference is context. When you paste SQL into an external AI tool, you lose critical metadata—time ranges, table schemas, indexes, execution plans, and live performance data. The Grafana Assistant already has all this loaded. It doesn't work from a static copy; it queries your actual Prometheus (metrics) and Loki (logs) sources, ensuring recommendations are based on real database states. There's no need to manually explain schema or time periods. The assistant also uses purpose-built analysis actions designed by database engineers, not generic prompts. Each action targets specific issues like slow queries or degradation, providing advice rooted in your data. This contextual depth makes diagnosis faster and more accurate.
How does the assistant handle cryptic wait events like wait/synch/mutex/innodb?
Database wait events often have opaque names such as wait/synch/mutex/innodb or io/table/sql/handler. Even experienced engineers may need to research these. The Grafana Assistant automatically interprets these events using its AI model. For example, if a wait event consumes 40% of execution time, the assistant explains what it means—specifically what resource the database is contending for—and connects it to your query's performance. It might say, "During this wait, the database is physically contending for a mutex lock in InnoDB. This often indicates contention on a hot row or index page." By translating technical names into plain English, the assistant reduces the need for external lookups and speeds up root cause analysis.
What are the pre-built analysis actions and how do they help?
Instead of forcing users to write their own prompts from scratch, the assistant includes out-of-the-box AI buttons for common scenarios. These purpose-built actions are designed by database engineers to tackle issues like slow queries, degraded performance, or recommendations for changes. For instance, a "Why is this query slow?" button automatically runs a multi-source analysis comparing metrics from Prometheus (CPU, wait events) with logs from Loki (execution samples). Another button may offer schema optimization suggestions. These guided analyses remove guesswork, providing consistent, actionable insights. You can still freely prompt in the chat box, but the pre-built actions ensure you quickly get expert-level diagnosis even if you're unfamiliar with query tuning.
How does the assistant handle data privacy and security?
Security is a priority. The assistant uses your query text and schema metadata only for the current analysis session. This data is not stored or used for model training. Every analysis runs against your own Prometheus and Loki sources within the time window you specify—no data is sent to external servers for training purposes. The integration respects your existing data governance policies. Additionally, the assistant's responses are generated server-side within Grafana Cloud, meaning your database details remain under your control. This approach ensures that while you benefit from AI-driven insights, sensitive information about your schema or queries never leaves your environment unnecessarily.
Can users still ask free-form questions in addition to using pre-built prompts?
Absolutely. While the pre-built analysis actions offer a guided experience, you are free to type any question directly into the assistant's chat box, just like you would with a generic AI chatbot. The difference is that your questions will be answered using the live context of your database—complete with current time range, schemas, indexes, and execution plans. For example, you could ask "What indexes should I add for this query?" and the assistant will analyze your explain plan and suggest missing indexes. Or you might ask "Why is the error rate climbing now but not yesterday?" and it will compare metrics across time windows. This flexibility makes the assistant suitable for both novices following guided paths and experts wanting deep custom analysis.