Why did errors rise after the deployment?
Use Anyshift’s Annie to investigate a PagerDuty alert with Kubernetes changes, Datadog signals and infrastructure dependencies.
Example
PagerDuty reports checkout-api errors after a rollout. The pods are ready, but requests fail. Ask Annie to compare the change with the application’s health signals.
- Resolve the affected service and read the alert’s time window.
- Compare rollout changes with available logs, metrics and dependencies.
- Return a likely cause with evidence, or explain what remains inconclusive.
A connection-pool change may explain database timeouts if the evidence supports it. Ready pods alone do not prove the application is healthy.
What you need
Provide the alert, service and time window. Reuse the project’s infrastructure and observability connections. PagerDuty is needed for its automatic trigger; Slack is needed only if you choose Slack delivery.
- PagerDuty: An incident.triggered webhook can start an investigation. Project automation rules decide which events Annie investigates.
- Kubernetes: The in-cluster integration supplies workloads and resource state. Helm-created resources are visible; chart versions and Helm revision history are not modeled as release entities.
- Datadog: Logs, metrics, monitors and traces require their corresponding application-key scopes. A successful connection check alone does not grant access to all signals.
- AWS changes: If the investigation needs AWS API activity, CloudTrail forwarding supplies that evidence. Reuse it when already configured.
Annie can only assess data available in the selected project and incident window. Missing permissions, unavailable telemetry or gaps in history must remain explicit. Nearby deployment and alert timestamps alone do not prove a root cause.
Try this prompt
In the intended Anyshift project, investigate [alert] for [service / environment] during [time window]. Confirm the project with list_projects. Use ask_annie with mode=rca, then poll the same investigation with get_annie_response until it completes or fails. Compare deployment changes with health signals. Return evidence, likely causes, missing data and the next check. Do not change infrastructure or send messages.