apiVersion: nines.dev/v1alpha1
kind: ServiceLevelObjective
metadata:
    name: inference-ttft
    project: production
spec:
    lifecycle: experimental
    measurement:
        type: latency
        metric: inference_time_to_first_token_seconds
        labels:
            service: inference
        thresholdSeconds: 0.5
        scrapeIntervalSeconds: 15
        observationPoint: Streams reaching a first token only. Failed or never-started streams are excluded; pair with inference availability. This is a conditional latency SLO.. Requires a classic histogram in seconds with an exact matching le bucket and count series; summaries and percentile gauges are unsupported.
    source: primary
    displayName: Time to first token
    service: inference
    target: 0.99
    windowSeconds: 2419200
    stepSeconds: 60