br_v0_2_example.yaml

# Schema version for this entire benchmark report
version: '0.2'

run: # These are details on the benchmark run that produced these results
  # Unique ID, only this benchmark report has this ID
  uid: 38b1f169ca178b756f7483523b17de61
  # Experiment ID, common to benchmark reports captured during this experiment
  eid: d9c5a5ddce6c2f885a9f8e6a6a6db0fb
  # Cluster ID (could be hash of cluster URL from context)
  cid: 09825952f60004aa59bb5b2a2eefa6d1
  # Pod ID, unique to a workload generating and/or data collecting pod
  pid: 052d5f654ea08edf5caa2c0293fa7fb3
  time:
    start: 2025-11-05T18:00:42Z
    end: 2025-11-05T18:17:03Z
    duration: PT49.97206788184121S
  user: namasluk
  description: "Baseline latency test for Qwen3-0.6B on inference-scheduling stack"
  keywords:
    - baseline
    - latency
    - qwen3

# System inputs, combination of component configurations and workload
scenario:
  stack:
  - metadata: # Details about this component
      schema_version: 0.0.1 # This is the version of this component's schema
      label: vllm-svc-0 # This is a unique name for this particular component
      cfg_id: cc73fc6b51a1d3b8128f312d70476d7c # This is a hash of this component's configuration
      description: "Optional description of this component"
    standardized: # These are common properties that apply to all instances of this kind of component
      kind: inference_engine # This describes what this component is
      tool: llm-d
      tool_version: ghcr.io/llm-d/llm-d-cuda:0.3.1
      role: decode # One of prefill, decode, aggregate
      replicas: 2
      model:
        name: Qwen/Qwen3-0.6B
      accelerator:
        model: NVIDIA-H100-80GB-HBM3
        count: 8
        parallelism:
          dp: 1
          dp_local: 1
          workers: 1
          ep: 1
          pp: 1
          tp: 8
    native: # These are the exact and specific arguments/variables/configuration details for this component
      args: # Arguments sent to executable
        --block-size: 1024
        --kv-transfer-config: {"kv_connector":"NixlConnector", "kv_role":"kv_both"}
        --no-enable-log-requests: null
        --disable-uvicorn-access-log: null
        --max-model-len: 1024
        --tensor-parallel-size: 8
      envars: # Relevant environment variables
        UCX_TLS: "rc,sm,cuda_ipc,cuda_copy,tcp"
        UCX_SOCKADDR_TLS_PRIORITY: "tcp"
        UCX_NET_DEVICES: mlx5_1:1
        NCCL_IB_HCA: mlx5_1
        VLLM_NIXL_SIDE_CHANNEL_PORT: "5557"
        VLLM_NIXL_SIDE_CHANNEL_HOST: 10.39.39.3
        VLLM_LOGGING_LEVEL: DEBUG
        VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"

  - metadata:
      schema_version: 0.0.1
      label: epp-0
      cfg_id: 47000e70c655e88198e0dc4e57d41d5f
      description: "This is a router, but no standardized component for this exists"
    # This component uses the ComponentStandardizedBase, where the standardized
    # section contains only "tool" and "tool_version", but otherwise this kind
    # has no other features.
    standardized:
      # kind, tool, and tool_version are the only required fields
      kind: generic
      tool: llm-d-inference-scheduler
      tool_version: ghcr.io/llm-d/llm-d-inference-scheduler:0.3.2
      # Other fields are not validated, and can be anything
      kind_draft: router
      some_config_param: 93
      another_thing:
        - a: 5
        - a: 1
        - b: 3
    native:
      config: # Configuration used (in this case a manifest of a k8s custom resource)
        apiVersion: inference.networking.x-k8s.io/v1alpha1
        kind: EndpointPickerConfig
        plugins:
        - type: single-profile-handler
        - type: decode-filter
        - parameters:
            blockSize: 16
            indexerConfig:
              kvBlockIndexConfig:
                enableMetrics: true
                metricsLoggingInterval: 60000000000
              tokenProcessorConfig:
                blockSize: 64
                hashSeed: '42'
            lruCapacityPerServer: 31250
            maxPrefixBlocksToMatch: 256
            mode: cache_tracking
          type: prefix-cache-scorer
        - type: kv-cache-scorer
        - type: queue-scorer
        - type: max-score-picker
        schedulingProfiles:
        - name: default
          plugins:
          - pluginRef: decode-filter
          - pluginRef: prefix-cache-scorer
            weight: 2.0
          - pluginRef: kv-cache-scorer
            weight: 1.0
          - pluginRef: queue-scorer
            weight: 1.0
          - pluginRef: max-score-picker

  load:
    metadata:
      schema_version: 0.0.1
      cfg_id: a4e18f265cc33786a42b8a3f7ac2edcb
      description: "Optional description of workload"
    standardized:
      tool: inference-perf
      tool_version: 0.3.0
      # For workload generators that have stages, we must specify which stage this report's data is from
      stage: 3
      input_seq_len:
        distribution: fixed
        value: 2148
      output_seq_len:
        distribution: gaussian
        value: 1000
        std_dev: 100
      prefix:
        prefix_len:
          distribution: fixed
          value: 1000
        num_groups: 32
        num_users_per_group: 10
        num_prefixes: 8
      multi_turn:
        enabled: true
        max_turns:
          distribution: fixed
          value: 20
      rate_qps: 10
      concurrency: .inf
      source: sampled
    native:
      config:
        api:
          headers: null
          streaming: true
          type: completion
        data:
          input_distribution: null
          output_distribution: null
          path: null
          shared_prefix:
            num_groups: 32
            num_prompts_per_group: 8
            output_len: 1000
            question_len: 100
            system_prompt_len: 2048
          type: shared_prefix
        load:
          interval: 1.0
          num_workers: 112
          stages:
          - duration: 50
            rate: 2.0
          - duration: 50
            rate: 5.0
          - duration: 50
            rate: 8.0
          - duration: 50
            rate: 10.0
          - duration: 50
            rate: 12.0
          - duration: 50
            rate: 15.0
          - duration: 50
            rate: 20.0
          sweep: null
          type: constant
          worker_max_concurrency: 100
          worker_max_tcp_connections: 2500
        metrics: null
        report:
          prometheus:
            per_stage: false
            summary: true
          request_lifecycle:
            per_request: true
            per_stage: true
            summary: true
        server:
          api_key: null
          base_url: http://infra-nam-release-inference-gateway.namasluk.svc.cluster.local:80/qwen-qwen3-0-6b
          ignore_eos: true
          model_name: Qwen/Qwen3-0.6B
          type: vllm
        storage:
          google_cloud_storage: null
          local_storage:
            path: /requests/inference-perf_1759339259-cache_tracking-run_100_1000_llm-d-0p6b-base
            report_file_prefix: null
          simple_storage_service: null
        tokenizer:
          pretrained_model_name_or_path: Qwen/Qwen3-0.6B
          token: null
          trust_remote_code: null

results:
  request_performance: # Stack performance as seen by users
    aggregate: # Statistical summary of results (required)
      requests:
        total: 500
        failures: 0
        input_length:
          units: count
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
        output_length:
          units: count
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
      throughput:
        input_token_rate:
          units: tokens/s
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
        output_token_rate:
          units: tokens/s
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
        request_rate:
          units: queries/s
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
        total_token_rate:
          units: tokens/s
          mean: 2262.448
          min: 2223.0
          p0p1: 2223.998
          p1: 2227.99
          p10: 2240.0
          p25: 2252.0
          p5: 2234.95
          p50: 2262.0
          p75: 2270.0
          p90: 2286.0
          p95: 2294.0
          p99: 2322.0
          p99p9: 2326.503
          max: 2328.0
      latency:
        time_to_first_token:
          units: s
          mean: 0.0368578234128654
          min: 0.027276142966002226
          p0p1: 0.027322400792501866
          p1: 0.02785794825293124
          p10: 0.031235878029838203
          p25: 0.033357357955537736
          p5: 0.030159439309500158
          p50: 0.03622930939309299
          p75: 0.03993474820163101
          p90: 0.04262634576298297
          p95: 0.045246614259667695
          p99: 0.05029489721637218
          p99p9: 0.057014757215045425
          max: 0.0574655020609498
        inter_token_latency:
          units: s/token
          mean: 0.0368578234128654
        time_per_output_token:
          units: s/token
          mean: 0.0368578234128654
        normalized_time_per_output_token:
          units: s/token
          mean: 0.0368578234128654
        request_latency:
          units: s
          mean: 0.0368578234128654

    # NOTE: Time series data is likely to be replaced with links to the actual data
    time_series: # Optional time series data
      throughput:
        input_token_rate:
          units: tokens/s
          series:
            - ts: 2025-11-05T18:01:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
            - ts: 2025-11-05T18:02:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
        output_token_rate:
          units: tokens/s
          series:
            - ts: 2025-11-05T18:01:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
            - ts: 2025-11-05T18:02:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
        total_token_rate:
          units: tokens/s
          series:
            - ts: 2025-11-05T18:01:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
            - ts: 2025-11-05T18:02:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
        request_rate:
          units: queries/s
          series:
            - ts: 2025-11-05T18:01:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
            - ts: 2025-11-05T18:02:00Z
              value: 204.0394
              p90: 0.04262634576298297
              p95: 0.045246614259667695
      latency:
        time_to_first_token:
          units: s
          series:
          - ts: 2025-11-05T18:01:00Z
            mean: 0.0368578234128654
            p90: 0.04262634576298297
            p95: 0.045246614259667695
          - ts: 2025-11-05T18:02:00Z
            mean: 0.0368578234128654
            p90: 0.04262634576298297
            p95: 0.045246614259667695

  observability:
    # Per-component observability metrics
    # Each entry references a component_label from scenario.stack[].metadata.label
    components:
      - component_label: vllm-svc-0  # References scenario.stack[0].metadata.label
        aggregate:
          kv_cache_usage:
            units: percent
            mean: 42.5
            min: 12.3
            p50: 41.8
            p75: 52.1
            p90: 63.4
            p95: 71.2
            p99: 85.6
            max: 92.1
          gpu_cache_usage:
            units: percent
            mean: 38.2
            min: 10.1
            p50: 37.5
            p90: 58.7
            p99: 78.3
            max: 84.0
          gpu_memory_usage:
            units: GiB
            mean: 62.4
            min: 58.1
            p50: 62.0
            p90: 66.8
            p99: 71.2
            max: 73.5
          gpu_utilization:
            units: percent
            mean: 76.0
            min: 45.2
            p50: 78.9
            p90: 89.3
            p99: 94.1
            max: 96.8
          running_requests:
            units: count
            mean: 24.6
            min: 1.0
            p50: 23.0
            p90: 42.0
            p99: 56.0
            max: 62.0
          waiting_requests:
            units: count
            mean: 3.2
            min: 0.0
            p50: 2.0
            p90: 8.0
            p99: 18.0
            max: 25.0
        time_series:
          kv_cache_usage:
            units: percent
            series:
              - ts: 2025-11-05T18:05:00Z
                mean: 35.2
                p90: 48.1
                p99: 62.5
              - ts: 2025-11-05T18:10:00Z
                mean: 52.8
                p90: 71.3
                p99: 85.6
          gpu_utilization:
            units: percent
            series:
              - ts: 2025-11-05T18:05:00Z
                mean: 73.2
                p90: 85.4
              - ts: 2025-11-05T18:10:00Z
                mean: 78.9
                p90: 91.2

      - component_label: epp-0  # References scenario.stack[1].metadata.label
        aggregate:
          cpu_utilization:
            units: percent
            mean: 18.4
            min: 5.2
            p50: 17.1
            p90: 28.9
            p99: 35.6
            max: 42.0
          cpu_memory_usage:
            units: GiB
            mean: 2.8
            min: 1.2
            p50: 2.6
            p90: 4.1
            p99: 5.3
            max: 5.8
          running_requests:
            units: count
            mean: 12.3
            min: 0.0
            p50: 11.0
            p90: 22.0
            p99: 34.0
            max: 40.0
          waiting_requests:
            units: count
            mean: 1.5
            min: 0.0
            p50: 1.0
            p90: 4.0
            p99: 9.0
            max: 12.0

    # Request drop rate across the system
    drop_rate:
      units: percent
      mean: 0.02
      min: 0.0
      p50: 0.0
      p90: 0.05
      p99: 0.12
      max: 0.18

    # Pod startup times (creation to Ready, per pod)
    # Collected continuously — new pods from autoscaling are detected each interval
    pod_startup_times:
      collected_at: '2025-11-05T18:17:00Z'
      pods:
        - name: qwen3-0p6b-decode-pod-1
          model: Qwen/Qwen3-0.6B
          role: decode
          node: gpu-node-01
          creation_timestamp: '2025-11-05T17:55:00Z'
          ready_timestamp: '2025-11-05T17:56:12Z'
          startup_seconds: 72.0
        - name: qwen3-0p6b-decode-pod-2
          model: Qwen/Qwen3-0.6B
          role: decode
          node: gpu-node-02
          creation_timestamp: '2025-11-05T17:55:00Z'
          ready_timestamp: '2025-11-05T17:56:25Z'
          startup_seconds: 85.0
        - name: qwen3-0p6b-decode-pod-3
          model: Qwen/Qwen3-0.6B
          role: decode
          node: gpu-node-03
          creation_timestamp: '2025-11-05T18:08:00Z'
          ready_timestamp: '2025-11-05T18:09:35Z'
          startup_seconds: 95.0
      # Aggregate statistics across all pod startups
      aggregate:
        units: s
        mean: 84.0
        min: 72.0
        p50: 85.0
        p90: 93.0
        p99: 94.8
        max: 95.0
      graph_path: metrics/graphs/pod_startup_times.png

    # Replica status — polled every 15s during experiment
    replica_status:
      namespace: benchmark-ns
      timestamp: '2025-11-05T18:17:00Z'
      # Latest snapshot
      controllers:
        - kind: Deployment
          name: qwen3-0p6b-decode
          model: Qwen/Qwen3-0.6B
          role: decode
          desired_replicas: 3
          available_replicas: 3
          ready_replicas: 3
          updated_replicas: 3
        - kind: Deployment
          name: qwen3-0p6b-prefill
          model: Qwen/Qwen3-0.6B
          role: prefill
          desired_replicas: 0
          available_replicas: 0
          ready_replicas: 0
          updated_replicas: 0
      # Time series of snapshots
      time_series:
        - timestamp: '2025-11-05T18:00:40Z'
          namespace: benchmark-ns
          controllers:
            - kind: Deployment
              name: qwen3-0p6b-decode
              model: Qwen/Qwen3-0.6B
              role: decode
              desired_replicas: 2
              available_replicas: 2
              ready_replicas: 2
              updated_replicas: 2
        - timestamp: '2025-11-05T18:08:10Z'
          namespace: benchmark-ns
          controllers:
            - kind: Deployment
              name: qwen3-0p6b-decode
              model: Qwen/Qwen3-0.6B
              role: decode
              desired_replicas: 3
              available_replicas: 2
              ready_replicas: 2
              updated_replicas: 2
        - timestamp: '2025-11-05T18:09:40Z'
          namespace: benchmark-ns
          controllers:
            - kind: Deployment
              name: qwen3-0p6b-decode
              model: Qwen/Qwen3-0.6B
              role: decode
              desired_replicas: 3
              available_replicas: 3
              ready_replicas: 3
              updated_replicas: 3
      # Aggregate ready replicas over time
      aggregate_ready_replicas:
        units: count
        mean: 2.4
        min: 2
        p50: 2
        p90: 3
        p99: 3
        max: 3
      graph_path: metrics/graphs/replica_status.png

    # -----------------------------------------------------------------------
    # Ad-hoc per-metric entries (produced by metrics_processor.py)
    # These are not yet formalized in the schema and pass via extra="allow".
    # They document what real benchmark reports look like in practice.
    # -----------------------------------------------------------------------

    # vLLM metrics — KV cache and scheduling
    vllm_kv_cache_usage_perc:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 42.5
            p50: 41.8
            p99: 85.6
            stddev: 15.2
            units: percent
            graph_path: metrics/graphs/vllm_kv_cache_usage_perc.png
      # Aggregated across all pods
      aggregated:
        mean: 40.3
        p50: 39.8
        p99: 82.1
        stddev: 14.8
        units: percent
    vllm_num_requests_running:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 24.6
            p50: 23.0
            p99: 56.0
            stddev: 12.1
            units: count
            graph_path: metrics/graphs/vllm_num_requests_running.png
      # Aggregated across all pods (present when multiple pods exist)
      aggregated:
        mean: 23.8
        p50: 22.0
        p99: 54.0
        stddev: 11.5
        units: count
    vllm_num_requests_waiting:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 3.2
            p50: 2.0
            p99: 18.0
            stddev: 4.5
            units: count
            graph_path: metrics/graphs/vllm_num_requests_waiting.png
    vllm_num_preemptions_total:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 0.5
            p50: 0.0
            p99: 3.0
            stddev: 0.8
            units: count
            graph_path: metrics/graphs/vllm_num_preemptions_total.png

    # vLLM metrics — Prefix cache
    vllm_prefix_cache_hits_total:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 12500.0
            p50: 12400.0
            p99: 14200.0
            stddev: 850.0
            units: tokens
            graph_path: metrics/graphs/vllm_prefix_cache_hits_total.png
    vllm_prefix_cache_queries_total:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 19100.0
            p50: 19000.0
            p99: 21500.0
            stddev: 1200.0
            units: tokens
            graph_path: metrics/graphs/vllm_prefix_cache_queries_total.png
    vllm_prefix_cache_hit_rate:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 65.3
            p50: 68.1
            p99: 92.4
            stddev: 18.7
            units: percent
            graph_path: metrics/graphs/vllm_prefix_cache_hit_rate.png

    # vLLM metrics — External prefix cache
    vllm_external_prefix_cache_hits_total:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 8200.0
            p50: 8100.0
            p99: 9500.0
            stddev: 620.0
            units: tokens
            graph_path: metrics/graphs/vllm_external_prefix_cache_hits_total.png
    vllm_external_prefix_cache_queries_total:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 19100.0
            p50: 19000.0
            p99: 21500.0
            stddev: 1200.0
            units: tokens
            graph_path: metrics/graphs/vllm_external_prefix_cache_queries_total.png
    vllm_external_prefix_cache_hit_rate:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 42.9
            p50: 44.2
            p99: 68.5
            stddev: 14.3
            units: percent
            graph_path: metrics/graphs/vllm_external_prefix_cache_hit_rate.png

    # vLLM metrics — NIXL KV transfer
    vllm_nixl_xfer_time_seconds_sum:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 0.0045
            p50: 0.0042
            p99: 0.0098
            stddev: 0.0015
            units: seconds
            graph_path: metrics/graphs/vllm_nixl_xfer_time_seconds_sum.png
    vllm_nixl_xfer_time_seconds_count:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 450.0
            p50: 440.0
            p99: 520.0
            stddev: 35.0
            units: count
            graph_path: metrics/graphs/vllm_nixl_xfer_time_seconds_count.png
    vllm_nixl_bytes_transferred_sum:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 1.25e+09
            p50: 1.22e+09
            p99: 1.48e+09
            stddev: 8.5e+07
            units: bytes
            graph_path: metrics/graphs/vllm_nixl_bytes_transferred_sum.png
    vllm_nixl_bytes_transferred_count:
      components:
        - component_id: decode-engine
          pod: qwen3-0p6b-decode-pod-1
          role: decode
          statistics:
            mean: 450.0
            p50: 440.0
            p99: 520.0
            stddev: 35.0
            units: count
            graph_path: metrics/graphs/vllm_nixl_bytes_transferred_count.png

    # EPP (inference scheduler) Prometheus metrics — pool-level gauges
    # These are pre-aggregated by the EPP across all endpoints.
    # Scraped from EPP pod's /metrics endpoint (port 9090).
    epp_pool_avg_kv_cache_utilization:
      components:
        - component_id: epp-0
          pod: qwen3-0p6b-epp-pod-1
          role: epp
          statistics:
            mean: 40.3
            p50: 39.8
            p99: 82.1
            stddev: 14.8
            units: percent
            graph_path: metrics/graphs/epp_pool_avg_kv_cache_utilization.png
    epp_pool_avg_queue_size:
      components:
        - component_id: epp-0
          pod: qwen3-0p6b-epp-pod-1
          role: epp
          statistics:
            mean: 4.7
            p50: 3.0
            p99: 22.0
            stddev: 5.8
            units: count
            graph_path: metrics/graphs/epp_pool_avg_queue_size.png
    epp_pool_avg_running_requests:
      components:
        - component_id: epp-0
          pod: qwen3-0p6b-epp-pod-1
          role: epp
          statistics:
            mean: 36.9
            p50: 35.0
            p99: 68.0
            stddev: 16.2
            units: count
            graph_path: metrics/graphs/epp_pool_avg_running_requests.png
    epp_pool_ready_pods:
      components:
        - component_id: epp-0
          pod: qwen3-0p6b-epp-pod-1
          role: epp
          statistics:
            mean: 2.4
            p50: 2.0
            p99: 3.0
            stddev: 0.5
            units: count
            graph_path: metrics/graphs/epp_pool_ready_pods.png

    # EPP (inference scheduler) log-derived metrics
    epp_dispatch_latency:
      statistics:
        mean: 0.0012
        p50: 0.0010
        p99: 0.0035
        stddev: 0.0008
        units: seconds
        graph_path: metrics/graphs/epp_dispatch_latency.png
    epp_endpoint_scores:
      components:
        - component_id: qwen3-0p6b-decode-pod-1
          statistics:
            mean: 0.82
            p50: 0.85
            p99: 0.95
            stddev: 0.08
            units: score
            graph_path: metrics/graphs/epp_endpoint_scores.png
    epp_request_distribution:
      components:
        - component_id: qwen3-0p6b-decode-pod-1
          statistics:
            count: 245
            units: count
            graph_path: metrics/graphs/epp_request_distribution.png
        - component_id: qwen3-0p6b-decode-pod-2
          statistics:
            count: 255
            units: count
            graph_path: metrics/graphs/epp_request_distribution.png

  profiling:
    anything_goes: 5 # The schema under profiling is currently open-ended

  component_health: # Component health and reliability metrics during benchmark
    - component_label: vllm-svc-0  # References scenario.stack[0].metadata.label
      total_restarts: 3 # This includes restarts for all replicas
      failed_replicas: 1 # Number of replicas that had a restart
      replica_health:
        - replica_id: vllm-svc-0-pod-1 # Some unique ID to identify a replica
          restarts: 1
          healthy: True # Status at completion of benchmark
          logs: https://logs.example.com/vllm-svc-0-pod-1
        - replica_id: vllm-svc-0-pod-2
          restarts: 2
          healthy: False
          logs: /path/to/logs/vllm.log
        - replica_id: vllm-svc-0-pod-3
          restarts: 0
          healthy: True
          logs: s3://path/to/logs/vllm.log

  # Session-level metrics for multi-turn workloads (inference-perf only).
  # Populated from *_session_lifecycle_metrics.json files; one report per stage.
  session_performance:
    sessions:
      total: 112
      succeeded: 110
      failed: 2
      total_events: 1340
      total_events_completed: 1320
      total_events_cancelled: 20
      session_rate:
        units: queries/s
        mean: 2.24
      session_duration:
        units: s
        mean: 48.3
        min: 12.1
        p50: 47.9
        p90: 61.2
        p99: 74.5
        max: 80.1
      events_per_session:
        units: count
        mean: 11.96
        min: 1.0
        p50: 12.0
        p99: 20.0
        max: 20.0
      events_cancelled_per_session:
        units: count
        mean: 0.18
        min: 0.0
        p50: 0.0
        p99: 2.0
        max: 3.0
      input_tokens_per_session:
        units: count
        mean: 25612.4
        min: 2148.0
        p50: 25400.0
        p99: 43020.0
        max: 46900.0
      output_tokens_per_session:
        units: count
        mean: 11800.2
        min: 980.0
        p50: 11700.0
        p99: 20100.0
        max: 22000.0