---
# Constant Rate Stress Test
# Test system stability under sustained load

test_scenario:
  name: "constant-rate-stress"
  type: "stress"
  description: "Test system stability and consistency under sustained constant-rate audio transcription load"

  # Test configuration
  backend: "openai-audio"
  endpoint: "/v1/audio/transcriptions"
  request_type: "audio_transcriptions"

  # Dataset configuration
  dataset:
    name: "openslr/librispeech_asr"
    config: "clean"
    split: "test"
    audio_column: "audio"
    num_samples: -1  # Use all available

  # Audio preprocessing
  audio_config:
    format: "mp3"
    bitrate: "64k"
    sample_rate: 16000
    mono: true
    max_duration: null

  # vLLM server configuration
  server:
    dtype: "float16"
    env_vars:
      VLLM_CPU_KVCACHE_SPACE: "2GiB"

  # Test stages - Different sustained rates
  # Note: Rates should be determined from baseline throughput tests
  # These are example values - adjust based on your system capacity
  stages:
    # Stage 1: Low sustained rate (30% of max capacity)
    - name: "low-rate-sustained"
      description: "Sustained load at 2 req/s for 5 minutes"
      profile: "constant"
      rate: 2  # requests per second
      max_seconds: 300  # 5 minutes
      result_filename_suffix: "constant-2rps"
      metrics_focus:
        - "Latency stability over time"
        - "Zero error rate"
        - "Consistent throughput"

    # Stage 2: Medium sustained rate (50% of max capacity)
    - name: "medium-rate-sustained"
      description: "Sustained load at 5 req/s for 5 minutes"
      profile: "constant"
      rate: 5  # requests per second
      max_seconds: 300  # 5 minutes
      result_filename_suffix: "constant-5rps"
      metrics_focus:
        - "Latency degradation over time"
        - "Memory stability"
        - "CPU utilization"

    # Stage 3: High sustained rate (75% of max capacity)
    - name: "high-rate-sustained"
      description: "Sustained load at 10 req/s for 5 minutes"
      profile: "constant"
      rate: 10  # requests per second
      max_seconds: 300  # 5 minutes
      result_filename_suffix: "constant-10rps"
      metrics_focus:
        - "System stability under high load"
        - "Queueing behavior"
        - "Resource exhaustion indicators"

    # Stage 4: Extended duration test
    - name: "extended-duration"
      description: "Extended test at medium rate for 15 minutes"
      profile: "constant"
      rate: 5  # requests per second
      max_seconds: 900  # 15 minutes
      result_filename_suffix: "constant-5rps-15min"
      metrics_focus:
        - "Memory leaks"
        - "Performance degradation over time"
        - "System stability"

  # Key metrics
  metrics:
    primary:
      # Throughput metrics
      - "Achieved rate (req/s)"
      - "Rate deviation from target (%)"

      # Latency metrics over time
      - "Mean latency (ms)"
      - "P95 latency (ms)"
      - "P99 latency (ms)"
      - "Latency trend (slope over time)"

      # Stability metrics
      - "Error rate (%)"
      - "Request completion rate (%)"
      - "Latency coefficient of variation"

    system_metrics:
      - "CPU utilization (%)"
      - "Memory usage (GB)"
      - "Memory growth rate (MB/min)"

    derived:
      - "Latency drift = (final_latency - initial_latency) / initial_latency"
      - "System stability score = 1 - (error_rate + latency_cv)"

  # Analysis
  analysis:
    time_series:
      - "Latency over time (detect degradation)"
      - "Throughput over time (detect drops)"
      - "Error rate over time"

    questions:
      - "Does latency remain stable over time?"
      - "Are there any memory leaks?"
      - "Does the system maintain target rate?"
      - "How does queueing behave under load?"

  # Success criteria
  success_criteria:
    - "Achieved rate within 5% of target rate"
    - "Error rate < 0.1%"
    - "Latency increase over time < 20%"
    - "No memory leaks (stable memory usage)"
    - "P99 latency < 5000ms"

  results:
    format: "json"
    location: "results/audio-models/{model}/constant-rate-stress/"
    graphs:
      - type: "latency_timeseries"
        title: "Latency Over Time by Rate"
        x_axis: "Time (seconds)"
        y_axis: "Latency (ms)"
        series:
          - "2 req/s"
          - "5 req/s"
          - "10 req/s"

      - type: "throughput_consistency"
        title: "Achieved vs Target Rate"
        x_axis: "Target Rate (req/s)"
        y_axis: "Achieved Rate (req/s)"
        reference_line: "y=x (perfect tracking)"

      - type: "resource_utilization"
        title: "Resource Utilization Over Time"
        x_axis: "Time (seconds)"
        y_axis_left: "CPU %"
        y_axis_right: "Memory GB"

# GuideLLM command examples
guidellm_commands:
  low_rate:
    cmd: |
      guidellm benchmark \
        --target "http://localhost:8000" \
        --request-type audio_transcriptions \
        --profile constant \
        --rate 2 \
        --max-seconds 300 \
        --data openslr/librispeech_asr \
        --data-args '{"name": "clean", "split": "test"}' \
        --data-column-mapper '{"audio_column": "audio"}'

  high_rate:
    cmd: |
      guidellm benchmark \
        --target "http://localhost:8000" \
        --request-type audio_transcriptions \
        --profile constant \
        --rate 10 \
        --max-seconds 300 \
        --data openslr/librispeech_asr \
        --data-args '{"name": "clean", "split": "test"}' \
        --data-column-mapper '{"audio_column": "audio"}'