drift.yaml

# =============================================================================
# Drift detection configuration
# =============================================================================
# Shared by the offline pipeline drift stage and online monitoring.
# All thresholds and severity levels are fully configurable.

drift:
  # Master switch — set to false to disable drift analysis entirely
  enabled: true

  # Which preprocessed split to use as reference baseline
  # Currently only "train" is supported
  reference_source: "train"

  # Statistical tests used by Evidently for drift detection
  stattest:
    numerical: "ks"              # Kolmogorov-Smirnov test for numerical features
    categorical: "chisquare"     # Chi-square test for categorical features

  # P-value thresholds below which drift is flagged per feature
  stattest_threshold:
    numerical: 0.05
    categorical: 0.05

  # Fraction of features that must drift for dataset_drift_detected=true
  drift_share: 0.5

  # Overall severity classification based on drift_share
  # drift_share <= low_max        → "low"
  # low_max < drift_share <= medium_max → "medium"
  # drift_share > medium_max      → "high"
  severity:
    low_max: 0.25
    medium_max: 0.50

  # Per-feature severity classification based on p-value
  # p-value < high_below          → "high"
  # high_below <= p-value < medium_below → "medium"
  # p-value >= medium_below (but < stattest_threshold) → "low"
  feature_severity:
    high_below: 0.001
    medium_below: 0.01

  # Online monitoring settings
  monitoring:
    enabled: true
    # Minimum rows in a batch to compute drift (below this, skip).
    # Matches the small-sample threshold used by the statistical tests —
    # below 50 rows KS / chi-square power degrades sharply, so results
    # carry a "small_sample" warning on the drift report.
    min_batch_size: 10
    # Notify user when severity reaches this level or higher
    alert_severity: "medium"
    # CI gate: in non-interactive (non-TTY) runs, exit non-zero (code 2) when
    # observed severity meets or exceeds this level. "never" disables the gate.
    # Allowed values: low | medium | high | never
    fail_on_severity: "high"

  # ==========================================================================
  # Image drift — ISP-based scenario simulation (raw DNG datasets only)
  # Only used when raw_input: true in the preprocessing config.
  # Ignored entirely for tabular and standard JPG/PNG image tasks.
  # ==========================================================================
  image:
    # Severity thresholds for per-channel Wasserstein image drift scores.
    # Defaults assume z-scored images (normalize=True in preprocessing).
    # score < medium              → "low"
    # medium <= score < high      → "medium"
    # score >= high               → "high"
    # Rule of thumb: medium ≈ 2× noise floor, high ≈ 5× noise floor.
    severity:
      medium: 0.10
      high: 0.25

    # Drift scenarios: each defines a name, optional description, and ISP
    # parameter overrides relative to the baseline preprocessing config.
    # Only explicitly listed parameters are changed; all others keep their
    # baseline values. Parameter values are clipped to parameter_bounds before
    # use and a warning is logged if clipping occurs.
    scenarios:
      - name: "wb_shift_warm"
        description: "White balance shifted toward warm tones (red gain increased, blue gain reduced)"
        isp:
          white_balance:
            r_gain: 1.8
            b_gain: 0.7

      - name: "wb_shift_cool"
        description: "White balance shifted toward cool tones (red gain reduced, blue gain increased)"
        isp:
          white_balance:
            r_gain: 0.7
            b_gain: 1.8

      - name: "high_noise"
        description: "Denoising disabled — simulates high-ISO sensor noise reaching the model"
        isp:
          denoising:
            strength: 0.0

      - name: "dark_gamma"
        description: "High gamma — simulates an underexposed or dark pipeline output"
        isp:
          gamma_correction:
            gamma: 2.8

      - name: "bright_gamma"
        description: "Low gamma — simulates an overexposed or bright pipeline output"
        isp:
          gamma_correction:
            gamma: 1.2

      - name: "heavy_sharpening"
        description: "Aggressive unsharp mask — simulates over-sharpened pipeline"
        isp:
          sharpening:
            amount: 2.5
            radius: 2.0

    # Bounds enforced on all scenario ISP overrides. Values outside these
    # ranges are clipped to the nearest bound and logged as warnings.
    # Keys use the format "isp_section.parameter_name".
    parameter_bounds:
      white_balance.r_gain:   [0.5, 2.5]
      white_balance.g_gain:   [0.5, 2.0]
      white_balance.b_gain:   [0.5, 2.0]
      denoising.strength:     [0.0, 1.0]
      sharpening.radius:      [0.1, 5.0]
      sharpening.amount:      [0.0, 3.0]
      gamma_correction.gamma: [1.0, 2.8]