Skip to content

Data Readiness

The nomad readiness command helps administrators determine when sufficient data has been collected for reliable ML predictions.

Quick Start

nomad readiness              # Basic readiness report
nomad readiness -v           # Verbose with feature details
nomad readiness --db mydata.db   # Specify database

What It Assesses

The readiness estimator evaluates four dimensions:

1. Sample Size

Level Jobs Score
Minimum 100+ 40%
Recommended 500+ 70%
Optimal 1000+ 100%

2. Class Balance

Compares successful vs failed jobs. Ideal ratio is between 60:40 and 90:10. Extremely imbalanced data (>95% one class) reduces prediction reliability.

3. Feature Coverage

Checks that all 17 feature dimensions have sufficient variance:

  • CPU efficiency
  • Memory utilization
  • I/O wait percentage
  • NFS write ratio
  • Runtime characteristics
  • Exit signals
  • And more...

Features with zero variance (all same value) don't contribute to predictions.

4. Data Recency

Recent data is weighted more heavily. The estimator checks:

  • Jobs from last 24 hours
  • Jobs from last 7 days
  • Jobs from last 30 days

Stale data (>30 days old only) triggers a warning.

Output Example

======================================================================
                       NOMAD Data Readiness
======================================================================
  Overall Score: 72% (Recommended)
======================================================================

  Sample Size       ================....  847 jobs (85%)
  Class Balance     ==============......  82:18 ratio (70%)
  Feature Coverage  ====================  17/17 features
  Data Recency      ==========..........  3 days old (50%)

======================================================================
  Forecast: At 125 jobs/day, optimal threshold (1000)
            will be reached in approximately 2 days.
======================================================================

Verbose Mode

With -v or --verbose, see per-feature statistics: 

nomad readiness -v

Shows coefficient of variation (CV%) for each feature, helping identify which metrics contribute most to predictions.

Recommendations

The command provides actionable recommendations:

  • Low sample size: "Continue collecting data. Run nomad collect to gather more jobs."
  • Class imbalance: "Dataset is heavily skewed toward successful jobs. Consider longer collection period to capture more failure modes."
  • Missing features: "GPU metrics unavailable. Install nvidia-smi for GPU-enabled nodes."
  • Stale data: "Most recent job is 15 days old. Ensure collectors are running."