Projection that computes statistical properties over event streams.

Traditional assertions check exact conditions. Statistical projections enable probabilistic assertions like:

• "p99 latency < 100ms"
• "error rate < 1%"
• "success rate > 99% over last 100 operations"

Why Statistical Projections?

Nemesis can inject latency, but we need statistical assertions to verify the system handles it gracefully:

• A single slow request isn't a bug
• 50% of requests being slow IS a bug
• 1% error rate might be acceptable; 10% is not

Usage

defmodule MyModel do
  def assertion_projections do
    [
      {PropertyDamage.Model.Projection.Statistics, [
        window_size: 100,
        assertions: [
          {:p99_latency_ms, :less_than, 500},
          {:error_rate, :less_than, 0.05},
          {:success_rate, :greater_than, 0.95}
        ]
      ]}
    ]
  end
end

Tracked Metrics

• :p50_latency_ms - Median latency
• :p95_latency_ms - 95th percentile latency
• :p99_latency_ms - 99th percentile latency
• :max_latency_ms - Maximum observed latency
• :mean_latency_ms - Average latency
• :success_rate - Ratio of successes to total
• :error_rate - Ratio of errors to total
• :throughput - Operations per second (requires time tracking)

Recording Metrics

Events should include latency information. The projection looks for:

• :latency_ms or :duration_ms fields for latency tracking
• Success events vs error events for rate calculation

You can also record metrics explicitly in your adapter.

Limitations

• Statistics only meaningful with sufficient sample size
• Shrinking statistical failures is problematic (minimal case may not reproduce)
• Thresholds are environment-dependent (CI vs production hardware)