Presets Guide

Presets are pre-built configuration profiles that set error rates, latency, burst patterns, and response generation for common testing scenarios. Instead of manually configuring dozens of fields, pick a preset that matches your testing goal.

Using Presets

# CLI
uv run chaosllm serve --preset=realistic
uv run chaosweb serve --preset=stress_scraping

# Python
from errorworks.llm.config import load_config
config = load_config(preset="realistic")

ChaosLLM Presets

Comparison Table

Preset	Rate Limit	Capacity 529	Server Errors	Connection Failures	Malformed	Burst	Latency
silent	0%	0%	0%	0%	0%	Off	10 +/- 5ms
gentle	1%	0.5%	0.5% svc unavail	0%	0%	Off	50 +/- 20ms
realistic	5%	2%	0.5% svc, 0.2% internal, 0.1% bad gw, 0.2% gw timeout	0.2% timeout, 0.1% reset, 1% slow	0.1% invalid JSON, 0.1% truncated, 0.1% missing fields	60s/5s (50% rl, 30% cap)	100 +/- 50ms
stress_aimd	15%	5%	2% svc, 0.5% internal, 0.2% bad gw, 0.3% gw timeout	0.5% slow	0%	30s/5s (80% rl, 50% cap)	30 +/- 15ms
stress_extreme	20%	10%	3% svc, 5% internal, 1% bad gw, 1% gw timeout	0%	3% invalid JSON, 2% truncated	15s/5s (90% rl, 70% cap)	10 +/- 5ms
chaos	6.25%	3.13%	1.88% svc, 1.88% internal, 1.25% bad gw, 1.25% gw timeout	1.25% timeout, 0.94% reset, 1.88% slow	1.25% invalid JSON, 0.94% truncated, 0.63% empty, 0.94% missing, 0.31% wrong ct	20s/8s (90% rl, 70% cap)	100 +/- 100ms

When to Use Each

Preset	Use When
silent	Measuring maximum pipeline throughput without noise
gentle	Verifying basic pipeline operation and debugging
realistic	Testing against production-like Azure OpenAI conditions
stress_aimd	Tuning AIMD throttle parameters and backoff behavior
stress_extreme	Verifying pipeline survives under harsh HTTP error rates
chaos	Achieving error handling coverage across every failure type

ChaosWeb Presets

Comparison Table

Preset	Rate Limit	Forbidden	Not Found	Connection Failures	Content Malform.	Redirects	Burst	Latency
silent	0%	0%	0%	0%	0%	0%	Off	200 +/- 100ms
gentle	1%	0%	1%	0%	0%	0%	Off	100 +/- 50ms
realistic	5%	3%	2%	0.5% timeout, 0.2% reset, 5% slow	1% wrong ct, 1% encoding, 0.5% truncated	0%	60s/8s (40% rl, 30% forbid)	300 +/- 150ms
stress_scraping	15%	10%	3%	5% timeout, 3% reset, 8% slow, 2% incomplete	3% wrong ct, 2% encoding, 2% truncated, 1% charset	1% SSRF	60s/10s (80% rl, 50% forbid)	500 +/- 200ms
stress_extreme	25%	15%	5%	10% timeout, 5% reset, 3% stall, 5% slow, 5% incomplete	5% wrong ct, 3% encoding, 3% truncated, 2% invalid enc, 2% charset, 1% meta	3% loop, 2% SSRF	30s/8s (90% rl, 70% forbid)	800 +/- 400ms

When to Use Each

Preset	Use When
silent	Measuring scraping throughput without error injection overhead
gentle	Verifying basic scraping functionality works before adding stress
realistic	Testing against typical web scraping production conditions
stress_scraping	Verifying retry logic, backoff, and error routing under pressure
stress_extreme	Finding failure modes and verifying graceful degradation

Combining Presets with Overrides

Presets provide the base configuration, but you can override any setting on top. Configuration precedence (highest to lowest):

1. CLI flags -- Override individual settings
2. Config file -- --config=my-config.yaml
3. Preset -- --preset=realistic
4. Defaults -- Built-in Pydantic defaults

Example: Preset with CLI Override

# Start with realistic, but increase rate limiting
uv run chaosllm serve --preset=realistic --rate-limit-pct=20.0

Example: Preset with Config File Override

# my-overrides.yaml
error_injection:
  rate_limit_pct: 20.0
  burst:
    interval_sec: 30  # More frequent bursts

uv run chaosllm serve --preset=realistic --config=my-overrides.yaml

The config file's rate_limit_pct: 20.0 overrides the preset's 5.0, and the burst interval_sec: 30 overrides the preset's 60. All other preset values are preserved.

Example: Python with Overrides

from errorworks.llm.config import load_config

config = load_config(
    preset="realistic",
    cli_overrides={
        "error_injection": {"rate_limit_pct": 20.0},
        "latency": {"base_ms": 200},
    },
)

Tip

Deep merge means you only need to specify the fields you want to change. Nested objects are merged recursively -- you do not need to repeat the entire burst section just to change interval_sec.

Listing Available Presets

from errorworks.llm.config import list_presets as list_llm_presets
from errorworks.web.config import list_presets as list_web_presets

print(list_llm_presets())  # ['chaos', 'gentle', 'realistic', 'silent', 'stress_aimd', 'stress_extreme']
print(list_web_presets())  # ['gentle', 'realistic', 'silent', 'stress_extreme', 'stress_scraping']

Related Pages

• ChaosLLM -- Full ChaosLLM endpoint and error injection reference
• ChaosWeb -- Full ChaosWeb endpoint and error injection reference
• Configuration -- YAML config file structure and precedence details