Use case 1: compare clients
Run native fetch, axios, ky, and ffetch against the exact same chaos profile and workload. This isolates client-level behavior under identical transport conditions.
chaos arena
The fetch client chaos arena
A browser-based benchmark that runs fetch, axios, ky, and ffetch side-by-side under identical controlled chaos. No install, no server, no account - opens directly in your browser.
Use case 2: diagnose resilience settings
Tune timeout, retries, and plugin configuration in a repeatable environment to see how your strategy shifts tail latency, success rate, and error shape.
What you can do in the arena
Every setting is live - change it and re-run. The entire arena state encodes into the URL so you can share a specific experiment with a link.
Configure the chaos layer
The built-in chaos engine (powered by chaos-fetch) intercepts every outbound request before it hits the network. You compose global rules that apply to all clients simultaneously:
- latency - fixed delay added to every request
- latencyRange - random delay between a min and max, simulating jitter
- failRandomly - probabilistic failures at a configurable rate and HTTP status
- failNth - fail every Nth request in sequence with a configurable status
- fail - always short-circuit with a fixed response, useful for testing circuit breakers
- rateLimit - limit requests per time window; when
retryAfterMsis set the 429 response includes aRetry-Afterheader that smart clients like ffetch and ky will respect - throttle - bandwidth shaping; limit response transfer rate in bytes/sec with configurable chunk size
Rules are executed in order and can be stacked. Start from one of eight presets (zero-config, light, api-instability, meltdown-recovery, rate-limited, slow-network, burst-traffic, brownout) or build your own.
Configure each client independently
Each of the four clients has its own configuration panel. Cards are draggable so you can reorder them visually.
fetch (native)
No config - the browser baseline. No timeout, no retry, no error handling. Shows what you get with zero effort.
axios
Configurable timeout. No retry built-in - shows where a popular client sits without a retry layer.
ky
Timeout, retry limit, retry methods, retry status codes, Retry-After awareness, backoff base and max. A capable baseline with sensible ergonomics.
ffetch
Everything ky has, plus three opt-in plugins you can toggle per run: dedupe (collapse identical in-flight requests), circuit breaker (stop retrying a broken upstream), and hedge (race a backup request to cut tail latency). Plugin execution order is configurable.
Run settings
Set the total request count (default 60), concurrency (default 6 simultaneous requests), and target URL. You can point the arena at any CORS-enabled endpoint - use a local dev server or a public API.
What gets measured
- Reliability score (0-100) - composite metric: success rate minus penalties for timeouts, thrown exceptions, and p95 tail latency
- Latency p50 / p95 / p99 - median, high-percentile, and tail latency in ms
- Error rate and error distribution - breakdown by outcome type: 2xx success, deduped-ok, TimeoutError, CircuitOpenError, HTTP 429, HTTP 500/503, other
- Throughput - requests per second across the full run
- Retry attempts - total transport calls vs successful outcomes, revealing amplification
- Peak and average in-flight - concurrency profile per client
Automated diagnosis
After each run a diagnostic engine analyzes each client against the chaos conditions and surfaces two things:
- Observations - detected problems with a confidence score, e.g. "timeout is too tight for observed p95" or "no retry configured but 20% of requests failed"
- Recommended tweaks - specific actionable suggestions, e.g. "increase timeout to 4500ms" or "enable circuit breaker with threshold 5"
The engine includes guard rules that suppress harmful recommendations - for example it won't suggest adding more retries if your timeout is already too tight, and won't suggest a circuit breaker for rate-limit problems. When a recommendation is suppressed the UI shows the reason.
Share and reproduce
Every chaos config, client setting, and run parameter is encoded into the URL. Copy the address bar to get a shareable link that reproduces the exact same experiment in anyone's browser.