Performance Benchmarks

Ablauf ships with a built-in benchmark endpoint that runs identical workloads on both engines side-by-side, so you can compare performance yourself.

Results

The following results were measured on a deployed Cloudflare Worker. Both engines executed the same 8-step workflow with identical CPU work per step, alternating execution order each round.

End-to-End Completion

Time from workflow creation to completion (lower is better).

Ablauf completes workflows in 4.5% of the time it takes native Cloudflare Workflows — roughly 22x faster.

Startup Latency

Time from creation request to the first step starting execution.

Ablauf starts executing steps in 3.4% of the time — roughly 29x faster startup.

Throughput

Steps executed per second (higher is better).

Ablauf sustains 13x higher throughput with significantly lower variance (stddev 0.96 vs 0.77).

Per-Step Orchestration Overhead

Cloudflare Workflows adds measurable orchestration overhead between steps. Ablauf runs all steps in a single execution pass with zero inter-step overhead.

Variance

Ablauf shows tight, predictable latency. Cloudflare Workflows has high variance across runs.

Benchmark Endpoint

The Worker exposes a benchmark endpoint that executes:

• A custom Ablauf workflow (benchmark-ablauf)
• A native Cloudflare Workflow (BenchmarkCloudflareWorkflow)

Both paths run the same payload, same number of steps, and same deterministic CPU work so the comparison is fair.

Usage

POST /benchmarks/workflows

This endpoint is token-protected and only runs when a BENCHMARK_TOKEN secret is configured.

npx wrangler secret put BENCHMARK_TOKEN

Then call the endpoint with x-benchmark-token.

curl -X POST https://your-worker.workers.dev/benchmarks/workflows \
  -H "content-type: application/json" \
  -H "x-benchmark-token: <your-token>" \
  -d '{
    "iterations": 10,
    "warmups": 4,
    "steps": 8,
    "workIterations": 5000,
    "pollIntervalMs": 10
  }'

Fairness Rules

• Alternating run order per measured round (ablauf -> cloudflare, then cloudflare -> ablauf)
• Warmup rounds are executed but discarded from final stats
• Identical payload shape and work units for both implementations
• Same polling interval and completion criteria

Metrics Returned

The response includes per-run and aggregated metrics for both engines:

• createMs — latency for workflow instance creation call
• startupMs — request-to-first-step-start latency
• completionMs — end-to-end latency until workflow completes
• runMs — workflow runtime measured inside the workflow implementation
• totalStepWallMs — sum of wall-clock step durations
• totalCallbackMs — sum of callback execution durations
• totalOrchestrationMs — estimated scheduler/orchestration overhead (wall - callback)
• throughputStepsPerSecond
• perStep — p50/p95/min/max/mean/stddev for each step

The comparison object reports mean deltas and ratios for key metrics.

Notes

• If BENCHMARK_TOKEN is missing, the endpoint returns 503.
• If the token header is wrong or missing, the endpoint returns 401.
• Limits for Worker CPU/memory and request behavior can be checked at Cloudflare Workers platform limits: https://developers.cloudflare.com/workers/platform/limits/