#!/usr/bin/env python3 """Rate-controlled paired scale test for 8 versus 12 API replicas.""" import json import os import random import statistics import time from pathlib import Path import track1_experiment as bench from track1_offered_load import run_sample from client import Client ROOT = Path(__file__).resolve().parents[1] OUT = ROOT / "evidence" REPEATS = 3 RATES = (1000,) REPLICAS = (8, 12) def main(): OUT.mkdir(exist_ok=True) env = os.environ.copy() env.update(API_CPUS="0.5", API_MEMORY="256m", DB_POOL_MAX="10", LB_CPUS="0.25") bench.run(bench.COMPOSE + ["up", "-d", "--build", "--scale", "api=8", "--wait", "--wait-timeout", "180"], env) bench.wait_ready() time.sleep(5) client = Client("http://127.0.0.1:8080") actor, _, _, chat = client.bootstrap() rng = random.Random(20261107) samples = [] last_replica_count = 8 for repeat in range(1, REPEATS + 1): block = [(replicas, rate) for replicas in REPLICAS for rate in RATES] rng.shuffle(block) for replicas, rate in block: if last_replica_count != 8: bench.run(bench.COMPOSE + ["up", "-d", "--build", "--scale", "api=8", "--wait", "--wait-timeout", "180"], env) bench.wait_ready() time.sleep(5) last_replica_count = 8 bench.reset_synthetic_messages() if replicas != last_replica_count: bench.run(bench.COMPOSE + ["up", "-d", "--build", "--scale", f"api={replicas}", "--wait", "--wait-timeout", "180"], env) bench.wait_ready() time.sleep(5) last_replica_count = replicas row = run_sample(actor, chat, rate, repeat, reset=False) row["replicas"] = replicas row["sample"] = repeat samples.append(row) print(json.dumps(row, ensure_ascii=False), flush=True) summary = [] for replicas in REPLICAS: for rate in RATES: group = [r for r in samples if r["replicas"] == replicas and r["offered_rate_rps"] == rate] summary.append({ "replicas": replicas, "offered_rate_rps": rate, "completed_success_rps_total_elapsed_median": round(statistics.median(r["completed_success_rps_total_elapsed"] for r in group), 3), "completed_success_rps_min": min(r["completed_success_rps_total_elapsed"] for r in group), "completed_success_rps_max": max(r["completed_success_rps_total_elapsed"] for r in group), "success_ratio_median": round(statistics.median(r["success_ratio"] for r in group), 6), "p95_ms_median": round(statistics.median(r["latency_successful_ms"]["p95"] for r in group), 3), "error_count_total": sum(r["failed_requests"] for r in group), "message_rows_after_median": int(statistics.median(r["message_rows_after"] for r in group if r["message_rows_after"] is not None)) if any(r["message_rows_after"] is not None for r in group) else None, }) result = { "method": "Randomized paired blocks at offered 1000 RPS, 15s per sample; three repeats; compare 8 versus 12 replicas at fixed 0.5 CPU/256 MiB per API replica; DB_POOL_MAX=10, PostgreSQL 2 CPU/1 GiB, HAProxy 0.25 CPU; HAProxy server-template supports 16 backends; scale to 8 before each DB reset to release idle connections, then scale to target; synthetic messages and idempotency keys truncated per sample; HTTP client reused per worker.", "summary": summary, "samples": samples, } (OUT / "offered-scale-results.json").write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n") print(json.dumps(summary, ensure_ascii=False, indent=2)) if __name__ == "__main__": main()