File size: 12,890 Bytes
7e9a520 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 | """AtlasOps Benchmark Runner.
Runs all 28 frozen scenarios (8 single-fault + 5 cascade + 5 multi-fault + 10 named replays)
against a model, scores them with the LLM judge, and outputs a comparison table.
Usage:
python bench/runner.py --model checkpoints/grpo_v3 --tag grpo_v3
python bench/runner.py --model checkpoints/AtlasOps_v2_baseline --tag baseline_v2
Output:
bench/results/<run_id>/results_per_episode.jsonl
bench/results/<run_id>/results_summary.json
bench/results/comparison_table.md (updates in place across runs)
"""
import argparse
import asyncio
import json
import logging
import os
import subprocess
import time
from datetime import datetime, timezone
from pathlib import Path
from agents.adversarial_designer import design_batch
from agents.coordinator import handle_incident
from agents.judge import judge_trajectory
from config.runtime import (
FROZEN_SCENARIOS,
evaluate_reward_contract,
bounded_speed_score as _bounded_speed_score,
)
# Backwards-compatible alias — tests import this name from bench.runner
_evaluate_episode_reward = evaluate_reward_contract
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
log = logging.getLogger("runner")
RESULTS_DIR = Path("bench/results")
MANIFESTS_DIR = Path("bench/chaos_manifests")
def apply_chaos(scenario_id: str) -> bool:
manifest = MANIFESTS_DIR / f"{scenario_id}.yaml"
if not manifest.exists():
log.error("manifest not found: %s", manifest)
return False
r = subprocess.run(["kubectl", "apply", "-f", str(manifest)], capture_output=True, text=True)
return r.returncode == 0
def reset_cluster() -> None:
subprocess.run(
["kubectl", "delete", "podchaos,networkchaos,stresschaos,dnschaos,iochaos,timechaos",
"--all", "-A"],
capture_output=True,
)
# Also remove any legacy deployment created by named replays
subprocess.run(["kubectl", "delete", "deployment", "checkoutservice-legacy",
"-n", "default", "--ignore-not-found=true"], capture_output=True)
time.sleep(60)
def wait_for_alert(timeout_s: int = 300) -> dict | None:
from agents.tools.alertmanager import alertmanager_list_alerts
deadline = time.time() + timeout_s
while time.time() < deadline:
result = alertmanager_list_alerts(active_only=True)
if result.get("success") and result.get("count", 0) > 0:
return {"commonLabels": {"alertname": result["alerts"][0]["alertname"]},
"alerts": result["alerts"]}
time.sleep(20)
log.warning("no alert fired within %ds — synthesising fallback", timeout_s)
return {"commonLabels": {"alertname": "BenchmarkTimeout"}, "alerts": [],
"scenario": "unknown", "synthetic": True}
async def run_scenario(scenario_id: str) -> dict:
t0 = time.time()
ok = apply_chaos(scenario_id)
if not ok:
return {"scenario_id": scenario_id, "status": "skip", "error": "manifest_apply_failed"}
alert = wait_for_alert()
alert["scenario_id"] = scenario_id
try:
incident = await handle_incident(alert)
judge_score = await judge_trajectory(incident, tier=tier)
except Exception as e:
log.exception("scenario %s failed: %s", scenario_id, e)
reset_cluster()
return {"scenario_id": scenario_id, "status": "error", "error": str(e)}
reset_cluster()
remediation = incident.get("remediation", {}).get("final", {})
triage = incident.get("triage", {}).get("final", {})
total_turns = sum(
len(incident.get(role, {}).get("trajectory", []))
for role in ("triage", "diagnosis", "remediation", "comms")
)
tier = scenario_id.split("/")[0] if "/" in scenario_id else "unknown"
episode = {
"scenario_id": scenario_id,
"tier": tier,
"status": "ok",
"outcome": remediation.get("outcome", "unknown"),
"resolved": remediation.get("outcome") == "resolved",
"time_to_resolve_s": remediation.get("time_to_resolve_seconds", round(time.time() - t0)),
"severity": triage.get("severity", "unknown"),
"total_turns": total_turns,
"judge": judge_score,
"postmortem_path": incident.get("comms", {}).get("final", {}).get("postmortem_path"),
}
# Keep reward evaluation centralized so train/eval/bench cannot drift.
episode["reward_contract"] = evaluate_reward_contract(episode)
return episode
def compute_summary(results: list[dict], tag: str, model: str) -> dict:
valid = [r for r in results if r.get("status") == "ok"]
resolved = [r for r in valid if r.get("resolved")]
cascades = [r for r in valid if r.get("tier") == "cascade"]
replays = [r for r in valid if r.get("tier") == "named_replays"]
def mean(xs, key, default=0.0):
vals = [x.get(key, default) for x in xs if x.get(key) is not None]
return round(sum(vals) / len(vals), 3) if vals else 0.0
judge_scores = [r.get("judge", {}).get("overall", 0) for r in valid if r.get("judge")]
contract_scores = [r.get("reward_contract", {}).get("total", 0) for r in valid]
penalties = [r.get("reward_contract", {}).get("penalty_total", 0) for r in valid]
per_tier = {}
tiers = sorted({r.get("tier", "unknown") for r in valid})
for tier in tiers:
trows = [r for r in valid if r.get("tier") == tier]
t_resolved = [r for r in trows if r.get("resolved")]
per_tier[tier] = {
"count": len(trows),
"resolution_rate": round(len(t_resolved) / max(len(trows), 1), 3),
"avg_time_to_resolve_s": mean(trows, "time_to_resolve_s"),
"avg_reward_contract": round(
sum(r.get("reward_contract", {}).get("total", 0) for r in trows) / max(len(trows), 1), 3
),
}
unsafe_action_count = sum(
1 for r in valid if r.get("reward_contract", {}).get("penalties", {}).get("unsafe_shortcut", 0) > 0
)
false_resolution_count = sum(
1 for r in valid if r.get("reward_contract", {}).get("penalties", {}).get("false_resolution", 0) > 0
)
hallucinated_evidence_count = sum(
1
for r in valid
if r.get("reward_contract", {}).get("penalties", {}).get("hallucinated_evidence", 0) > 0
)
return {
"tag": tag,
"model": model,
"run_date": datetime.now(timezone.utc).isoformat(),
"total_scenarios": len(results),
"resolution_rate": round(len(resolved) / max(len(valid), 1), 3),
"avg_reward": round(sum(judge_scores) / max(len(judge_scores), 1), 3),
"avg_reward_contract": round(sum(contract_scores) / max(len(contract_scores), 1), 3),
"avg_penalty": round(sum(penalties) / max(len(penalties), 1), 3),
"avg_turns": mean(valid, "total_turns"),
"avg_time_to_resolve_s": mean(valid, "time_to_resolve_s"),
"cascade_resolution_rate": round(
len([r for r in cascades if r.get("resolved")]) / max(len(cascades), 1), 3
),
"named_replay_resolution_rate": round(
len([r for r in replays if r.get("resolved")]) / max(len(replays), 1), 3
),
"unsafe_action_count": unsafe_action_count,
"false_resolution_count": false_resolution_count,
"hallucinated_evidence_count": hallucinated_evidence_count,
"per_tier": per_tier,
}
def write_comparison_table(summary: dict) -> None:
table_path = RESULTS_DIR / "comparison_table.md"
existing_runs: list[dict] = []
if table_path.exists():
# naive parse — rebuild from stored JSON summaries
for d in RESULTS_DIR.iterdir():
s_file = d / "results_summary.json"
if s_file.exists():
existing_runs.append(json.loads(s_file.read_text()))
existing_runs = [r for r in existing_runs if r.get("tag") != summary["tag"]]
existing_runs.append(summary)
existing_runs.sort(key=lambda x: x.get("run_date", ""))
header = (
"| Tag | Model | Resolution | Reward (Judge) | Reward (Contract) | Avg Penalty | Avg Turns "
"| Cascade Res. | Replay Res. | Date |\n"
)
header += "|---|---|---|---|---|---|---|---|---|---|\n"
rows = ""
for r in existing_runs:
rows += (
f"| {r['tag']} | `{Path(r['model']).name}` "
f"| {r['resolution_rate']:.0%} "
f"| {r['avg_reward']:.3f} "
f"| {r.get('avg_reward_contract', 0):.3f} "
f"| {r.get('avg_penalty', 0):.3f} "
f"| {r['avg_turns']:.1f} "
f"| {r['cascade_resolution_rate']:.0%} "
f"| {r['named_replay_resolution_rate']:.0%} "
f"| {r['run_date'][:10]} |\n"
)
per_tier_lines = ["\n## Per-tier Breakdown\n"]
for r in existing_runs:
per_tier_lines.append(f"\n### {r['tag']}\n")
per_tier_lines.append("| Tier | Count | Resolution | Avg TTR (s) | Avg Contract Reward |\n")
per_tier_lines.append("|---|---|---|---|---|\n")
for tier, item in sorted((r.get("per_tier") or {}).items()):
per_tier_lines.append(
f"| {tier} | {item.get('count', 0)} | {item.get('resolution_rate', 0):.0%} "
f"| {item.get('avg_time_to_resolve_s', 0):.1f} | {item.get('avg_reward_contract', 0):.3f} |\n"
)
per_tier_lines.append(
f"\n- unsafe actions: `{r.get('unsafe_action_count', 0)}`"
f", false resolutions: `{r.get('false_resolution_count', 0)}`"
f", hallucinated evidence: `{r.get('hallucinated_evidence_count', 0)}`\n"
)
table_path.write_text(
f"# AtlasOps — Benchmark Results\n\n{header}{rows}{''.join(per_tier_lines)}",
encoding="utf-8",
)
log.info("comparison table updated: %s", table_path)
async def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--model", required=True, help="Model path or HF ID")
parser.add_argument("--tag", default="", help="Run label (e.g. grpo_v3, baseline_v2)")
parser.add_argument("--scenarios", nargs="*", help="Override scenario list")
parser.add_argument("--output", default="", help="Override output dir")
parser.add_argument("--adversarial", type=int, default=10,
help="Number of dynamic adversarial scenarios to generate (0 to skip)")
args = parser.parse_args()
os.environ["AGENT_MODEL"] = args.model
tag = args.tag or f"run-{int(time.time())}"
run_id = f"{tag}-{datetime.now(timezone.utc).strftime('%Y%m%d_%H%M%S')}"
out_dir = Path(args.output) if args.output else (RESULTS_DIR / run_id)
out_dir.mkdir(parents=True, exist_ok=True)
scenarios = list(args.scenarios or FROZEN_SCENARIOS)
# Generate fresh adversarial scenarios from 72B judge before running frozen set
if args.adversarial > 0:
log.info("generating %d dynamic adversarial scenarios via 72B judge...", args.adversarial)
# Seed with any existing failure history from prior runs
prior_failures = []
for d in RESULTS_DIR.iterdir():
ep_file = d / "results_per_episode.jsonl"
if ep_file.exists():
for line in ep_file.read_text().splitlines():
try:
ep = json.loads(line)
if not ep.get("resolved"):
prior_failures.append(ep)
except json.JSONDecodeError:
pass
adv_results = await design_batch(prior_failures, count=args.adversarial)
for adv in adv_results:
# Add generated manifest path as a runnable scenario
rel = str(Path(adv["manifest_path"]).relative_to(Path("bench/chaos_manifests")))
rel = rel.replace("\\", "/").removesuffix(".yaml")
scenarios.append(rel)
log.info("added %d adversarial scenarios to run", len(adv_results))
log.info("running %d scenarios for tag=%s model=%s", len(scenarios), tag, args.model)
results = []
episodes_file = out_dir / "results_per_episode.jsonl"
with episodes_file.open("w", encoding="utf-8") as f:
for i, s in enumerate(scenarios, 1):
log.info("[%d/%d] %s", i, len(scenarios), s)
r = await run_scenario(s)
results.append(r)
f.write(json.dumps(r) + "\n")
f.flush()
summary = compute_summary(results, tag, args.model)
(out_dir / "results_summary.json").write_text(json.dumps(summary, indent=2), encoding="utf-8")
write_comparison_table(summary)
log.info("=== Benchmark complete ===")
log.info(" Resolution rate : %.1f%%", summary["resolution_rate"] * 100)
log.info(" Avg reward : %.3f", summary["avg_reward"])
log.info(" Results : %s", out_dir)
if __name__ == "__main__":
asyncio.run(main())
|