|
| 1 | +""" |
| 2 | +Experiment: Find the parallelism level that overloads the server. |
| 3 | +
|
| 4 | +Strategy: |
| 5 | +- For each parallelism level (starting low, escalating), query ALL states |
| 6 | + at that level with sleep between each query. |
| 7 | +- Each state gets a unique time window (never reused across runs). |
| 8 | +- Log wall-clock time, outcome, and any signs of server stress (slow responses, |
| 9 | + errors, partial results). |
| 10 | +- Sleep between queries to stay below rate limit and isolate the effect of |
| 11 | + parallelism from quota exhaustion. |
| 12 | +
|
| 13 | +Usage: |
| 14 | + python experiments/overload_threshold_experiment.py [--level N] [--sleep 10] |
| 15 | +
|
| 16 | + If --level is not given, it auto-escalates through 1, 2, 4, 8, 16, 32. |
| 17 | +""" |
| 18 | + |
| 19 | +from __future__ import annotations |
| 20 | + |
| 21 | +import json |
| 22 | +import signal |
| 23 | +import sys |
| 24 | +import time |
| 25 | +from datetime import datetime, timezone |
| 26 | +from pathlib import Path |
| 27 | + |
| 28 | +sys.path.insert(0, str(Path(__file__).resolve().parents[1])) |
| 29 | + |
| 30 | +from dataretrieval import waterdata |
| 31 | +from dataretrieval.ogc.interruptions import ChunkInterrupted, QuotaExhausted |
| 32 | + |
| 33 | +LOG_FILE = Path(__file__).with_name("overload_threshold_log.jsonl") |
| 34 | + |
| 35 | +# Large states — these put real pressure on the server |
| 36 | +STATES = ["Ohio", "Pennsylvania", "Connecticut", "New Hampshire", "Vermont"] |
| 37 | + |
| 38 | +# Time windows pool — we rotate through these, never reusing a (state, window) |
| 39 | +# pair across the entire log. |
| 40 | +WINDOWS = [ |
| 41 | + "2000-01-01/2000-12-31", |
| 42 | + "2001-01-01/2001-12-31", |
| 43 | + "2002-01-01/2002-12-31", |
| 44 | + "2003-01-01/2003-12-31", |
| 45 | + "2004-01-01/2004-12-31", |
| 46 | + "2005-01-01/2005-12-31", |
| 47 | + "2006-01-01/2006-12-31", |
| 48 | + "2007-01-01/2007-12-31", |
| 49 | + "2008-01-01/2008-12-31", |
| 50 | + "2009-01-01/2009-12-31", |
| 51 | + "2010-01-01/2010-12-31", |
| 52 | + "2011-01-01/2011-12-31", |
| 53 | + "2012-01-01/2012-12-31", |
| 54 | + "2013-01-01/2013-12-31", |
| 55 | + "2014-01-01/2014-12-31", |
| 56 | + "2015-01-01/2015-12-31", |
| 57 | + "2016-01-01/2016-12-31", |
| 58 | + "2017-01-01/2017-12-31", |
| 59 | + "2018-01-01/2018-12-31", |
| 60 | + "2019-01-01/2019-12-31", |
| 61 | + "2020-01-01/2020-12-31", |
| 62 | + "2021-01-01/2021-12-31", |
| 63 | + "2022-01-01/2022-12-31", |
| 64 | + "2023-01-01/2023-12-31", |
| 65 | + "2024-01-01/2024-12-31", |
| 66 | +] |
| 67 | + |
| 68 | + |
| 69 | +class HangTimeout(Exception): |
| 70 | + pass |
| 71 | + |
| 72 | + |
| 73 | +def _timeout_handler(signum, frame): |
| 74 | + raise HangTimeout("hang") |
| 75 | + |
| 76 | + |
| 77 | +def _load_log() -> list[dict]: |
| 78 | + if not LOG_FILE.exists(): |
| 79 | + return [] |
| 80 | + entries = [] |
| 81 | + with open(LOG_FILE) as f: |
| 82 | + for line in f: |
| 83 | + line = line.strip() |
| 84 | + if line: |
| 85 | + entries.append(json.loads(line)) |
| 86 | + return entries |
| 87 | + |
| 88 | + |
| 89 | +def _append_log(entry: dict) -> None: |
| 90 | + with open(LOG_FILE, "a") as f: |
| 91 | + f.write(json.dumps(entry) + "\n") |
| 92 | + |
| 93 | + |
| 94 | +def _used_windows_for_state(entries: list[dict], state: str) -> set[str]: |
| 95 | + """Return windows already used for a given state in the log.""" |
| 96 | + return {e["time_window"] for e in entries if e.get("state") == state} |
| 97 | + |
| 98 | + |
| 99 | +def _next_window(state: str, used: set[str]) -> str | None: |
| 100 | + """Pick the next unused window for a state.""" |
| 101 | + for w in WINDOWS: |
| 102 | + if w not in used: |
| 103 | + return w |
| 104 | + return None |
| 105 | + |
| 106 | + |
| 107 | +def run_level(level: int, sleep_s: int, timeout_s: int) -> list[dict]: |
| 108 | + """Run all states at a given parallelism level. Returns the trial results.""" |
| 109 | + entries = _load_log() |
| 110 | + |
| 111 | + # Preload sites |
| 112 | + site_cache: dict[str, list[str]] = {} |
| 113 | + for state in STATES: |
| 114 | + print(f" Fetching {state} sites...", end=" ", flush=True) |
| 115 | + sites, _ = waterdata.get_monitoring_locations(state=state, site_type_code="ST") |
| 116 | + site_cache[state] = sites["monitoring_location_id"].tolist() |
| 117 | + print(f"{len(site_cache[state])} sites") |
| 118 | + |
| 119 | + print( |
| 120 | + f"\n Running all {len(STATES)} states at n={level}" |
| 121 | + f" (sleep={sleep_s}s between):\n" |
| 122 | + ) |
| 123 | + |
| 124 | + results = [] |
| 125 | + for i, state in enumerate(STATES): |
| 126 | + # Pick a fresh window |
| 127 | + used = _used_windows_for_state(entries + results, state) |
| 128 | + window = _next_window(state, used) |
| 129 | + if window is None: |
| 130 | + print(f" {state}: no fresh windows left, skipping") |
| 131 | + continue |
| 132 | + |
| 133 | + site_ids = site_cache[state] |
| 134 | + print( |
| 135 | + f" [{i + 1}/{len(STATES)}] {state:<15} ({len(site_ids):>4} sites) " |
| 136 | + f"{window} n={level}...", |
| 137 | + end=" ", |
| 138 | + flush=True, |
| 139 | + ) |
| 140 | + |
| 141 | + old = signal.signal(signal.SIGALRM, _timeout_handler) |
| 142 | + signal.alarm(timeout_s) |
| 143 | + t0 = time.perf_counter() |
| 144 | + |
| 145 | + entry = { |
| 146 | + "timestamp": datetime.now(timezone.utc).isoformat(), |
| 147 | + "state": state, |
| 148 | + "n_sites": len(site_ids), |
| 149 | + "time_window": window, |
| 150 | + "parallelism": level, |
| 151 | + "timeout_s": timeout_s, |
| 152 | + "sleep_s": sleep_s, |
| 153 | + "outcome": None, |
| 154 | + "wall_clock_s": None, |
| 155 | + "n_records": None, |
| 156 | + "error_type": None, |
| 157 | + "error_msg": None, |
| 158 | + } |
| 159 | + |
| 160 | + try: |
| 161 | + if level <= 1: |
| 162 | + df, md = waterdata.get_daily( |
| 163 | + monitoring_location_id=site_ids, |
| 164 | + parameter_code="00060", |
| 165 | + time=window, |
| 166 | + ) |
| 167 | + else: |
| 168 | + with waterdata.parallel_chunks(level): |
| 169 | + df, md = waterdata.get_daily( |
| 170 | + monitoring_location_id=site_ids, |
| 171 | + parameter_code="00060", |
| 172 | + time=window, |
| 173 | + ) |
| 174 | + |
| 175 | + elapsed = time.perf_counter() - t0 |
| 176 | + entry["outcome"] = "success" |
| 177 | + entry["wall_clock_s"] = round(elapsed, 2) |
| 178 | + entry["n_records"] = len(df) |
| 179 | + print(f"✓ {elapsed:.1f}s, {len(df):,} records") |
| 180 | + |
| 181 | + except HangTimeout: |
| 182 | + elapsed = time.perf_counter() - t0 |
| 183 | + entry["outcome"] = "timeout" |
| 184 | + entry["wall_clock_s"] = round(elapsed, 2) |
| 185 | + entry["error_type"] = "HangTimeout" |
| 186 | + entry["error_msg"] = f"No response within {timeout_s}s" |
| 187 | + print(f"✗ HANG after {elapsed:.0f}s") |
| 188 | + |
| 189 | + except QuotaExhausted as exc: |
| 190 | + elapsed = time.perf_counter() - t0 |
| 191 | + entry["outcome"] = "quota_exhausted" |
| 192 | + entry["wall_clock_s"] = round(elapsed, 2) |
| 193 | + entry["error_type"] = "QuotaExhausted" |
| 194 | + entry["error_msg"] = f"429, retry_after={exc.retry_after}" |
| 195 | + print(f"✗ 429 after {elapsed:.1f}s (retry_after={exc.retry_after})") |
| 196 | + |
| 197 | + except ChunkInterrupted as exc: |
| 198 | + elapsed = time.perf_counter() - t0 |
| 199 | + entry["outcome"] = "interrupted" |
| 200 | + entry["wall_clock_s"] = round(elapsed, 2) |
| 201 | + entry["error_type"] = type(exc).__name__ |
| 202 | + entry["error_msg"] = str(exc)[:200] |
| 203 | + print(f"✗ {type(exc).__name__} after {elapsed:.1f}s") |
| 204 | + |
| 205 | + except Exception as exc: |
| 206 | + elapsed = time.perf_counter() - t0 |
| 207 | + entry["outcome"] = "error" |
| 208 | + entry["wall_clock_s"] = round(elapsed, 2) |
| 209 | + entry["error_type"] = type(exc).__name__ |
| 210 | + entry["error_msg"] = str(exc)[:200] |
| 211 | + print(f"✗ {type(exc).__name__} after {elapsed:.1f}s: {str(exc)[:80]}") |
| 212 | + |
| 213 | + finally: |
| 214 | + signal.alarm(0) |
| 215 | + signal.signal(signal.SIGALRM, old) |
| 216 | + |
| 217 | + _append_log(entry) |
| 218 | + results.append(entry) |
| 219 | + |
| 220 | + # Sleep between queries to stay under rate limit |
| 221 | + if i < len(STATES) - 1: |
| 222 | + print(f" (sleeping {sleep_s}s...)", flush=True) |
| 223 | + time.sleep(sleep_s) |
| 224 | + |
| 225 | + return results |
| 226 | + |
| 227 | + |
| 228 | +def print_summary(): |
| 229 | + """Print summary grouped by parallelism level.""" |
| 230 | + entries = _load_log() |
| 231 | + if not entries: |
| 232 | + print("No log entries.") |
| 233 | + return |
| 234 | + |
| 235 | + from collections import defaultdict |
| 236 | + |
| 237 | + by_level = defaultdict(list) |
| 238 | + for e in entries: |
| 239 | + by_level[e["parallelism"]].append(e) |
| 240 | + |
| 241 | + print("\n" + "=" * 90) |
| 242 | + print("OVERLOAD THRESHOLD EXPERIMENT — all runs") |
| 243 | + print("=" * 90) |
| 244 | + print( |
| 245 | + f"{'level':>5} | {'trials':>6} | {'ok':>4} | {'429':>4} | " |
| 246 | + f"{'hang':>4} | {'err':>4} | {'avg_s':>7} | {'max_s':>7} | " |
| 247 | + f"{'avg_records':>11} | {'total_records':>13}" |
| 248 | + ) |
| 249 | + print("-" * 90) |
| 250 | + |
| 251 | + for level in sorted(by_level.keys()): |
| 252 | + trials = by_level[level] |
| 253 | + ok = [t for t in trials if t["outcome"] == "success"] |
| 254 | + quota = [t for t in trials if t["outcome"] == "quota_exhausted"] |
| 255 | + hangs = [t for t in trials if t["outcome"] == "timeout"] |
| 256 | + errs = [t for t in trials if t["outcome"] in ("error", "interrupted")] |
| 257 | + |
| 258 | + times = [t["wall_clock_s"] for t in ok] |
| 259 | + avg_t = f"{sum(times) / len(times):.1f}" if times else "—" |
| 260 | + max_t = f"{max(times):.1f}" if times else "—" |
| 261 | + avg_r = int(sum(t["n_records"] for t in ok) / len(ok)) if ok else 0 |
| 262 | + total_r = sum(t["n_records"] for t in ok) if ok else 0 |
| 263 | + |
| 264 | + print( |
| 265 | + f"{level:>5} | {len(trials):>6} | {len(ok):>4} | {len(quota):>4} | " |
| 266 | + f"{len(hangs):>4} | {len(errs):>4} | {avg_t:>7} | {max_t:>7} | " |
| 267 | + f"{avg_r:>11,} | {total_r:>13,}" |
| 268 | + ) |
| 269 | + |
| 270 | + print("=" * 90) |
| 271 | + |
| 272 | + # Detail per level+state |
| 273 | + print("\nPer-state breakdown (successful trials only):") |
| 274 | + print( |
| 275 | + f" {'level':>5} | {'state':<15} | {'time_s':>7} | {'records':>10} | {'window'}" |
| 276 | + ) |
| 277 | + print(" " + "-" * 70) |
| 278 | + for level in sorted(by_level.keys()): |
| 279 | + for e in sorted(by_level[level], key=lambda x: x.get("state", "")): |
| 280 | + if e["outcome"] == "success": |
| 281 | + print( |
| 282 | + f" {level:>5} | {e['state']:<15} | " |
| 283 | + f"{e['wall_clock_s']:>7.1f} | {e['n_records']:>10,} | " |
| 284 | + f"{e['time_window']}" |
| 285 | + ) |
| 286 | + |
| 287 | + |
| 288 | +if __name__ == "__main__": |
| 289 | + import argparse |
| 290 | + |
| 291 | + parser = argparse.ArgumentParser() |
| 292 | + parser.add_argument("--level", type=int, default=None, help="Single level to test") |
| 293 | + parser.add_argument( |
| 294 | + "--levels", type=str, default=None, help="Comma-separated levels" |
| 295 | + ) |
| 296 | + parser.add_argument( |
| 297 | + "--sleep", type=int, default=10, help="Seconds between queries (default: 10)" |
| 298 | + ) |
| 299 | + parser.add_argument( |
| 300 | + "--timeout", type=int, default=120, help="Per-query timeout (default: 120)" |
| 301 | + ) |
| 302 | + parser.add_argument("--summary", action="store_true", help="Print summary only") |
| 303 | + args = parser.parse_args() |
| 304 | + |
| 305 | + if args.summary: |
| 306 | + print_summary() |
| 307 | + sys.exit(0) |
| 308 | + |
| 309 | + if args.level is not None: |
| 310 | + levels = [args.level] |
| 311 | + elif args.levels is not None: |
| 312 | + levels = [int(x) for x in args.levels.split(",")] |
| 313 | + else: |
| 314 | + levels = [1, 2, 4, 8, 16, 32] |
| 315 | + |
| 316 | + print("Overload threshold experiment") |
| 317 | + print(f" Levels to test: {levels}") |
| 318 | + print(f" States: {STATES}") |
| 319 | + print(f" Sleep between queries: {args.sleep}s") |
| 320 | + print(f" Timeout: {args.timeout}s") |
| 321 | + print(f" Log: {LOG_FILE}") |
| 322 | + print() |
| 323 | + |
| 324 | + for level in levels: |
| 325 | + print(f"{'=' * 60}") |
| 326 | + print(f" LEVEL {level}") |
| 327 | + print(f"{'=' * 60}") |
| 328 | + results = run_level(level, args.sleep, args.timeout) |
| 329 | + |
| 330 | + # Check if we got throttled — if so, stop escalating |
| 331 | + failures = [r for r in results if r["outcome"] != "success"] |
| 332 | + if failures: |
| 333 | + print(f"\n ⚠ Failures at level {level} — stopping escalation.") |
| 334 | + break |
| 335 | + |
| 336 | + print() |
| 337 | + |
| 338 | + print_summary() |
0 commit comments