|
| 1 | +#!/usr/bin/env python3 |
| 2 | +"""Verify that every message in a subnet-blob-msg run was delivered to all nodes.""" |
| 3 | + |
| 4 | +from __future__ import annotations |
| 5 | + |
| 6 | +import argparse |
| 7 | +import json |
| 8 | +import sys |
| 9 | +from collections import defaultdict |
| 10 | +from pathlib import Path |
| 11 | + |
| 12 | + |
| 13 | +def parse_args() -> argparse.Namespace: |
| 14 | + parser = argparse.ArgumentParser( |
| 15 | + description=( |
| 16 | + "Validate that every message in a subnet-blob-msg Shadow run " |
| 17 | + "was delivered to all nodes." |
| 18 | + ) |
| 19 | + ) |
| 20 | + parser.add_argument( |
| 21 | + "shadow_output", |
| 22 | + help="Path to the Shadow output directory (the one containing the hosts/ folder).", |
| 23 | + ) |
| 24 | + parser.add_argument( |
| 25 | + "--min-reach", |
| 26 | + type=float, |
| 27 | + default=1.0, |
| 28 | + help="Minimum fraction of non-publisher nodes that must receive each message (default: 1.0).", |
| 29 | + ) |
| 30 | + parser.add_argument( |
| 31 | + "--skip", |
| 32 | + type=int, |
| 33 | + default=4, |
| 34 | + help="Number of initial (warmup) messages to skip when checking reach (default: 4).", |
| 35 | + ) |
| 36 | + return parser.parse_args() |
| 37 | + |
| 38 | + |
| 39 | +def iter_stdout_logs(hosts_dir: Path): |
| 40 | + """Yield all stdout log files under the given hosts directory.""" |
| 41 | + for stdout_file in sorted(hosts_dir.rglob("*.stdout")): |
| 42 | + if stdout_file.is_file(): |
| 43 | + yield stdout_file |
| 44 | + |
| 45 | + |
| 46 | +def parse_logs(hosts_dir: Path): |
| 47 | + """Parse all stdout logs and return per-message delivery sets and total node count. |
| 48 | +
|
| 49 | + Returns: |
| 50 | + (message_deliveries, node_count) where message_deliveries maps |
| 51 | + message_id -> set of node_ids that received it, ordered by first |
| 52 | + delivery time across nodes. |
| 53 | + """ |
| 54 | + # message_id -> set of node_ids |
| 55 | + deliveries: dict[str, set[str]] = defaultdict(set) |
| 56 | + # message_id -> earliest timestamp string (for ordering) |
| 57 | + first_seen: dict[str, str] = {} |
| 58 | + node_ids: set[str] = set() |
| 59 | + |
| 60 | + for log_path in iter_stdout_logs(hosts_dir): |
| 61 | + node_name = log_path.parent.name # e.g. "node0" |
| 62 | + current_node_id: str | None = None |
| 63 | + |
| 64 | + with log_path.open("r", encoding="utf-8", errors="replace") as fh: |
| 65 | + for line in fh: |
| 66 | + try: |
| 67 | + entry = json.loads(line) |
| 68 | + except (json.JSONDecodeError, ValueError): |
| 69 | + continue |
| 70 | + |
| 71 | + msg = entry.get("msg") |
| 72 | + if msg == "PeerID": |
| 73 | + current_node_id = str(entry.get("node_id", node_name)) |
| 74 | + node_ids.add(current_node_id) |
| 75 | + elif msg == "Received Message": |
| 76 | + mid = entry.get("id", "") |
| 77 | + if not mid: |
| 78 | + continue |
| 79 | + nid = current_node_id or node_name |
| 80 | + node_ids.add(nid) |
| 81 | + deliveries[mid].add(nid) |
| 82 | + ts = entry.get("time", "") |
| 83 | + if mid not in first_seen or ts < first_seen[mid]: |
| 84 | + first_seen[mid] = ts |
| 85 | + |
| 86 | + # Order messages by first delivery time |
| 87 | + ordered_ids = sorted(deliveries.keys(), key=lambda m: first_seen.get(m, "")) |
| 88 | + return deliveries, ordered_ids, len(node_ids) |
| 89 | + |
| 90 | + |
| 91 | +def main() -> int: |
| 92 | + args = parse_args() |
| 93 | + base_dir = Path(args.shadow_output).expanduser().resolve() |
| 94 | + if not base_dir.exists(): |
| 95 | + print(f"shadow output directory does not exist: {base_dir}", file=sys.stderr) |
| 96 | + return 1 |
| 97 | + |
| 98 | + hosts_dir = base_dir / "hosts" |
| 99 | + if not hosts_dir.is_dir(): |
| 100 | + print(f"hosts directory not found under: {base_dir}", file=sys.stderr) |
| 101 | + return 1 |
| 102 | + |
| 103 | + deliveries, ordered_ids, node_count = parse_logs(hosts_dir) |
| 104 | + |
| 105 | + if not ordered_ids: |
| 106 | + print("no messages found in logs", file=sys.stderr) |
| 107 | + return 1 |
| 108 | + |
| 109 | + if node_count == 0: |
| 110 | + print("no nodes found in logs", file=sys.stderr) |
| 111 | + return 1 |
| 112 | + |
| 113 | + # Skip warmup messages |
| 114 | + check_ids = ordered_ids[args.skip:] |
| 115 | + if not check_ids: |
| 116 | + print( |
| 117 | + f"no messages left after skipping {args.skip} warmup messages " |
| 118 | + f"(total messages: {len(ordered_ids)})", |
| 119 | + file=sys.stderr, |
| 120 | + ) |
| 121 | + return 1 |
| 122 | + |
| 123 | + # For each message, one node is the publisher so max receivers = node_count - 1 |
| 124 | + expected_receivers = node_count - 1 |
| 125 | + failures: list[tuple[str, int, float]] = [] |
| 126 | + |
| 127 | + for mid in check_ids: |
| 128 | + receivers = len(deliveries[mid]) |
| 129 | + # The publisher also appears in "Received Message" sometimes, so cap at node_count |
| 130 | + reach = min(receivers, expected_receivers) / expected_receivers if expected_receivers > 0 else 0.0 |
| 131 | + if reach < args.min_reach: |
| 132 | + failures.append((mid, receivers, reach)) |
| 133 | + |
| 134 | + print(f"Nodes: {node_count}") |
| 135 | + print(f"Total messages: {len(ordered_ids)} (skipped {args.skip} warmup)") |
| 136 | + print(f"Checked messages: {len(check_ids)}") |
| 137 | + print(f"Required reach: {args.min_reach:.0%}") |
| 138 | + print() |
| 139 | + |
| 140 | + for mid in check_ids: |
| 141 | + receivers = len(deliveries[mid]) |
| 142 | + reach = min(receivers, expected_receivers) / expected_receivers if expected_receivers > 0 else 0.0 |
| 143 | + status = "OK" if reach >= args.min_reach else "FAIL" |
| 144 | + print(f" [{status}] {mid}: {receivers}/{expected_receivers} nodes ({reach:.0%})") |
| 145 | + |
| 146 | + print() |
| 147 | + if failures: |
| 148 | + print( |
| 149 | + f"FAILED: {len(failures)}/{len(check_ids)} messages did not reach " |
| 150 | + f"{args.min_reach:.0%} of nodes.", |
| 151 | + file=sys.stderr, |
| 152 | + ) |
| 153 | + return 1 |
| 154 | + |
| 155 | + print(f"PASSED: all {len(check_ids)} messages reached {args.min_reach:.0%} of nodes.") |
| 156 | + return 0 |
| 157 | + |
| 158 | + |
| 159 | +if __name__ == "__main__": |
| 160 | + sys.exit(main()) |
0 commit comments