Files
mtusi2026lab3-fork/scripts/failover_all.py
2026-11-04 17:20:59 +03:00

149 lines
6.4 KiB
Python

"""Сценарий поочерёдного отказа всех API-реплик под записью."""
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
import json
import os
import re
import subprocess
from threading import Event, Lock
import time
from urllib.parse import quote
from uuid import uuid4
from scripts.deployment import ComposeProject, ROOT, deployment_lock
from tests.client import Client, Failure, arguments, check, run
def failure_kind(error: Failure) -> str:
"""Различает HTTP-отказы, ошибки соединения и проверки контракта."""
match = re.search(r"received (\d{3})$", str(error))
if match:
return "http_" + match.group(1)
return "connection" if "connection/TLS failure" in str(error) else "contract"
def snapshot(project: ComposeProject, phase: str, target: str) -> list:
"""Печатает состояние API без окружения и секретов."""
replicas = project.replicas()
print(json.dumps({
"scenario": "replica_failure", "phase": phase, "target": target,
"replicas": [{"instance": item.instance_id, "running": item.running, "health": item.health} for item in replicas],
}), flush=True)
return replicas
def history(client: Client, actor: dict, chat: dict) -> list[dict]:
"""Читает всю историю без повторов курсора."""
path = f"/api/v1/chats/{chat['id']}/messages?limit=100"
items = []
cursors = set()
for _ in range(1000):
page = client.call("GET", path, actor=actor, schema="MessagePage")["json"]
items.extend(page["items"])
cursor = page["next_cursor"]
if cursor is None:
return items
check(cursor not in cursors, "cursor cycle")
cursors.add(cursor)
path = f"/api/v1/chats/{chat['id']}/messages?limit=100&cursor={quote(cursor, safe='')}"
raise Failure("history exceeded safety bound")
def exercise(project: ComposeProject, args) -> None:
"""Останавливает каждую реплику и восстанавливает её перед следующей."""
replicas = project.replicas()
check(len(replicas) >= 2, "at least two API replicas are required")
project.wait_healthy({item.id for item in replicas})
client = Client(args.base_url, args.ca_file, timeout=2)
for item in replicas:
project.wait_routed(item.instance_id, client.base, args.ca_file)
actor, _, _, chat = client.bootstrap()
confirmed = set()
for target in replicas:
snapshot(project, "before", target.instance_id)
stop = Event()
lock = Lock()
successes = []
errors = []
def worker():
local = Client(client.base, args.ca_file, timeout=2)
key = uuid4().hex
while not stop.is_set():
started = time.monotonic()
try:
result = local.message(actor, chat, "failover-" + key, key=key)
with lock:
successes.append((started, time.monotonic(), result["json"]["id"]))
key = uuid4().hex
except Failure as error:
with lock:
errors.append(failure_kind(error))
time.sleep(0.1)
started = time.monotonic()
with ThreadPoolExecutor(max_workers=args.concurrency) as pool:
jobs = [pool.submit(worker) for _ in range(args.concurrency)]
try:
time.sleep(2)
stop_started = time.monotonic()
subprocess.run(
[str(ROOT / "scripts/stop-one.sh")], cwd=ROOT,
env={**os.environ, "TARGET_INSTANCE_ID": target.instance_id},
check=True, timeout=10, capture_output=True,
)
stop_finished = time.monotonic()
stopped = snapshot(project, "stopped", target.instance_id)
check(any(item.id == target.id and not item.running for item in stopped), "target API was not stopped")
check(any(item.id != target.id and item.running and item.health == "healthy" for item in stopped), "no healthy surviving API")
time.sleep(args.duration)
finally:
stop.set()
try:
for job in jobs:
job.result(timeout=5)
finally:
project.docker("start", target.id)
project.wait_healthy({item.id for item in replicas})
project.wait_routed(target.instance_id, client.base, args.ca_file)
snapshot(project, "restored", target.instance_id)
elapsed = max(end for _, end, _ in successes) - started if successes else 0
recovered = [end for begin, end, _ in successes if begin >= stop_finished]
check(bool(recovered), "no confirmed writes after failure")
recovery = min(recovered) - stop_started
check(recovery <= 10, "writes did not recover within 10 seconds")
confirmed.update(id for _, _, id in successes)
items = history(client, actor, chat)
check(confirmed <= {item["id"] for item in items}, "confirmed messages were lost")
texts = [item["text"] for item in items]
check(len(texts) == len(set(texts)), "uncertain retry created duplicates")
print(json.dumps({
"instance": target.instance_id,
"successful_writes": len(successes),
"failed_attempts": len(errors),
"errors_by_kind": dict(Counter(errors)),
"elapsed_seconds": round(elapsed, 3),
"successful_rps": round(len(successes) / elapsed, 3) if elapsed else 0,
"first_recovery_seconds": round(recovery, 3),
}), flush=True)
def main() -> None:
"""Готовит параметры эксперимента отказа реплик."""
def configure(parser):
parser.add_argument("--duration", type=float, default=15)
parser.add_argument("--concurrency", type=int, default=2)
args = arguments("Fail each API replica under write load", configure)
check(1 <= args.duration <= 120 and 1 <= args.concurrency <= 8, "invalid workload")
project = ComposeProject()
with deployment_lock(project):
exercise(project, args)
if __name__ == "__main__":
run(main)