bench: трек 1 — оркестратор серий, k6 open-loop, py-spy sidecar

- bench/track1/run.py: на каждую конфигурацию x повтор — свежий стенд
  (отдельный compose-проект, пустые volumes), образ API из git-коммита
  (git archive), ступени constant-arrival-rate; CPU каждого контейнера,
  ожидание пула (STATS_LOG_SECONDS), pg_stat_database/pg_stat_statements,
  счётчики HAProxy; raw JSON + results.csv, без токенов.
- compose.bench.yaml: cpuset для k6 / LB+Valkey / PostgreSQL / API, чтобы
  генератор и LB не отнимали CPU у реплик.
- bench/track1/profile.py + pyspy/: py-spy с хешем пакета, профиль одной
  реплики под нагрузкой.
- API: STATS_LOG_SECONDS (по умолчанию выключено) — счётчики пула и CPU
  процесса в лог; API_IMAGE в compose для образа из коммита.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Alexey Bodanyuk
2026-10-16 13:51:37 +03:00
co-authored by Claude Opus 5.5
parent c050642ec2
commit 2033440cdb
9 changed files with 628 additions and 4 deletions
+21 -2
View File
@@ -338,6 +338,22 @@ def create_app(settings: Settings | None = None) -> Starlette:
except ApiError:
pass # database unavailable right now; next round
stats_log = logging.getLogger("messenger.stats")
async def stats_loop(every: float) -> None:
"""Experiments only: DB pool saturation (waiting requests, wait time) and process CPU."""
cpu = time.process_time()
db.pool.pop_stats()
while True:
await asyncio.sleep(every)
stats = db.pool.pop_stats()
now_cpu = time.process_time()
stats_log.info("pool_stats", extra={"fields": {
"interval_s": every, "cpu_s": round(now_cpu - cpu, 4),
**{k: stats.get(k, 0) for k in ("pool_size", "pool_available", "requests_waiting", "requests_num",
"requests_queued", "requests_wait_ms", "usage_ms", "requests_errors")}}})
cpu = now_cpu
@contextlib.asynccontextmanager
async def lifespan(app: Starlette):
passwords.configure(settings.hash_concurrency)
@@ -346,11 +362,14 @@ def create_app(settings: Settings | None = None) -> Starlette:
"session_store": sessions.name, "session_ttl_s": settings.session_ttl_seconds,
"db_pool_max": settings.db_pool_max, "db_timeout_s": settings.db_timeout_seconds,
"instance_id": settings.instance_id, "drain_s": settings.drain_seconds}})
purge = asyncio.create_task(purge_loop())
tasks = [asyncio.create_task(purge_loop())]
if settings.stats_log_seconds > 0:
tasks.append(asyncio.create_task(stats_loop(settings.stats_log_seconds)))
try:
yield
finally:
purge.cancel()
for task in tasks:
task.cancel()
await sessions.close()
await db.close()
+3
View File
@@ -106,6 +106,7 @@ class Settings:
instance_id: str = "local"
drain_seconds: float = 0.0
idempotency_ttl_seconds: int = 86400
stats_log_seconds: float = 0.0
@property
def db_statement_timeout_ms(self) -> int:
@@ -151,4 +152,6 @@ class Settings:
drain_seconds=_float("DRAIN_SECONDS", 0.0, 0.0, 10.0),
# Contract: a replay within 24 h returns the original result.
idempotency_ttl_seconds=_int("IDEMPOTENCY_TTL_SECONDS", 86400, 86400, 30 * 24 * 3600),
# Experiments (bench/track1): log DB pool counters and process CPU every N s; 0 = off.
stats_log_seconds=_float("STATS_LOG_SECONDS", 0.0, 0.0, 60.0),
)