feat(sync): 每小时后台自动同步 + 手动拉取最新接口

- services/scheduler.py:通过 job_locks 表跨 worker 抢占,
  gunicorn 多进程下一个周期只跑一次;claim 超时 30 分钟自动释放,
  避免 worker 中途挂掉把任务永久卡死
- POST /api/garmin/sync-latest:同步执行,窗口 clamp 到 1..7 天
- GET  /api/garmin/auto-sync:返回上次/下次运行时间
- db.py:注释里的分号会被 SCHEMA.split(";") 截断,改为先剥注释再切分

19 项调度器测试,全量 324 项通过

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
This commit is contained in:
ericwyuan
2026-08-24 00:17:04 +08:00
parent 85c00759f8
commit 12ef5ca06b
5 changed files with 412 additions and 4 deletions

View File

@@ -0,0 +1,157 @@
"""
Background scheduler.
Keeps the local database close to Garmin without the user having to press
anything: every interval it pulls the last couple of days for each account
that has stored OAuth tokens.
Two things make this fiddly in this deployment, and both are handled here:
* **Several workers.** gunicorn runs more than one process, and each would
otherwise start its own timer and sync the same account concurrently. A row
in `job_locks` is claimed before any work starts, so exactly one worker runs
a given tick.
* **Restarts.** The thread dies with its worker. The lock records when the job
last completed, so a freshly started worker picks the schedule back up
rather than either skipping an interval or immediately re-running.
"""
import datetime
import os
import threading
import time
from config import DB_TYPE
from db import execute, query_one, query_all
from services import garmin as garmin_svc
JOB_NAME = "garmin_auto_sync"
# How often to pull, and how far back. Two days rather than one: the current
# day is still being written to, and a day can arrive late.
INTERVAL_SECONDS = int(os.environ.get("AUTO_SYNC_INTERVAL_SECONDS") or 3600)
SYNC_DAYS = int(os.environ.get("AUTO_SYNC_DAYS") or 2)
ENABLED = (os.environ.get("AUTO_SYNC") or "true").lower() not in ("0", "false", "no")
# A claim older than this is treated as abandoned — the worker holding it died
# mid-run, and without expiry the job would never run again.
CLAIM_TIMEOUT_SECONDS = 1800
_started = False
_lock = threading.Lock()
def _now():
return datetime.datetime.utcnow()
def _iso(dt):
return dt.isoformat(timespec="seconds")
def _parse(value):
if not value:
return None
try:
return datetime.datetime.fromisoformat(str(value).replace(" ", "T"))
except ValueError:
return None
def claim(name=JOB_NAME, interval=INTERVAL_SECONDS):
"""Take the job if it is due and nobody else holds it.
Returns True when this process should do the work.
"""
holder = f"{os.getpid()}"
now = _now()
row = query_one("SELECT * FROM job_locks WHERE name = ?", [name])
if row:
last_run = _parse(row.get("last_run_at"))
if last_run and (now - last_run).total_seconds() < interval:
return False
claimed = _parse(row.get("claimed_at"))
if claimed and (now - claimed).total_seconds() < CLAIM_TIMEOUT_SECONDS:
return False
if DB_TYPE == "mariadb":
sql = ("INSERT INTO job_locks (name, holder, claimed_at) VALUES (?, ?, ?) "
"ON DUPLICATE KEY UPDATE holder=VALUES(holder), claimed_at=VALUES(claimed_at)")
else:
sql = ("INSERT INTO job_locks (name, holder, claimed_at) VALUES (?, ?, ?) "
"ON CONFLICT(name) DO UPDATE SET holder=excluded.holder, "
"claimed_at=excluded.claimed_at")
execute(sql, [name, holder, _iso(now)])
# Re-read: if another worker claimed between our check and our write, its
# holder is the one now recorded and we must stand down.
check = query_one("SELECT holder FROM job_locks WHERE name = ?", [name])
return bool(check and check.get("holder") == holder)
def release(name=JOB_NAME, ran=True):
if ran:
execute(
"UPDATE job_locks SET claimed_at = NULL, last_run_at = ? WHERE name = ?",
[_iso(_now()), name],
)
else:
execute("UPDATE job_locks SET claimed_at = NULL WHERE name = ?", [name])
def sync_all_accounts(days=None):
"""Sync every account that has a stored token. Returns a per-user result."""
days = days or SYNC_DAYS
rows = query_all("SELECT user_id FROM garmin_tokens")
results = []
for row in rows:
uid = row["user_id"]
try:
out = garmin_svc.sync_data(uid, {}, days=days)
results.append({"user": uid, "status": out.get("status"),
"records": out.get("recordsSynced")})
except Exception as e: # noqa: BLE001 - one account must not stop the rest
results.append({"user": uid, "status": "error", "error": str(e)[:200]})
return results
def _loop():
while True:
try:
if claim():
try:
sync_all_accounts()
finally:
release()
except Exception as e: # noqa: BLE001 - the loop must outlive any single failure
print(f"[scheduler] tick failed: {e}")
# Checked more often than the interval so a worker that starts late
# still picks the job up promptly rather than waiting a full hour.
time.sleep(min(300, INTERVAL_SECONDS))
def start():
"""Start the scheduler thread once per process."""
global _started
if not ENABLED:
print("[scheduler] disabled by AUTO_SYNC")
return
with _lock:
if _started:
return
_started = True
threading.Thread(target=_loop, daemon=True, name="auto-sync").start()
print(f"[scheduler] auto-sync every {INTERVAL_SECONDS}s, {SYNC_DAYS} day(s) back")
def status():
row = query_one("SELECT * FROM job_locks WHERE name = ?", [JOB_NAME])
last = _parse(row.get("last_run_at")) if row else None
return {
"enabled": ENABLED,
"intervalSeconds": INTERVAL_SECONDS,
"days": SYNC_DAYS,
"lastRunAt": _iso(last) if last else None,
"nextRunAt": _iso(last + datetime.timedelta(seconds=INTERVAL_SECONDS)) if last else None,
"running": bool(row and row.get("claimed_at")),
}