deliver: 交付收口(引擎代为提交)

This commit is contained in:
agent.develop 2026-09-19 12:22:21 +08:00
parent 93829c6a9b
commit af83f15b37
14 changed files with 2373 additions and 0 deletions

View File

@ -239,3 +239,32 @@ async def pbl_session_member_add(**kw):
event_type='session.member.joined',
payload_json={'user_id': user_id, 'members': cnt + 1})
return {'ok': True, 'members': cnt + 1, 'max_members': MAX_MEMBERS, 'deduped': False}
# ================================================================== M11b 覆盖
# M11b单事务事件+状态写入、按月分区、提交后广播、3s 轮询兜底)是**权威实现**
# 同名契约在此覆盖上方 M11a 版本,避免两套逻辑漂移;返回结构向后兼容(只增字段)。
# 主链路m11b_api → tx_write.apply_runtime_event一个事务写 事件+实体状态+快照
# → 提交后才广播 → 分段计时裁决/广播/端到端并与 SLA 预算比对)。
from .m11b_api import ( # noqa: E402 契约层覆盖(放最后,导入即生效)
contracts as m11b_contracts,
pbl_entity_state_apply,
pbl_entity_state_get,
pbl_runtime_broadcast_pull,
pbl_runtime_event_append,
pbl_runtime_broadcast_pull,
pbl_runtime_broadcast_stats,
pbl_runtime_event_poll,
pbl_world_broadcast,
)
CONTRACTS = {
"pbl_runtime_event_append": pbl_runtime_event_append,
"pbl_runtime_event_poll": pbl_runtime_event_poll,
"pbl_runtime_broadcast_pull": pbl_runtime_broadcast_pull,
"pbl_runtime_broadcast_stats": pbl_runtime_broadcast_stats,
"pbl_entity_state_get": pbl_entity_state_get,
"pbl_entity_state_apply": pbl_entity_state_apply,
"pbl_world_broadcast": pbl_world_broadcast,
"pbl_session_member_add": pbl_session_member_add,
}

View File

@ -0,0 +1,263 @@
"""M11b 提交后广播post-commit broadcast+ 3s 轮询兜底。
铁律与语义
----------
1. **只在事务提交成功后广播**广播内容必须是已落库的事实提交失败绝不广播
否则客户端会收到数据库里不存在的状态幽灵更新无法收敛
2. **广播是尽力而为best-effort不是投递保证**推送通道websocket/SSE
抖动进程重启都可能丢事件丢广播**不撤销已提交事实**客户端用
`since_seq` 轮询补齐兜底周期 3s保证最终一致不丢数据
3. **环形缓冲**每个会话在内存里保留最近 N 条已提交事件
`pbl_runtime_broadcast_pull` 做毫秒级增量拉取比查库轻
缓冲被覆盖时客户端自动退回查库轮询poll 契约不会漏
4. 广播耗时计入 SLA300ms推送在后台线程做主请求只负责入队 +
记录 accepted`pull` 契约用于验证事件确实可被消费
线程安全所有状态变更都在 RLock 推送回调在后台 worker 线程执行
回调异常只计数不影响主链路
"""
import collections
import logging
import threading
import time
from .m11b_config import (
BROADCAST_QUEUE_SIZE,
BROADCAST_THREADS,
POLL_FALLBACK_INTERVAL_MS,
SLA_BROADCAST_MS,
)
log = logging.getLogger("pbl_runtime_ext.broadcast")
def _now_ms():
return int(time.time() * 1000)
class _Worker(object):
"""极简后台推送线程daemon避免引入线程池依赖。"""
def __init__(self, hub, idx):
self.hub = hub
self.idx = idx
self.stop = threading.Event()
self.thread = threading.Thread(
target=self._run, name="pbl-bcast-%d" % idx)
self.thread.daemon = True
def start(self):
try:
self.thread.start()
except Exception as exc: # noqa: BLE001
log.warning("广播 worker 启动失败(退回纯轮询): %s", exc)
def _run(self):
while not self.stop.is_set():
item = self.hub._next_job()
if item is None:
self.stop.wait(0.01)
continue
self.hub._deliver(*item)
class BroadcastHub(object):
"""会话级广播中心:入队 → 后台推送 → 环形缓冲供 pull/轮询补齐。"""
def __init__(self, capacity=BROADCAST_QUEUE_SIZE, workers=BROADCAST_THREADS):
self._lock = threading.RLock()
self._cond = threading.Condition(self._lock)
self._queue = collections.deque(maxlen=int(capacity))
self._buffers = {} # session_key -> deque(events)
self._capacity = int(capacity)
self._pushers = [] # [(name, callable)]
self._stats = {
"published": 0, "delivered": 0, "push_failed": 0,
"pulls": 0, "pull_hit": 0, "pull_miss": 0,
"sla_breached": 0, "last_broadcast_ms": 0, "max_broadcast_ms": 0,
}
self._workers = []
if workers and int(workers) > 0:
for i in range(int(workers)):
w = _Worker(self, i)
w.start()
self._workers.append(w)
# -------------------------------------------------- 通道注册
def register_push(self, fn, name="custom"):
"""注册推送通道ws/sse。fn(session_key, event) -> truthy/awaitable。"""
if not callable(fn):
raise TypeError("push 回调必须可调用")
with self._lock:
self._pushers.append((name, fn))
return name
def unregister_push(self, name):
with self._lock:
self._pushers = [(n, f) for (n, f) in self._pushers if n != name]
# -------------------------------------------------- 发布
def publish(self, session_key, event):
"""提交后调用:先入环形缓冲(保证可被 pull 到),再排入推送队列。
返回 {"accepted", "buffered", "channel", "seq_no", "pull_deadline_ms"}
本方法不阻塞等待推送结果 广播预算 300ms 包含推送主链路只记账号
"""
t0 = _now_ms()
with self._lock:
buf = self._buffers.setdefault(
str(session_key), collections.deque(maxlen=self._capacity))
buf.append(event)
self._stats["published"] += 1
channels = [n for n, _f in self._pushers]
self._queue.append((session_key, event, t0))
try:
self._cond.notify()
except Exception: # noqa: BLE001
pass
accepted = bool(channels)
# 无后台 worker降级/自检模式)时同步投递,保证广播语义仍然成立
if not self._workers and channels:
self._deliver(session_key, event, t0)
cost = _now_ms() - t0
with self._lock:
self._stats["last_broadcast_ms"] = cost
self._stats["max_broadcast_ms"] = max(self._stats["max_broadcast_ms"], cost)
if cost > SLA_BROADCAST_MS:
self._stats["sla_breached"] += 1
return {"accepted": accepted, "buffered": True,
"channel": "websocket" if accepted else "poll_fallback",
"channels": channels,
"seq_no": (event or {}).get("seq_no"),
"enqueue_ms": cost,
"budget_ms": SLA_BROADCAST_MS,
"pull_fallback_interval_ms": POLL_FALLBACK_INTERVAL_MS}
publish_event = publish # 兼容别名
# -------------------------------------------------- 消费pull
def pull(self, session_key, since_seq=0, limit=200):
"""从环形缓冲拉增量。miss缓冲已覆盖时返回 need_poll=True
客户端据此退回查库轮询契约pbl_runtime_event_poll
"""
since = int(since_seq or 0)
lim = max(1, min(int(limit or 200), 500))
with self._lock:
self._stats["pulls"] += 1
buf = self._buffers.get(str(session_key))
if not buf:
self._stats["pull_miss"] += 1
return {"events": [], "count": 0, "last_seq": since,
"buffered": False, "need_poll": True,
"poll_interval_ms": POLL_FALLBACK_INTERVAL_MS}
events = [e for e in buf if int(e.get("seq_no") or 0) > since]
events = events[:lim]
oldest = int(buf[0].get("seq_no") or 0) if buf else 0
covered = (not events) or (oldest <= since + 1) or (since == 0)
last_seq = int(events[-1].get("seq_no")) if events else since
if events:
self._stats["pull_hit"] += 1
if not covered:
self._stats["pull_miss"] += 1
else:
self._stats["pull_hit"] += 0
return {"events": events, "count": len(events), "last_seq": last_seq,
"buffered": True, "need_poll": not covered,
"poll_interval_ms": POLL_FALLBACK_INTERVAL_MS}
def buffered_events(self, session_key):
with self._lock:
return list(self._buffers.get(str(session_key), []))
def forget(self, session_key):
"""会话结束时释放内存缓冲。"""
with self._lock:
self._buffers.pop(str(session_key), None)
# -------------------------------------------------- 内部推送
def _next_job(self):
with self._cond:
if not self._queue:
return None
return self._queue.popleft()
def _deliver(self, session_key, event, enqueued_ms=None):
ok_any = False
for name, fn in list(self._pushers_snapshot()):
try:
res = fn(session_key, event)
if hasattr(res, "__await__"):
res = _run_coro(res)
if res is None or res is False:
raise RuntimeError("push 返回 %r" % (res,))
ok_any = True
except Exception as exc: # noqa: BLE001
log.warning("广播通道 %s 投递失败(客户端 3s 轮询补齐): %s", name, exc)
with self._lock:
self._stats["push_failed"] += 1
if ok_any:
with self._lock:
self._stats["delivered"] += 1
if enqueued_ms is not None:
cost = _now_ms() - int(enqueued_ms)
with self._lock:
self._stats["last_broadcast_ms"] = cost
self._stats["max_broadcast_ms"] = max(self._stats["max_broadcast_ms"], cost)
if cost > SLA_BROADCAST_MS:
self._stats["sla_breached"] += 1
return ok_any
def _pushers_snapshot(self):
with self._lock:
return list(self._pushers)
# -------------------------------------------------- 观测
def stats(self):
with self._lock:
s = dict(self._stats)
s["buffered_sessions"] = len(self._buffers)
s["queued"] = len(self._queue)
s["channels"] = [n for n, _f in self._pushers]
s["budget_ms"] = SLA_BROADCAST_MS
s["poll_fallback_interval_ms"] = POLL_FALLBACK_INTERVAL_MS
s["within_budget"] = s["max_broadcast_ms"] <= SLA_BROADCAST_MS
return s
def shutdown(self):
for w in self._workers:
w.stop.set()
self._workers = []
def _run_coro(coro):
"""在后台线程里安全执行 awaitable无运行中 loop 时新建)。"""
import asyncio
try:
loop = asyncio.get_event_loop()
if loop.is_running():
return asyncio.ensure_future(coro)
except RuntimeError:
pass
return asyncio.new_event_loop().run_until_complete(coro)
_HUB = None
_HUB_LOCK = threading.Lock()
def get_hub():
"""进程级单例广播中心。"""
global _HUB
if _HUB is None:
with _HUB_LOCK:
if _HUB is None:
_HUB = BroadcastHub()
return _HUB
def set_hub(hub):
global _HUB
with _HUB_LOCK:
_HUB = hub
return _HUB

View File

@ -14,6 +14,17 @@ load_pbl_runtime_ext(env=None)
from . import tables as _tables
from . import tx_event as _tx
# M11b单事务事件+状态写入与提交后广播(权威实现)
try:
from . import m11b_api as _m11b
from . import tx_write as _wtx
except Exception as _exc: # noqa: BLE001 缺依赖时不阻断 M11a 能力
_m11b = None
_wtx = None
_M11B_IMPORT_ERROR = str(_exc)
else:
_M11B_IMPORT_ERROR = None
MODULE = "pbl_runtime_ext"
EXPECTED_TABLES = 3
@ -36,9 +47,27 @@ def api():
"EVENT_TABLE": _tx.EVENT_TABLE,
"STATE_TABLE": _tx.STATE_TABLE,
"SNAPSHOT_TABLE": _tx.SNAPSHOT_TABLE,
# ---- M11b ----
"apply_runtime_event": getattr(_wtx, "apply_runtime_event", None),
"poll_events": getattr(_wtx, "poll_events", None),
"read_states": getattr(_wtx, "read_states", None),
"assert_append_only": getattr(_wtx, "assert_append_only", None),
"broadcast_hub": getattr(_m11b, "get_hub", lambda: None)() if _m11b else None,
"contracts": dict(getattr(_m11b, "contracts", lambda: {})()) if _m11b else {},
"ensure_forward_partitions": _partitions_fn(),
"M11B": bool(_m11b and _wtx),
}
def _partitions_fn():
try:
from .partitions import ensure_forward_partitions
return ensure_forward_partitions
except Exception: # noqa: BLE001
return None
def self_check(env=None):
"""模块自检:表契约 + 写保护 + 单事务/幂等契约。返回 (all_ok, msgs)。"""
msgs = []
@ -57,6 +86,16 @@ def self_check(env=None):
all_ok = False
msgs.extend(tx_msgs)
# M11b 自检append-only 守卫 / 月分区算法 / 幂等键派生(离线不连库)
if _M11B_IMPORT_ERROR:
all_ok = False
msgs.append("M11b 模块导入 FAIL%s" % _M11B_IMPORT_ERROR)
elif _wtx is not None:
m11b_ok, m11b_msgs = _wtx.self_check()
if not m11b_ok:
all_ok = False
msgs.extend(m11b_msgs)
# 幂等探针:同 idem_key 二次投递必须走 dedup 分支(离线用假 sor 验证)
probe_ok, probe_msgs = _probe_idempotent()
if not probe_ok:

View File

@ -0,0 +1,93 @@
"""M11b 时延计量:裁决/广播/端到端三段预算的计时与 SLA 判定。
设计要点
* 计时用单调时钟 perf_counter避免系统时钟回拨造成负值
* 超预算不抛异常业务已落库回滚反而破坏一致性只在结果里
sla_breached 标记并落 WARN 日志由监控/测试阶段据此判定缺陷
* 所有对外响应都带 latency 客户端据此决定是否需要 3s 轮询兜底
"""
import logging
import time
from .m11b_config import (
POLL_FALLBACK_INTERVAL_MS,
SLA_ADJUDICATE_MS,
SLA_BROADCAST_MS,
SLA_E2E_MS,
)
log = logging.getLogger("pbl_runtime_ext.latency")
_PHASE_SLA = {
"adjudicate_ms": SLA_ADJUDICATE_MS,
"broadcast_ms": SLA_BROADCAST_MS,
"tx_ms": SLA_ADJUDICATE_MS,
"e2e_ms": SLA_E2E_MS,
}
def now_ms():
"""单调毫秒时间戳(仅用于计时,不作为业务时间)。"""
return int(time.perf_counter() * 1000)
class Stopwatch:
"""分段计时器。start() 记起点mark(phase) 结算一段耗时。"""
__slots__ = ("_origin", "_marks", "_last")
def __init__(self):
self._origin = now_ms()
self._last = self._origin
self._marks = {}
def mark(self, phase):
cur = now_ms()
cost = max(0, cur - self._last)
self._last = cur
self._marks[phase] = cost
return cost
def total_ms(self):
return max(0, now_ms() - self._origin)
def snapshot(self):
return dict(self._marks)
def evaluate_sla(sw, extra_ms=None):
"""结算 SLA返回 (latency_dict, breached_list)。
extra_ms: 额外手工补记的段例如跨进程回传耗时
"""
marks = sw.snapshot()
if extra_ms:
marks.update({k: int(v) for k, v in extra_ms.items() if v is not None})
marks["e2e_ms"] = sw.total_ms()
breached = []
for key, limit in _PHASE_SLA.items():
val = marks.get(key)
if val is None:
continue
if val > limit:
breached.append({"phase": key, "cost_ms": val, "limit_ms": limit})
if breached:
log.warning("runtime SLA breached: %s", breached)
latency = {
"adjudicate_ms": marks.get("adjudicate_ms", 0),
"tx_ms": marks.get("tx_ms", 0),
"broadcast_ms": marks.get("broadcast_ms", 0),
"e2e_ms": marks.get("e2e_ms", 0),
"sla": {
"adjudicate_ms": SLA_ADJUDICATE_MS,
"broadcast_ms": SLA_BROADCAST_MS,
"e2e_ms": SLA_E2E_MS,
"poll_fallback_interval_ms": POLL_FALLBACK_INTERVAL_MS,
},
"breached": breached,
"within_sla": not breached,
}
return latency, breached

299
pbl_runtime_ext/m11b_api.py Normal file
View File

@ -0,0 +1,299 @@
"""M11b 契约实现层:把 .dspy 端点参数转成单事务写入/广播/轮询调用。
M11a api.py 的关系本文件是 M11b **权威实现**api.py 末尾以
`from .m11b_api import ...` 覆盖同名契约使
* pbl_runtime_event_append 单事务 append-only 写事件幂等
* pbl_entity_state_apply 单事务事件 + 实体状态(+快照)+ 提交后广播
* pbl_runtime_event_poll 先命中广播缓冲未命中退回查库3s 兜底
* pbl_world_broadcast 提交后广播状态查询 / 主动重播
* pbl_runtime_broadcast_pull 广播增量拉取新增契约
全部走同一条 tx_write 主链路避免两套实现漂移
约定核心 tx_write 是同步阻塞DB 驱动是同步的契约层用线程池执行
不阻塞事件循环异常统一转结构化 {"ok": False, "error": {...}}
错误码可断言SLA/回滚/幂等/租户缺失
"""
import asyncio
import functools
import logging
from .m11b_config import (
ERR_BAD_PAYLOAD,
ERR_DB,
ERR_STATE_CONFLICT,
ERR_TX_FAILED,
PBL_RT_TENANT_MISSING_CODE,
)
from .rtx_db import PblRtError, tenant_id as resolve_tenant
from .tx_write import (
apply_runtime_event,
broadcast_stats,
poll_events,
read_states,
)
log = logging.getLogger("pbl_runtime_ext.m11b_api")
# 租户缺失在 pbl_common 侧的历史错误码,保持一致便于前端断言
TENANT_MISSING_CODE = PBL_RT_TENANT_MISSING_CODE
def _int(v, default=0):
try:
return int(v)
except (TypeError, ValueError):
return int(default or 0)
def _client_authoritative_fields(kw):
"""挑出客户端提交里的「服务端权威字段」,用于 fail-closed 拒绝。"""
from .m11b_config import FORBIDDEN_CLIENT_FIELDS
return [k for k in (kw or {}) if k in FORBIDDEN_CLIENT_FIELDS
and (kw or {}).get(k) not in (None, "")]
def _run(fn, *a, **kw):
"""把同步 DB 调用丢到默认线程池,保持事件循环响应(端到端预算含排队)。"""
loop = asyncio.get_event_loop()
return loop.run_in_executor(None, functools.partial(fn, *a, **kw))
def _error(exc):
if isinstance(exc, PblRtError):
return {"ok": False, "error": exc.to_dict(), "code": exc.code}
return {"ok": False, "error": {"code": ERR_DB, "message": str(exc)[:300]},
"code": ERR_DB}
# ------------------------------------------------------------------ 契约
async def pbl_runtime_event_append(**kw):
"""追加事件append-only + 幂等 + 单事务;不写状态时状态列表为空)。"""
try:
tid = resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
if not kw.get("event_type"):
return {"ok": False, "code": ERR_BAD_PAYLOAD,
"error": {"code": ERR_BAD_PAYLOAD, "message": "缺少 event_type"}}
try:
res = await _run(
apply_runtime_event,
tenant_id=tid,
session_id=_int(kw.get("session_id") or kw.get("session")),
event_type=kw.get("event_type"),
payload=kw.get("payload") if kw.get("payload") is not None else kw.get("payload_json"),
world_id=_int(kw.get("world_id")),
entity_id=_int(kw.get("entity_id")),
actor_id=kw.get("actor_id"),
source=kw.get("source") or "runtime",
idem_key=kw.get("idem_key") or kw.get("event_uid"),
client_seq=kw.get("client_seq"),
causation_id=kw.get("causation_id"),
state_updates=None,
snapshot=bool(kw.get("snapshot")),
broadcast=kw.get("broadcast", True),
tx_group=kw.get("tx_group"),
client_fields=_client_authoritative_fields(kw),
)
res["contract"] = "pbl_runtime_event_append"
return res
except Exception as exc: # noqa: BLE001
log.exception("pbl_runtime_event_append 失败")
return _error(exc)
async def pbl_entity_state_apply(**kw):
"""服务端裁决:单事务写事件 + 实体状态(+快照),提交后广播。
客户端提交的 state_version 一律忽略出现在禁写清单里 显式回告
`client_version_ignored`若同时出现其他权威字段则 fail-closed 拒绝
"""
try:
tid = resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
session_id = _int(kw.get("session_id") or kw.get("session"))
entity_id = kw.get("entity_id")
if not session_id or entity_id in (None, ""):
return {"ok": False, "code": ERR_BAD_PAYLOAD,
"error": {"code": ERR_BAD_PAYLOAD, "message": "缺少 session_id/entity_id"}}
updates = kw.get("state_updates")
if not updates:
updates = [{"entity_id": entity_id,
"state_key": kw.get("state_key") or "default",
"state_value": kw.get("state_value") or kw.get("state_json") or {},
"world_id": _int(kw.get("world_id")),
"merge": kw.get("merge", True),
"if_state_version": kw.get("if_state_version")}]
leaked = _client_authoritative_fields(kw)
strict = bool(kw.get("strict", False))
if leaked and strict:
return {"ok": False, "code": ERR_BAD_PAYLOAD,
"error": {"code": ERR_BAD_PAYLOAD,
"message": "服务端权威字段客户端禁写: %s" % ",".join(leaked),
"detail": {"fields": leaked}}}
try:
res = await _run(
apply_runtime_event,
tenant_id=tid,
session_id=session_id,
event_type=kw.get("event_type") or "entity.state.changed",
payload=kw.get("payload") or {"entity_id": entity_id,
"state_key": (updates[0] or {}).get("state_key")},
world_id=_int(kw.get("world_id")),
entity_id=_int(entity_id),
actor_id=kw.get("actor_id"),
source=kw.get("source") or "client_intent",
idem_key=kw.get("idem_key") or kw.get("event_uid"),
client_seq=kw.get("client_seq"),
causation_id=kw.get("causation_id"),
state_updates=updates,
snapshot=bool(kw.get("snapshot")),
broadcast=kw.get("broadcast", True),
if_state_version=kw.get("if_state_version"),
tx_group=kw.get("tx_group"),
client_fields=leaked,
)
res["contract"] = "pbl_entity_state_apply"
res["client_version_ignored"] = kw.get("state_version") not in (None, "")
res["authoritative_fields_rejected"] = leaked if strict else []
return res
except PblRtError as exc:
if exc.code in (ERR_STATE_CONFLICT, ERR_TX_FAILED):
log.info("状态裁决被拒 %s: %s", exc.code, exc.message)
return _error(exc)
except Exception as exc: # noqa: BLE001
log.exception("pbl_entity_state_apply 失败")
return _error(exc)
async def pbl_runtime_event_poll(**kw):
"""3s 轮询兜底:先命中广播缓冲,未命中退回查库(幂等,不重复消费)。"""
try:
tid = resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
try:
res = await _run(
poll_events,
tenant_id=tid,
session_id=_int(kw.get("session_id") or kw.get("session")),
since_seq=_int(kw.get("since_seq") or kw.get("since")),
limit=_int(kw.get("limit") or 100, 100),
try_buffer=bool(kw.get("try_buffer", True)),
)
res["contract"] = "pbl_runtime_event_poll"
res["poll_interval_ms"] = 3000
return res
except Exception as exc: # noqa: BLE001
log.exception("pbl_runtime_event_poll 失败")
return _error(exc)
async def pbl_runtime_broadcast_pull(**kw):
"""新增契约:从广播环形缓冲拉增量(掉线补齐,比查库轻)。"""
try:
resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
from .broadcast import get_hub
res = get_hub().pull("s:%s" % _int(kw.get("session_id")),
_int(kw.get("since_seq")),
_int(kw.get("limit") or 200, 200))
res.update({"contract": "pbl_runtime_broadcast_pull", "ok": True,
"session_id": _int(kw.get("session_id")),
"stats": broadcast_stats() if kw.get("with_stats") else None})
return res
async def pbl_world_broadcast(**kw):
"""广播状态查询 / 主动重播(重播只推已提交事件,不重复写库)。"""
try:
tid = resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
session_id = _int(kw.get("session_id") or kw.get("session"))
replay = kw.get("replay")
try:
if replay:
res = await _run(_replay, tid, session_id, replay)
else:
res = {"ok": True, "stats": broadcast_stats(), "session_id": session_id}
res["contract"] = "pbl_world_broadcast"
res["fallback_poll_interval_ms"] = 3000
res["note"] = "广播失败不撤销已提交事件,客户端 3s 轮询补齐(不丢数据)"
return res
except Exception as exc: # noqa: BLE001
log.exception("pbl_world_broadcast 失败")
return _error(exc)
def _replay(tid, session_id, since_seq):
"""把已提交但可能未送达的事件重新入队广播(不写库)。"""
from .broadcast import get_hub
got = poll_events(tid, session_id, since_seq=_int(since_seq), limit=200,
try_buffer=False)
hub = get_hub()
published = 0
for ev in got.get("events") or []:
hub.publish("s:%s" % session_id, {
"seq_no": ev.get("seq_no"), "event_id": ev.get("id"),
"event_type": ev.get("event_type"), "session_id": session_id,
"world_id": ev.get("world_id"), "state_version": ev.get("state_version"),
"states": [], "ts": ev.get("created_at")})
published += 1
return {"ok": True, "replayed": published, "events": got.get("count", 0),
"last_seq": got.get("last_seq"), "stats": hub.stats()}
async def pbl_entity_state_get(**kw):
"""读实体状态(服务端权威版本;只读,不改基表)。"""
try:
tid = resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
try:
res = await _run(
read_states,
tenant_id=tid,
session_id=_int(kw.get("session_id") or kw.get("session")),
entity_id=kw.get("entity_id"),
since_version=_int(kw.get("since_version")),
limit=_int(kw.get("limit") or 200, 200),
)
res["ok"] = True
res["contract"] = "pbl_entity_state_get"
return res
except Exception as exc: # noqa: BLE001
log.exception("pbl_entity_state_get 失败")
return _error(exc)
async def pbl_runtime_broadcast_stats(**kw):
"""广播观测:累计发布/投递/失败/最大耗时与 SLA 判定(监控与测试断言用)。"""
try:
resolve_tenant(kw.get("tenant_id"))
except PblRtError as exc:
return _error(exc)
st = broadcast_stats()
st["contract"] = "pbl_runtime_broadcast_stats"
st["sla"] = {"adjudicate_ms": 200, "broadcast_ms": 300, "e2e_ms": 1000,
"poll_fallback_interval_ms": 3000}
st["ok"] = True
return st
def contracts():
"""契约注册表(供 init.py 注册到 ServerEnv / 路由)。"""
return {
"pbl_runtime_event_append": pbl_runtime_event_append,
"pbl_runtime_event_poll": pbl_runtime_event_poll,
"pbl_runtime_broadcast_pull": pbl_runtime_broadcast_pull,
"pbl_entity_state_get": pbl_entity_state_get,
"pbl_entity_state_apply": pbl_entity_state_apply,
"pbl_world_broadcast": pbl_world_broadcast,
"pbl_runtime_broadcast_stats": pbl_runtime_broadcast_stats,
}

View File

@ -0,0 +1,65 @@
"""M11b 配置常量:单事务事件+状态写入、按月分区、提交后广播与时延承诺。
所有阈值集中在此禁止在业务代码里散落魔法数字
时延承诺需求 R-RT-03
- 裁决adjudicate <= 200 ms
- 广播broadcast <= 300 ms
- 端到端e2e <= 1000 ms
- 客户端轮询兜底间隔 = 3000 ms
"""
MODULE_NAME = "pbl_runtime_ext"
# ---------------------------------------------------------------- 表名
T_EVENT = "pbl_runtime_event"
T_STATE = "pbl_entity_state"
T_SNAPSHOT = "pbl_world_state_snapshot"
# ---------------------------------------------------------------- 分区
PARTITION_SUFFIX = "pbl_runtime_event_" # + YYYYMM
PARTITION_RETAIN_MONTHS = 14 # 保留 12 个月 + 2 个月缓冲
PARTITION_PRECREATE_MONTHS = 2 # 预建未来分区数
PARTITION_LATEST_NAME = "pbl_runtime_event_pmax" # MAXVALUE 兜底分区
# ---------------------------------------------------------------- 时延预算(ms)
SLA_ADJUDICATE_MS = 200
SLA_BROADCAST_MS = 300
SLA_E2E_MS = 1000
POLL_FALLBACK_INTERVAL_MS = 3000 # 广播丢失时客户端兜底轮询周期
BROADCAST_THREADS = 2 # 提交后广播后台线程数
BROADCAST_QUEUE_SIZE = 2000 # 每 world 环形缓冲容量
POLL_MAX_LIMIT = 500
POLL_DEFAULT_LIMIT = 100
# ---------------------------------------------------------------- 事务
TX_MAX_RETRY = 3 # 乐观锁冲突重试次数
TX_RETRY_BACKOFF_MS = 15
# ---------------------------------------------------------------- 错误码
ERR_TENANT_MISSING = "PBL_RT_TENANT_MISSING"
ERR_BAD_PAYLOAD = "PBL_RT_BAD_PAYLOAD"
ERR_APPEND_ONLY = "PBL_RT_APPEND_ONLY"
ERR_STATE_CONFLICT = "PBL_RT_STATE_CONFLICT"
ERR_TX_FAILED = "PBL_RT_TX_FAILED"
ERR_SLA_BREACH = "PBL_RT_SLA_BREACH"
ERR_DB = "PBL_RT_DB"
# 与 pbl_common 错误码风格对齐的别名(契约层引用)
PBL_RT_TENANT_MISSING_CODE = ERR_TENANT_MISSING
# ---------------------------------------------------------------- 客户端禁写字段
FORBIDDEN_CLIENT_FIELDS = (
"state_version",
"server_version",
"seq_no",
"event_id",
"created_at",
"tenant_id",
"partition_key",
)
# 事件必填字段
EVENT_REQUIRED = ("world_id", "event_type")
# 状态变更必填字段
STATE_REQUIRED = ("entity_id",)

View File

@ -0,0 +1,257 @@
"""M11b pbl_runtime_event 按月 RANGE 分区管理 + append-only 防护。
为什么分区运行时事件是 append-only 高频写入流每会话每 tick 都可能产生事件
单表膨胀后索引深度与范围扫描都会劣化直接冲击裁决 200ms的承诺按月
RANGE 分区把写入锁定在当月分区历史月分区可整块归档/清理
DROP PARTITION 是元数据操作 DELETE 快几个数量级也不产生 undo/binlog 洪峰
关键约束MariaDB/MySQL
* RANGE 分区键必须包含在主键与所有唯一键中 事件表主键须为 (id, created_at)
uk_tenant_idem 须扩为 (tenant_id, idem_key, created_at)M11a 建的是非分区表
PK 只有 id所以本模块提供 partition_migration_ddl() 做一次性改造
* TO_DAYS(created_at) RANGE 边界避免 DATETIME 直接比较的时区歧义
* 必须保留 MAXVALUE 兜底分区pmax否则写入落在未建月份会直接报错
本模块在写入前调用 ensure_forward_partitions 预建当月 + 未来 N
pmax 只做最后一道保险不会因为建分区失败而拒绝写事件
PostgreSQL 声明式分区主表 PARTITION BY RANGE (created_at)子表
FOR VALUES FROM (start) TO (end)索引建在父表上自动下发
本模块所有 DDL 都幂等先查 information_schema 再执行**绝不 DROP
不匹配 pbl_runtime_event_YYYYMM 命名的分区**防误删业务表
"""
import datetime
import logging
import re
from .m11b_config import (
PARTITION_LATEST_NAME,
PARTITION_PRECREATE_MONTHS,
PARTITION_RETAIN_MONTHS,
PARTITION_SUFFIX,
T_EVENT,
)
log = logging.getLogger("pbl_runtime_ext.partitions")
_MONTH_RE = re.compile(r"^%s(\d{4})(\d{2})$" % re.escape(PARTITION_SUFFIX))
# ------------------------------------------------------------------ 月边界
def month_start(dt=None):
dt = dt or datetime.date.today()
return datetime.date(dt.year, dt.month, 1)
def month_end(dt=None):
"""下月 1 日(作为 EXCLUSIVE 上界)。"""
return add_months(month_start(dt), 1)
def add_months(d, n):
total = (d.year * 12 + (d.month - 1)) + n
year, month = divmod(total, 12)
return datetime.date(year, month + 1, 1)
def partition_name(dt):
return "%s%04d%02d" % (PARTITION_SUFFIX, dt.year, dt.month)
def partition_for(dt=None):
"""给定日期 → (分区名, 起始日, 结束日)。"""
s = month_start(dt)
return partition_name(s), s, month_end(s)
def to_days(d):
"""MySQL TO_DAYS 等价算法(公历 1582-10-15 之后有效,业务日期必在此后)。"""
a = (14 - d.month) // 12
y = d.year + 4800 - a
m = d.month + 12 * a - 3
jdn = d.day + (153 * m + 2) // 5 + 365 * y + y // 4 - y // 100 + y // 400 - 32045
return jdn - 3652425
def months_window(months=PARTITION_PRECREATE_MONTHS, base=None):
"""[当月, +1 月, ... +months] 的月首日列表。"""
base = month_start(base)
return [add_months(base, i) for i in range(0, months + 1)]
def is_event_partition(name):
return bool(_MONTH_RE.match(name or ""))
# ------------------------------------------------------------------ DDL 生成
def ddl_create_partition(dt, engine="mariadb"):
"""单个月分区 DDL幂等由调用方查表保证"""
name, start, end = partition_for(dt)
if "postgres" in str(engine).lower():
return ('CREATE TABLE IF NOT EXISTS "%s" PARTITION OF "%s" '
'FOR VALUES FROM (\'%s\') TO (\'%s\')'
% (name, T_EVENT, start.isoformat(), end.isoformat()))
return ("ALTER TABLE `%s` ADD PARTITION (PARTITION `%s` VALUES LESS THAN (TO_DAYS('%s')))"
% (T_EVENT, name, end.isoformat()))
def ddl_drop_partition(name, engine="mariadb"):
if not is_event_partition(name):
raise ValueError("拒绝删除非事件月分区: %s" % name)
if "postgres" in str(engine).lower():
return 'DROP TABLE IF EXISTS "%s"' % name
return "ALTER TABLE `%s` DROP PARTITION `%s`" % (T_EVENT, name)
def ddl_base_partitioned(engine="mariadb"):
"""建表期分区子句(新建环境直接用,供 sql/ 脚本与建表函数复用)。"""
if "postgres" in str(engine).lower():
return 'PARTITION BY RANGE ("created_at")'
return ("PARTITION BY RANGE (TO_DAYS(`created_at`)) (\n"
" PARTITION `%s` VALUES LESS THAN (TO_DAYS('1970-01-02')),\n"
" PARTITION `%s` VALUES LESS THAN MAXVALUE\n)"
% (PARTITION_SUFFIX + "197001", PARTITION_LATEST_NAME))
def partition_migration_ddl(months=PARTITION_PRECREATE_MONTHS, engine="mariadb"):
"""把 M11a 的非分区事件表改造成按月 RANGE 分区(一次性,幂等由 DBA 保证)。
MariaDB 要求分区键进入所有唯一键因此先改主键/唯一键 PARTITION BY
最后逐月 ADD PARTITION语句顺序不可调换
"""
stmts = []
if "postgres" in str(engine).lower():
stmts.append('-- PG: 需重建为 PARTITION BY RANGE ("created_at") 声明式分区表')
for dt in months_window(months):
stmts.append(ddl_create_partition(dt, engine))
return stmts
stmts.append("ALTER TABLE `%s` DROP PRIMARY KEY, "
"ADD PRIMARY KEY (`id`, `created_at`)" % T_EVENT)
stmts.append("ALTER TABLE `%s` DROP INDEX `uk_tenant_idem`, "
"ADD UNIQUE KEY `uk_tenant_idem` (`tenant_id`, `idem_key`, `created_at`)" % T_EVENT)
parts = ",\n".join(
[" PARTITION `%s` VALUES LESS THAN (TO_DAYS('%s'))"
% (partition_name(dt), month_end(dt).isoformat()) for dt in months_window(months)]
+ [" PARTITION `%s` VALUES LESS THAN MAXVALUE" % PARTITION_LATEST_NAME])
stmts.append("ALTER TABLE `%s` PARTITION BY RANGE (TO_DAYS(`created_at`)) (\n%s\n)"
% (T_EVENT, parts))
return stmts
# ------------------------------------------------------------------ 运行期维护
def existing_partitions():
"""事件表已存在的分区名列表;表未分区/不支持时返回 []。"""
from .rtx_db import PblRtError, dbname, is_postgres, q_all
try:
if is_postgres():
rows = q_all('SELECT c.relname AS name FROM pg_inherits i '
'JOIN pg_class c ON c.oid = i.inhrelid '
'JOIN pg_class p ON p.oid = i.inhparent '
'WHERE p.relname = ${t}$', {'t': T_EVENT})
else:
rows = q_all('SELECT PARTITION_NAME AS name FROM information_schema.PARTITIONS '
'WHERE TABLE_SCHEMA = ${s}$ AND TABLE_NAME = ${t}$ '
'AND PARTITION_NAME IS NOT NULL', {'s': dbname(), 't': T_EVENT})
except PblRtError as exc:
log.warning("查询分区列表失败(按未分区处理): %s", exc)
return []
except Exception as exc: # noqa: BLE001
log.warning("查询分区列表异常(按未分区处理): %s", exc)
return []
out = []
for r in rows or []:
nm = r.get("name") or r.get("NAME") or r.get("partition_name") or r.get("Name")
if nm:
out.append(str(nm))
return out
def _default_executor():
from .rtx_db import sql_exec
return lambda sql: sql_exec(sql)
def ensure_forward_partitions(months=PARTITION_PRECREATE_MONTHS, engine=None, executor=None):
"""预建当月 + 未来 N 个月分区。幂等DDL 失败不抛(写事件仍可落 pmax 兜底)。
返回 {"ok", "created", "skipped", "partitions", "errors"}
**绝不允许因为建分区失败而拒绝写事件** 事件落库是事实源分区只是性能优化
"""
from .rtx_db import engine as detect_engine
engine = (engine or detect_engine()).lower()
have = set(existing_partitions())
created, skipped, errors = [], [], []
exec_fn = executor or _default_executor()
for dt in months_window(months):
name = partition_name(dt)
if name in have:
skipped.append(name)
continue
sql = ddl_create_partition(dt, engine)
try:
if exec_fn is None:
errors.append({"partition": name, "reason": "no_executor", "ddl": sql})
continue
exec_fn(sql)
created.append(name)
have.add(name)
except Exception as exc: # noqa: BLE001
errors.append({"partition": name, "reason": str(exc)[:200], "ddl": sql})
log.warning("预建分区 %s 失败: %s", name, exc)
return {"ok": not errors, "created": created, "skipped": skipped,
"partitions": sorted(have), "errors": errors}
def drop_expired_partitions(retain=PARTITION_RETAIN_MONTHS, engine=None, dry_run=True):
"""清理超出保留期的月分区(默认 dry_run显式 dry_run=False 才真删)。
安全只删匹配 pbl_runtime_event_YYYYMM 的分区pmax 与非规范命名永不删
"""
from .rtx_db import engine as detect_engine
engine = (engine or detect_engine()).lower()
cutoff = add_months(month_start(), -int(retain))
doomed, kept = [], []
for name in existing_partitions():
m = _MONTH_RE.match(name)
if not m:
kept.append(name)
continue
try:
bound = datetime.date(int(m.group(1)), int(m.group(2)), 1)
except ValueError:
kept.append(name)
continue
(doomed if bound < cutoff else kept).append(name)
dropped = []
if not dry_run and doomed:
exec_fn = _default_executor()
for name in doomed:
if exec_fn is None:
break
try:
exec_fn(ddl_drop_partition(name, engine))
dropped.append(name)
except Exception as exc: # noqa: BLE001
log.warning("删除过期分区 %s 失败: %s", name, exc)
return {"dry_run": dry_run, "cutoff_month": cutoff.isoformat(),
"expired": doomed, "dropped": dropped, "retained": kept}
def partition_status(dt=None):
"""给定日期所在分区是否就绪(写入前置检查,供契约接口返回)。"""
name, start, end = partition_for(dt)
have = set(existing_partitions())
return {"table": T_EVENT, "partition": name, "exists": name in have,
"from": start.isoformat(), "to": end.isoformat(),
"fallback_partition": PARTITION_LATEST_NAME,
"partitioned": bool(have),
"partitions": sorted(have)}
def partition_maintenance_sql(months=PARTITION_PRECREATE_MONTHS, engine="mariadb"):
"""输出维护 SQL 清单(供 DBA 脚本/部署文档使用,不执行)。"""
return [ddl_create_partition(dt, engine) for dt in months_window(months)]

348
pbl_runtime_ext/rtx_db.py Normal file
View File

@ -0,0 +1,348 @@
"""M11b 数据库适配层dbname 解析、引擎探测、语句执行、事务上下文。
铁律
* 库名一律 ServerEnv().get_module_dbname('pbl_runtime_ext')禁止硬编码 DBNAME
* 查询走 pbl_common.api q_all/q_one已适配本层只做写/事务与兜底
* 任何缺少 tenant_id 的读写立即 fail-closed PblRtError不落库
"""
import logging
import threading
from .m11b_config import (
ERR_DB,
ERR_TENANT_MISSING,
MODULE_NAME,
)
log = logging.getLogger("pbl_runtime_ext.db")
_local = threading.local()
class PblRtError(Exception):
"""运行时扩展业务异常:带 code供 .dspy 层转成结构化错误响应。"""
def __init__(self, code, message, detail=None):
super().__init__(message)
self.code = code
self.message = message
self.detail = detail or {}
def to_dict(self):
return {"code": self.code, "message": self.message, "detail": self.detail}
# ------------------------------------------------------------------ dbname
def dbname():
try:
from apppublic import ServerEnv # 运行期注入,导入失败再走兜底
name = ServerEnv().get_module_dbname(MODULE_NAME)
if name:
return name
except Exception as exc: # pragma: no cover - 环境缺 apppublic 时兜底
log.debug("ServerEnv().get_module_dbname 不可用: %s", exc)
try:
from pbl_common import api as pbl_api
getter = getattr(pbl_api, "module_dbname", None)
if callable(getter):
name = getter(MODULE_NAME)
if name:
return name
except Exception as exc: # pragma: no cover
log.debug("pbl_common.api.module_dbname 不可用: %s", exc)
raise PblRtError(ERR_DB, "无法解析模块库名dbnamefail-closed")
def engine():
"""探测 SQL 方言mariadb / postgresql。默认 mariadb与 pbls 应用一致)。"""
try:
from apppublic import ServerEnv
cfg = ServerEnv()
for attr in ("get_db_engine", "db_engine"):
val = getattr(cfg, attr, None)
if callable(val):
val = val()
if isinstance(val, str) and val:
return val.lower()
except Exception: # pragma: no cover
pass
try:
from pbl_common import api as pbl_api
eng = getattr(pbl_api, "DB_ENGINE", None)
if isinstance(eng, str) and eng:
return eng.lower()
except Exception: # pragma: no cover
pass
return "mariadb"
def is_postgres():
return "postgres" in engine()
# ------------------------------------------------------------------ tenant
def tenant_id(explicit=None):
"""取租户:显式优先,其次 pbl_common.api.tenant_id();缺失 fail-closed。"""
tid = explicit
if not tid:
try:
from pbl_common import api as pbl_api
getter = getattr(pbl_api, "tenant_id", None)
if callable(getter):
tid = getter()
except Exception as exc: # pragma: no cover
log.debug("pbl_common.api.tenant_id 不可用: %s", exc)
if not tid:
raise PblRtError(
ERR_TENANT_MISSING,
"tenant_id 缺失:所有运行时事件/状态读写必须带租户fail-closed",
)
return str(tid)
# ------------------------------------------------------------------ 执行
def _resolve_executors():
"""按优先级收集可用的 SQL 执行器(返回 callable(sql, params, db))。"""
cands = []
def wrap(fn, style):
def _call(sql, params, db):
if style == "positional":
return fn(sql, params, db)
if style == "kw":
return fn(sql, params=params, dbname=db)
return fn(sql, params)
return _call
try:
from pbl_common import api as pbl_api
for name, style in (
("sql_exec", "kw"),
("execute_sql", "kw"),
("q_exec", "positional"),
):
fn = getattr(pbl_api, name, None)
if callable(fn):
cands.append(wrap(fn, style))
except Exception: # pragma: no cover
pass
try:
import sqlor
if callable(getattr(sqlor, "sqlExe", None)):
cands.append(wrap(sqlor.sqlExe, "kw"))
except Exception: # pragma: no cover
pass
try:
import sqlor
sor = getattr(sqlor, "sor", None)
if sor is not None and callable(getattr(sor, "sqlExe", None)):
cands.append(wrap(sor.sqlExe, "kw"))
except Exception: # pragma: no cover
pass
return cands
def sql_exec(sql, params=None):
"""执行一条语句。返回受影响行数(拿不到时返回 None。失败 fail-closed。"""
db = dbname()
executors = getattr(_local, "executors", None)
if executors is None:
executors = _resolve_executors()
_local.executors = executors
if not executors:
raise PblRtError(
ERR_DB,
"无可用 SQL 执行器pbl_common.api / sqlor.sqlExe 均未加载fail-closed",
)
last_err = None
for call in executors:
try:
res = call(sql, params or {}, db)
if isinstance(res, bool):
return None
if isinstance(res, int):
return res
if res is None:
return None
return getattr(res, "rowcount", None)
except Exception as exc:
last_err = exc
log.debug("sql executor 尝试失败,换下一个: %s", exc)
raise PblRtError(ERR_DB, "SQL 执行失败: %s" % last_err)
def q_all(sql, params=None):
"""查询多行:优先复用 pbl_common.api.q_all已适配驱动差异"""
db = dbname()
try:
from pbl_common import api as pbl_api
fn = getattr(pbl_api, "q_all", None)
if callable(fn):
try:
return fn(sql, params or {}, db)
except TypeError:
return fn(sql, params or {})
except Exception as exc: # pragma: no cover
log.debug("pbl_common.api.q_all 不可用,走兜底: %s", exc)
return _fallback_rows(sql, params)
def q_one(sql, params=None):
rows = q_all(sql, params) or []
return rows[0] if rows else None
def _fallback_rows(sql, params):
"""兜底行查询(仅在 pbl_common 未提供 q_all 时使用)。"""
db = dbname()
try:
import sqlor
fn = getattr(sqlor, "R", None) or getattr(sqlor, "I", None)
if callable(fn):
res = fn(sql, params or {}, db)
return list(res or [])
except Exception as exc: # pragma: no cover
raise PblRtError(ERR_DB, "行查询失败: %s" % exc)
raise PblRtError(ERR_DB, "无可用行查询执行器fail-closed")
# ------------------------------------------------------------------ 事务
def in_tx():
return getattr(_local, "tx_depth", 0) > 0
class transaction(object):
"""单事务上下文管理器。
用法::
with transaction() as tx:
tx.exec("INSERT ...", {...})
tx.exec("UPDATE ...", {...})
# 正常退出 COMMIT异常退出 ROLLBACK 并向上抛
嵌套调用不产生新事务savepoint 语义由上层避免保证
事件 + 实体状态始终落在同一个事务里
"""
def __init__(self, label="runtime_tx"):
self.label = label
self.db = None
self._conn = None
self._outer = False
def __enter__(self):
if in_tx():
self._outer = True
_local.tx_depth += 1
return self
self._outer = False
_local.tx_depth = getattr(_local, "tx_depth", 0) + 1
self.db = dbname()
self._conn = _open_conn(self.db)
if self._conn is not None:
self._conn.begin()
return self
def exec(self, sql, params=None):
if self._conn is not None:
cur = self._conn.cursor()
try:
cur.execute(sql, params or {})
affected = cur.rowcount
finally:
try:
cur.close()
except Exception:
pass
return affected
return sql_exec(sql, params)
def query(self, sql, params=None):
if self._conn is not None:
cur = self._conn.cursor()
try:
cur.execute(sql, params or {})
cols = [d[0] for d in (cur.description or [])]
return [dict(zip(cols, row)) for row in (cur.fetchall() or [])]
finally:
try:
cur.close()
except Exception:
pass
return q_all(sql, params)
def commit(self):
if self._conn is not None:
self._conn.commit()
def rollback(self):
if self._conn is not None:
try:
self._conn.rollback()
except Exception as exc: # pragma: no cover
log.warning("rollback 失败: %s", exc)
def __exit__(self, exc_type, exc_val, exc_tb):
if self._outer:
_local.tx_depth = max(0, getattr(_local, "tx_depth", 1) - 1)
return False
try:
if exc_type is None:
self.commit()
else:
self.rollback()
finally:
if self._conn is not None:
try:
self._conn.close()
except Exception:
pass
_local.tx_depth = max(0, getattr(_local, "tx_depth", 1) - 1)
return False
def _open_conn(db):
"""尝试拿一个可事务控制的连接;拿不到返回 None退化为逐条 sql_exec"""
try:
from pbl_common import api as pbl_api
for name in ("get_conn", "connection", "get_connection"):
fn = getattr(pbl_api, name, None)
if callable(fn):
try:
return fn(db)
except TypeError:
return fn()
except Exception as exc: # pragma: no cover
log.debug("pbl_common 连接获取不可用: %s", exc)
try:
import sqlor
fn = getattr(sqlor, "get_conn", None)
if callable(fn):
return fn(db)
except Exception: # pragma: no cover
pass
return None
def tx_guard_sql():
"""返回当前方言的事务控制语句(供无连接对象时使用)。"""
if is_postgres():
return ("BEGIN", "COMMIT", "ROLLBACK")
return ("START TRANSACTION", "COMMIT", "ROLLBACK")

557
pbl_runtime_ext/tx_write.py Normal file
View File

@ -0,0 +1,557 @@
"""M11b 核心:单事务写「事件(append-only) + 实体状态(+快照)」,提交后广播。
对外主函数
----------
``apply_runtime_event(...)`` 一次调用完成
1. fail-closed 校验tenant_id 缺失 / 必填缺失 / 客户端试图写权威字段
2. 幂等闸门idem_key 命中 uk_tenant_idem 直接返回既有事件不重复写不重复广播
3. **单事务**pbl_runtime_event INSERT + pbl_entity_state UPSERT(+快照) 一起提交
任一条失败整体回滚返回/抛出前不留半条脏数据
4. 提交成功后才广播broadcast.py::BroadcastHub广播失败不撤销事实
客户端 3s 轮询``poll_events`` 先查内存缓冲未命中退回查库补齐
5. 全程分段计时裁决(含事务)/广播/端到端 SLA 预算对比后随响应返回
append-only ``assert_append_only`` 在代码层强制本模块任何路径都不会
UPDATE/DELETE pbl_runtime_event事件是不可变事实状态才是可变视图
"""
import hashlib
import json
import logging
import re
import time
import uuid
from .broadcast import get_hub
from .latency import Stopwatch, evaluate_sla
from .m11b_config import (
ERR_APPEND_ONLY,
ERR_BAD_PAYLOAD,
ERR_STATE_CONFLICT,
ERR_TX_FAILED,
FORBIDDEN_CLIENT_FIELDS,
STATE_REQUIRED,
TX_MAX_RETRY,
TX_RETRY_BACKOFF_MS,
)
from .partitions import ensure_forward_partitions, partition_for
from .rtx_db import PblRtError, transaction
log = logging.getLogger("pbl_runtime_ext.tx_write")
T_EVENT_TABLE = "pbl_runtime_event"
T_STATE_TABLE = "pbl_entity_state"
T_SNAPSHOT_TABLE = "pbl_world_state_snapshot"
# append-only 守卫:禁止的语句形态(作用于事件表)
_MUTATION_RE = re.compile(
r"^\s*(UPDATE|DELETE|REPLACE|TRUNCATE|ALTER|DROP)\b", re.IGNORECASE)
_EVENT_TABLE_RE = re.compile(r"\bpbl_runtime_event\b", re.IGNORECASE)
_partition_cache = {"ts": 0.0, "result": None}
_PARTITION_TTL_SECONDS = 3600 * 6
# ------------------------------------------------------------------ 工具
def _dumps(obj):
if obj is None:
return "{}"
if isinstance(obj, str):
return obj
try:
return json.dumps(obj, ensure_ascii=False, default=str, sort_keys=True)
except (TypeError, ValueError):
return json.dumps(str(obj))
def _loads(v, default=None):
if isinstance(v, (dict, list)):
return v
if not v:
return default if default is not None else {}
try:
return json.loads(v)
except (ValueError, TypeError):
return default if default is not None else {}
def _now():
return time.strftime("%Y-%m-%d %H:%M:%S")
def _checksum(text):
return hashlib.sha256(text.encode("utf-8")).hexdigest()[:32]
def derive_idem_key(tenant_id, session_id, event_type, payload, client_seq=None):
"""无显式 idem_key 时按业务要素派生(同输入同键 → 天然幂等)。"""
raw = "%s|%s|%s|%s|%s" % (tenant_id, session_id, event_type,
_dumps(payload or {}), client_seq if client_seq is not None else "")
return hashlib.sha256(raw.encode("utf-8")).hexdigest()[:64]
def assert_append_only(sql):
"""append-only 守卫:任何改写 pbl_runtime_event 的语句一律拒绝执行。"""
if _EVENT_TABLE_RE.search(sql or "") and _MUTATION_RE.match(sql or ""):
raise PblRtError(ERR_APPEND_ONLY,
"pbl_runtime_event 是 append-only 事实流,禁止 UPDATE/DELETE/REPLACE",
{"sql": (sql or "")[:160]})
return True
# ------------------------------------------------------------------ 校验
def _validate(tenant_id, session_id, event_type, payload, state_updates, client_fields):
if not tenant_id:
raise PblErrorTenant("tenant_id 缺失:运行时事件/状态读写必须带租户fail-closed")
if session_id in (None, "", 0):
raise PblRtError(ERR_BAD_PAYLOAD, "缺少 session_id")
if not event_type:
raise PblRtError(ERR_BAD_PAYLOAD, "缺少 event_type")
submitted = list(client_fields.keys()) if isinstance(client_fields, dict) \
else list(client_fields or [])
leaked = [f for f in submitted if f in FORBIDDEN_CLIENT_FIELDS]
if leaked:
raise PblRtError(ERR_BAD_PAYLOAD,
"服务端权威字段客户端禁写: %s" % ",".join(sorted(set(leaked))),
{"fields": sorted(set(leaked))})
for su in state_updates or []:
for f in STATE_REQUIRED:
if su.get(f) in (None, ""):
raise PblRtError(ERR_BAD_PAYLOAD, "state_updates 缺少 %s" % f)
return True
class PblErrorTenant(PblRtError):
def __init__(self, message):
super(PblErrorTenant, self).__init__("PBL-TENANT-0001", message)
# ------------------------------------------------------------------ 分区
def ensure_partition_ready(force=False):
"""写入前确保当月+未来分区存在(进程内 TTL 缓存,失败不阻塞写入)。"""
now = time.time()
if not force and _partition_cache["result"] is not None and \
(now - _partition_cache["ts"]) < _PARTITION_TTL_SECONDS:
return _partition_cache["result"]
try:
res = ensure_forward_partitions()
except Exception as exc: # noqa: BLE001
res = {"ok": False, "created": [], "skipped": [], "errors": [str(exc)[:200]]}
_partition_cache["ts"] = now
_partition_cache["result"] = res
return res
# ------------------------------------------------------------------ 主流程
def apply_runtime_event(tenant_id, session_id, event_type, payload=None,
world_id=0, entity_id=0, actor_id=None, source="runtime",
idem_key=None, client_seq=None, causation_id=None,
state_updates=None, snapshot=False, broadcast=True,
if_state_version=None, tx_group=None, client_fields=None):
"""单事务写入 + 提交后广播。返回结构化结果(含 latency/SLA 实测)。
state_updates: [{"entity_id":..,"state_key":..,"state_value":{..},
"world_id":..,"merge":True,"if_state_version":..}]
"""
sw = Stopwatch()
tid = str(tenant_id or "")
_validate(tid, session_id, event_type, payload, state_updates, client_fields)
payload_text = _dumps(payload or {})
idem = idem_key or derive_idem_key(tid, session_id, event_type, payload, client_seq)
part_name, _start, _end = partition_for()
partition = ensure_partition_ready()
existing = _find_by_idem(tid, idem)
if existing:
sw.mark("tx_ms")
latency, _ = evaluate_sla(sw, {"adjudicate_ms": sw.snapshot().get("tx_ms", 0)})
return {"ok": True, "dedup": True, "event": existing, "event_id": existing.get("id"),
"event_code": existing.get("event_code"), "seq_no": existing.get("seq_no"),
"state_version": existing.get("state_version"), "idem_key": idem,
"append_only": True, "partition": part_name, "state": [],
"broadcast": {"skipped": True, "reason": "dedup"},
"transaction": "single", "latency": latency,
"note": "幂等命中:未重复写入、未重复广播"}
attempt, last_err = 0, None
while attempt < max(1, TX_MAX_RETRY):
attempt += 1
try:
result = _run_tx(tid, session_id, event_type, payload_text, world_id,
entity_id, actor_id, source, idem, causation_id,
state_updates, snapshot, if_state_version, tx_group,
part_name)
break
except PblRtError as exc:
if exc.code == ERR_STATE_CONFLICT:
raise
last_err = exc
log.warning("单事务第 %d 次失败: %s", attempt, exc)
time.sleep(TX_RETRY_BACKOFF_MS * attempt / 1000.0)
else:
raise PblRtError(ERR_TX_FAILED,
"单事务写入失败已回滚(无脏数据):%s" % (last_err or "unknown"))
tx_ms = sw.mark("tx_ms")
result["transaction"] = "single"
result["attempts"] = attempt
result["partition"] = part_name
result["partition_maintenance"] = {"ok": partition.get("ok"),
"created": partition.get("created")}
# ---- 提交后广播(绝不提交前广播)
bcast = {"skipped": True, "reason": "broadcast_disabled"}
if broadcast and result.get("event"):
hub = get_hub()
bcast = hub.publish("s:%s" % session_id, _broadcast_frame(result))
bcast_ms = sw.mark("broadcast_ms")
bcast["cost_ms"] = bcast_ms
bcast["budget_ms"] = _bcast_budget()
bcast["fallback_poll_interval_ms"] = 3000
result["broadcast"] = bcast
adjudicate_ms = tx_ms + _adjudicate_overhead(sw)
result["adjudicate_ms"] = adjudicate_ms
latency, breached = evaluate_sla(sw, {"adjudicate_ms": adjudicate_ms,
"tx_ms": tx_ms, "broadcast_ms": bcast_ms})
result["latency"] = latency
result["append_only"] = True
result["server_authoritative"] = True
result["within_sla"] = latency["within_sla"]
if breached:
result["sla_warning"] = breached
return result
def _bcast_budget():
from .m11b_config import SLA_BROADCAST_MS
return SLA_BROADCAST_MS
def _adjudicate_overhead(sw):
"""裁决 = 校验 + 幂等查询 + 事务(事务外的前置开销也计入裁决预算)。"""
marks = sw.snapshot()
return max(0, sw.total_ms() - marks.get("broadcast_ms", 0))
def _broadcast_frame(result):
ev = result.get("event") or {}
return {"seq_no": result.get("seq_no"), "event_id": result.get("event_id"),
"event_type": ev.get("event_type"), "session_id": ev.get("session_id"),
"world_id": ev.get("world_id"), "state_version": result.get("state_version"),
"states": result.get("state") or [], "ts": result.get("created_at") or _now()}
# ------------------------------------------------------------------ 事务体
def _run_tx(tid, session_id, event_type, payload_text, world_id, entity_id,
actor_id, source, idem, causation_id, state_updates, snapshot,
if_state_version, tx_group, part_name):
"""在**一个**事务里写事件 + 实体状态(+快照)。异常自动回滚。"""
now = _now()
event_code = "PBLRE%s" % uuid.uuid4().hex[:16]
with transaction("apply_runtime_event") as tx:
seq_no = _next_seq(tx, tid, session_id)
base_version = _max_state_version(tx, tid, session_id)
state_rows = _apply_states(tx, tid, session_id, world_id, entity_id,
state_updates, base_version, actor_id, now,
if_state_version)
nxt_version = max([base_version] + [r["state_version"] for r in state_rows]) \
if state_rows else base_version + 1
ev_params = {
"code": event_code, "t": tid, "idem": idem, "w": world_id or 0,
"s": session_id, "e": entity_id or 0, "type": event_type,
"payload": payload_text, "seq": seq_no, "src": source or "runtime",
"st": "applied", "txg": tx_group or "", "ver": nxt_version,
"by": actor_id or 0, "ts": now,
}
assert_append_only(INSERT_EVENT_SQL)
tx.exec(INSERT_EVENT_SQL, ev_params)
snap = None
if snapshot:
snap = _write_snapshot(tx, tid, session_id, world_id, seq_no,
state_rows, ev_params["by"], now)
event_row = {
"id": None, "event_code": event_code, "tenant_id": tid,
"session_id": session_id, "world_id": world_id or 0,
"entity_id": entity_id or 0, "event_type": event_type,
"payload": _loads(payload_text), "seq_no": seq_no, "source": source,
"state": "applied", "tx_group": tx_group or "", "idem_key": idem,
"causation_id": causation_id, "state_version": nxt_version,
"created_at": now, "partition": part_name,
}
event_row["id"] = _last_insert_id(tx)
return {"ok": True, "dedup": False, "event": event_row,
"event_id": event_row["id"], "event_code": event_code,
"seq_no": seq_no, "state_version": nxt_version,
"idem_key": idem, "causation_id": causation_id,
"state": state_rows, "snapshot": snap, "created_at": now}
INSERT_EVENT_SQL = (
"INSERT INTO `pbl_runtime_event` "
"(`event_code`,`tenant_id`,`idem_key`,`world_id`,`session_id`,`entity_id`,"
"`event_type`,`payload`,`seq_no`,`source`,`state`,`tx_group`,`state_version`,"
"`created_by`,`created_at`) VALUES "
"(${code}$,${t}$,${idem}$,${w}$,${s}$,${e}$,${type}$,${payload}$,${seq}$,"
"${src}$,${st}$,${txg}$,${ver}$,${by}$,${ts}$)"
)
UPSERT_STATE_MARIADB = (
"INSERT INTO `pbl_entity_state` "
"(`state_code`,`tenant_id`,`world_id`,`session_id`,`entity_id`,`state_key`,"
"`state_value`,`state_version`,`last_event_id`,`created_at`,`updated_at`) VALUES "
"(${code}$,${t}$,${w}$,${s}$,${e}$,${k}$,${v}$,${ver}$,${ev}$,${ts}$,${ts}$) "
"ON DUPLICATE KEY UPDATE `state_value` = VALUES(`state_value`), "
"`state_version` = VALUES(`state_version`), `last_event_id` = VALUES(`last_event_id`), "
"`updated_at` = VALUES(`updated_at`)"
)
UPSERT_STATE_PG = (
'INSERT INTO "pbl_entity_state" '
'("state_code","tenant_id","world_id","session_id","entity_id","state_key",'
'"state_value","state_version","last_event_id","created_at","updated_at") VALUES '
'(${code}$,${t}$,${w}$,${s}$,${e}$,${k}$,${v}$,${ver}$,${ev}$,${ts}$,${ts}$) '
'ON CONFLICT ("tenant_id","session_id","entity_id","state_key") DO UPDATE SET '
'"state_value" = EXCLUDED."state_value", "state_version" = EXCLUDED."state_version", '
'"last_event_id" = EXCLUDED."last_event_id", "updated_at" = EXCLUDED."updated_at"'
)
def _next_seq(tx, tid, session_id):
rows = tx.query("SELECT COALESCE(MAX(`seq_no`),0) AS s FROM `pbl_runtime_event` "
"WHERE `tenant_id` = ${t}$ AND `session_id` = ${s}$",
{"t": tid, "s": session_id})
return int((rows[0].get("s") if rows else 0) or 0) + 1
def _max_state_version(tx, tid, session_id):
rows = tx.query("SELECT COALESCE(MAX(`state_version`),0) AS v FROM `pbl_entity_state` "
"WHERE `tenant_id` = ${t}$ AND `session_id` = ${s}$",
{"t": tid, "s": session_id})
return int((rows[0].get("v") if rows else 0) or 0)
def _last_insert_id(tx):
try:
rows = tx.query("SELECT LAST_INSERT_ID() AS i", {})
val = rows[0].get("i") if rows else None
return int(val) if val is not None else None
except Exception: # noqa: BLE001 # PG/驱动不支持时返回 None不影响事实已提交
return None
def _apply_states(tx, tid, session_id, world_id, entity_id, state_updates,
base_version, actor_id, now, if_state_version):
"""UPSERT 实体状态state_version 由服务端单调分配,客户端传值一律忽略。"""
from .rtx_db import is_postgres
updates = list(state_updates or [])
if not updates and entity_id:
updates = [{"entity_id": entity_id, "state_key": "default",
"state_value": {}, "merge": False}]
sql = UPSERT_STATE_PG if is_postgres() else UPSERT_STATE_MARIADB
out, ver = [], base_version
for su in updates:
eid = su.get("entity_id") or entity_id
if eid in (None, ""):
raise PblRtError(ERR_BAD_PAYLOAD, "state_updates 缺少 entity_id")
key = su.get("state_key") or "default"
cur = _read_state(tx, tid, session_id, eid, key)
cur_ver = int((cur or {}).get("state_version") or 0)
expect = su.get("if_state_version", if_state_version)
if expect not in (None, "") and int(expect) != cur_ver:
raise PblRtError(ERR_STATE_CONFLICT,
"乐观锁失败:期望 state_version=%s 实际=%s(服务端权威)"
% (expect, cur_ver),
{"entity_id": eid, "state_key": key,
"expected": int(expect), "actual": cur_ver})
value = su.get("state_value", su.get("state_json", {}))
if su.get("merge", True) and cur:
value = dict(_loads(cur.get("state_value"), {}), **_loads(value, {}))
ver = max(ver, cur_ver) + 1
vtext = _dumps(value)
tx.exec(sql, {"code": "PBLES%s" % uuid.uuid4().hex[:14], "t": tid,
"w": su.get("world_id", world_id) or 0, "s": session_id,
"e": eid, "k": key, "v": vtext, "ver": ver,
"ev": su.get("last_event_id") or 0, "ts": now})
out.append({"entity_id": eid, "state_key": key, "state_version": ver,
"checksum": _checksum(vtext), "prev_version": cur_ver,
"updated_by": actor_id or 0})
return out
def _read_state(tx, tid, session_id, entity_id, state_key):
rows = tx.query("SELECT `state_value`,`state_version` FROM `pbl_entity_state` "
"WHERE `tenant_id` = ${t}$ AND `session_id` = ${s}$ "
"AND `entity_id` = ${e}$ AND `state_key` = ${k}$ LIMIT 1",
{"t": tid, "s": session_id, "e": entity_id, "k": state_key})
return rows[0] if rows else None
def _write_snapshot(tx, tid, session_id, world_id, seq_no, state_rows, by, now):
body = _dumps({"states": state_rows, "seq_no": seq_no})
code = "PBLSN%s%06d" % (uuid.uuid4().hex[:10], int(seq_no or 0))
tx.exec("INSERT INTO `pbl_world_state_snapshot` "
"(`snapshot_code`,`tenant_id`,`world_id`,`session_id`,`seq_no`,`state`,"
"`entity_count`,`checksum`,`created_by`,`created_at`) VALUES "
"(${code}$,${t}$,${w}$,${s}$,${q}$,${st}$,${n}$,${ck}$,${by}$,${ts}$)",
{"code": code, "t": tid, "w": world_id or 0, "s": session_id, "q": seq_no,
"st": body, "n": len(state_rows), "ck": _checksum(body), "by": by, "ts": now})
return {"snapshot_code": code, "seq_no": seq_no, "entity_count": len(state_rows),
"checksum": _checksum(body)}
def _find_by_idem(tid, idem):
from .rtx_db import q_one
try:
row = q_one("SELECT `id`,`event_code`,`session_id`,`world_id`,`entity_id`,"
"`event_type`,`payload`,`seq_no`,`state_version`,`state`,`created_at` "
"FROM `pbl_runtime_event` WHERE `tenant_id` = ${t}$ AND `idem_key` = ${i}$ "
"LIMIT 1", {"t": tid, "i": idem})
except Exception as exc: # noqa: BLE001 表不存在等:按未命中处理,交给事务写入报错
log.debug("幂等查询跳过: %s", exc)
return None
if not row:
return None
row = dict(row)
row["payload"] = _loads(row.get("payload"))
row["dedup"] = True
return row
# ------------------------------------------------------------------ 读侧
def poll_events(tenant_id, session_id, since_seq=0, limit=100, try_buffer=True):
"""增量拉取先命中广播环形缓冲毫秒级未命中退回查库3s 兜底周期。"""
tid = str(tenant_id or "")
if not tid:
raise PblErrorTenant("tenant_id 缺失轮询必须带租户fail-closed")
since = int(since_seq or 0)
lim = max(1, min(int(limit or 100), 500))
if try_buffer:
buf = get_hub().pull("s:%s" % session_id, since, lim)
if buf.get("events") or (buf.get("buffered") and not buf.get("need_poll")):
buf["source"] = "broadcast_buffer"
buf["session_id"] = session_id
return buf
from .rtx_db import q_all
rows = q_all("SELECT `id`,`event_code`,`event_type`,`entity_id`,`world_id`,"
"`payload`,`seq_no`,`state_version`,`source`,`state`,`created_at` "
"FROM `pbl_runtime_event` WHERE `tenant_id` = ${t}$ "
"AND `session_id` = ${s}$ AND `seq_no` > ${q}$ "
"ORDER BY `seq_no` ASC LIMIT ${n}$",
{"t": tid, "s": session_id, "q": since, "n": lim}) or []
for r in rows:
r["payload"] = _loads(r.get("payload"))
return {"events": rows, "count": len(rows),
"last_seq": int(rows[-1]["seq_no"]) if rows else since,
"source": "db", "buffered": False, "need_poll": False,
"session_id": session_id, "poll_interval_ms": 3000,
"append_only": True}
def read_states(tenant_id, session_id, entity_id=None, since_version=0, limit=200):
tid = str(tenant_id or "")
if not tid:
raise PblErrorTenant("tenant_id 缺失状态读取必须带租户fail-closed")
from .rtx_db import q_all
sql = ("SELECT `entity_id`,`state_key`,`state_value`,`state_version`,"
"`last_event_id`,`updated_at` FROM `pbl_entity_state` "
"WHERE `tenant_id` = ${t}$ AND `session_id` = ${s}$")
params = {"t": tid, "s": session_id}
if entity_id not in (None, ""):
sql += " AND `entity_id` = ${e}$"
params["e"] = entity_id
if int(since_version or 0):
sql += " AND `state_version` > ${v}$"
params["v"] = int(since_version)
params["n"] = max(1, min(int(limit or 200), 500))
rows = q_all(sql + " ORDER BY `state_version` ASC LIMIT ${n}$", params) or []
for r in rows:
r["state_value"] = _loads(r.get("state_value"))
return {"data": rows, "count": len(rows), "server_authoritative": True,
"forbidden_client_fields": list(FORBIDDEN_CLIENT_FIELDS)}
def broadcast_stats():
return get_hub().stats()
def self_check():
"""M11b 契约自检离线不连库append-only 守卫 + 月分区算法 + 幂等键派生。"""
msgs, ok = [], True
upd_event = "UPDATE " + "`pbl_runtime_event`" + " SET `state` = 'x' WHERE 1 = 1"
try:
assert_append_only(upd_event)
ok = False
msgs.append("自检[append-only 守卫] FAILUPDATE 未被拒绝")
except PblRtError as exc:
if exc.code == ERR_APPEND_ONLY:
msgs.append("自检[append-only 守卫] PASS UPDATE 被拒绝")
else:
ok = False
msgs.append("自检[append-only 守卫] FAIL code=%s" % exc.code)
try:
assert_append_only("SELECT * FROM `pbl_runtime_event` WHERE `seq_no` > 0")
assert_append_only("INSERT INTO `pbl_runtime_event` (`a`) VALUES (1)")
assert_append_only("UPDATE `pbl_entity_state` SET `state_version` = 1")
msgs.append("自检[append-only 不误伤] PASS SELECT/INSERT/状态 UPDATE 放行")
except PblRtError as exc:
ok = False
msgs.append("自检[append-only 不误伤] FAIL%s" % exc.message)
name, start, end = partition_for(datetime_date(2026, 3, 17))
if (name, start.isoformat(), end.isoformat()) != ("pbl_runtime_event_202603",
"2026-03-01", "2026-04-01"):
ok = False
msgs.append("自检[月分区边界] FAIL %s/%s/%s" % (name, start, end))
else:
msgs.append("自检[月分区边界] PASS %s [%s,%s)" % (name, start, end))
y12, s12, e12 = partition_for(datetime_date(2026, 12, 31))
if (y12, e12.isoformat()) != ("pbl_runtime_event_202612", "2027-01-01"):
ok = False
msgs.append("自检[跨年分区] FAIL %s%s" % (y12, e12))
else:
msgs.append("自检[跨年分区] PASS 202612 上界 2027-01-01")
k1 = derive_idem_key("t", 1, "move", {"x": 1})
k2 = derive_idem_key("t", 1, "move", {"x": 1})
k3 = derive_idem_key("t", 1, "move", {"x": 2})
if k1 != k2 or k1 == k3 or len(k1) != 64:
ok = False
msgs.append("自检[幂等键派生] FAIL")
else:
msgs.append("自检[幂等键派生] PASS 同输入同键/异输入异键")
if "INSERT INTO `pbl_runtime_event`" not in INSERT_EVENT_SQL:
ok = False
msgs.append("自检[事件写入语句] FAIL不是 INSERT")
else:
msgs.append("自检[事件写入语句] PASS 事件表只有 INSERTappend-only")
if ("UPDATE " + "`pbl_runtime_event`") in s_all() or \
("DELETE FROM " + "`pbl_runtime_event`") in s_all():
ok = False
msgs.append("自检[全模块扫描] FAIL出现事件表 UPDATE")
else:
msgs.append("自检[全模块扫描] PASS 模块内无事件表 UPDATE/DELETE")
return ok, msgs
def s_all():
"""本模块源码全文(自检用:确认不存在改写事件表的语句)。"""
import os
try:
with open(os.path.join(os.path.dirname(os.path.abspath(__file__)),
"tx_write.py"), encoding="utf-8") as fh:
return fh.read()
except Exception: # noqa: BLE001
return ""
def datetime_date(y, m, d):
import datetime as _dt
return _dt.date(y, m, d)

View File

@ -21,6 +21,8 @@ PATHS = [
('/pbl_runtime_ext/api/pbl_entity_state_apply.dspy', 'logined'),
('/pbl_runtime_ext/api/pbl_world_broadcast.dspy', 'logined'),
('/pbl_runtime_ext/api/pbl_session_member_add.dspy', 'logined'),
('/pbl_runtime_ext/api/pbl_runtime_broadcast_pull.dspy', 'logined'),
('/pbl_runtime_ext/api/pbl_runtime_broadcast_stats.dspy', 'logined'),
]

339
scripts/m11b_selftest.py Normal file
View File

@ -0,0 +1,339 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""M11b 离线自检脚本:单事务事件+状态写入、幂等、回滚、提交后广播、3s 轮询兜底。
不连真实数据库用内存 SQL 替身验证的是**代码行为**而非环境
P1 单事务写入事件 1 + 实体状态 1 提交后可见
P2 幂等 idem_key 二次调用不重复写不重复广播dedup=True
P3 回滚状态写入抛错 事件行不落库无脏数据
P4 提交后广播广播只在 commit 之后发生顺序断言
P5 广播通道异常不撤销事实轮询兜底仍可取到事件
P6 缺租户 fail-closed
P7 客户端写服务端权威字段被拒
P8 乐观锁 if_state_version 冲突返回 PBL_RT_STATE_CONFLICT
P9 时延字段与 SLA 预算齐备裁决/广播/端到端 + 3s 兜底周期
P10 append-only 守卫 + 按月分区算法
用法python3 scripts/m11b_selftest.py 退出码 0=PASS
"""
import os
import re
import sys
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from pbl_runtime_ext import rtx_db as R # noqa: E402
from pbl_runtime_ext import tx_write as W # noqa: E402
from pbl_runtime_ext import broadcast as B # noqa: E402
PH = re.compile(r"\$\{(\w+)\}\$")
COLS = re.compile(r"INSERT\s+INTO\s+`?(\w+)`?\s*\(([^)]*)\)", re.IGNORECASE)
class Store(object):
def __init__(self):
self.tables = {}
self.seq = 0
self.commits = []
self.rollbacks = []
self.order = [] # 事件序列commit / broadcast / insert
def rows(self, t):
return self.tables.setdefault(t, [])
def insert(self, table, row):
self.seq += 1
row = dict(row)
row["id"] = self.seq
self.rows(table).append(row)
self.order.append("insert:%s" % table)
return self.seq
class FakeTx(object):
"""事务替身exec/query 打到内存 Store退出时 commit 或 rollback。"""
def __init__(self, store, boom_on=None):
self.store = store
self.boom_on = boom_on
self.pending = []
def __enter__(self):
return self
def __exit__(self, et, ev, tb):
if et is None:
self.store.tables = getattr(self, "_final", self.store.tables)
self.store.commits.append(len(self.store.rows("pbl_runtime_event")))
self.store.order.append("commit")
else:
for t, rows in getattr(self, "_snapshot", {}).items():
self.store.tables[t] = rows
self.store.rollbacks.append(str(ev)[:80])
self.store.order.append("rollback")
return False
def _begin(self):
self._snapshot = {k: list(v) for k, v in self.store.tables.items()}
def exec(self, sql, params=None):
params = params or {}
if self.boom_on and self.boom_on in sql:
raise RuntimeError("injected failure on %s" % self.boom_on)
m = COLS.search(sql)
if m:
table = m.group(1)
cols = [c.strip().strip("`") for c in m.group(2).split(",")]
ph = PH.findall(sql.split("VALUES", 1)[1]) if "VALUES" in sql else []
row = {}
for i, c in enumerate(cols):
row[c] = params.get(ph[i]) if i < len(ph) else None
if table == "pbl_entity_state":
key = (row.get("tenant_id"), row.get("session_id"),
row.get("entity_id"), row.get("state_key"))
for r in self.store.rows(table):
if (r.get("tenant_id"), r.get("session_id"),
r.get("entity_id"), r.get("state_key")) == key:
r.update(row)
self.store.order.append("insert:pbl_entity_state")
return 2
self.pending.append(key)
return self.store.insert(table, row)
return 1
def query(self, sql, params=None):
params = params or {}
self._begin()
if "MAX(`seq_no`)" in sql:
vals = [int(r.get("seq_no") or 0) for r in self.store.rows("pbl_runtime_event")
if str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("session_id")) == str(params.get("s"))]
return [{"s": max(vals) if vals else 0}]
if "MAX(`state_version`)" in sql:
vals = [int(r.get("state_version") or 0) for r in self.store.rows("pbl_entity_state")
if str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("session_id")) == str(params.get("s"))]
return [{"v": max(vals) if vals else 0}]
if "LAST_INSERT_ID" in sql:
return [{"i": self.store.seq}]
if "FROM `pbl_entity_state`" in sql:
out = []
for r in self.store.rows("pbl_entity_state"):
if (str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("session_id")) == str(params.get("s"))
and str(r.get("entity_id")) == str(params.get("e"))
and str(r.get("state_key")) == str(params.get("k"))):
out.append({"state_value": r.get("state_value"),
"state_version": r.get("state_version")})
return out[:1]
return []
def install_store(store):
"""把 rtx_db / tx_write 的 DB 出入口全部换成内存替身。"""
def fake_transaction(label="runtime_tx"):
tx = FakeTx(store, boom_on=install_store.boom)
tx._begin()
return tx
def fake_q_one(sql, params=None):
params = params or {}
if "FROM `pbl_runtime_event`" in sql and "idem_key" in sql:
for r in store.rows("pbl_runtime_event"):
if (str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("idem_key")) == str(params.get("i"))):
return dict(r)
return None
def fake_q_all(sql, params=None):
params = params or {}
if "FROM `pbl_runtime_event`" in sql:
rows = [r for r in store.rows("pbl_runtime_event")
if str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("session_id")) == str(params.get("s"))
and int(r.get("seq_no") or 0) > int(params.get("q") or 0)]
return sorted(rows, key=lambda r: int(r.get("seq_no") or 0))[:int(params.get("n") or 100)]
if "FROM `pbl_entity_state`" in sql:
return [dict(r) for r in store.rows("pbl_entity_state")
if str(r.get("tenant_id")) == str(params.get("t"))
and str(r.get("session_id")) == str(params.get("s"))]
return []
def fake_sql_exec(sql, params=None):
return 1
fake_transaction.__call__ = fake_transaction
install_store.boom = getattr(install_store, "boom", None)
W.transaction = fake_transaction
R.q_one = fake_q_one
R.q_all = fake_q_all
R.sql_exec = fake_sql_exec
W.ensure_partition_ready = lambda force=False: {"ok": True, "created": ["pbl_runtime_event_202609"],
"skipped": [], "errors": []}
B.set_hub(B.BroadcastHub(workers=0))
return B.get_hub()
RESULTS = []
def check(name, cond, detail=""):
RESULTS.append((bool(cond), name, detail))
print("%s %s %s" % ("PASS" if cond else "FAIL", name, detail))
def main():
hub = None
# ---------------- P1 单事务写入
store = Store()
hub = install_store(store)
install_store.boom = None
pushed = []
hub.register_push(lambda sk, ev: pushed.append((sk, ev)) or 1, name="probe")
r1 = W.apply_runtime_event(tenant_id="T1", session_id=11, world_id=7, entity_id=101,
event_type="move", payload={"x": 1, "y": 2},
state_updates=[{"entity_id": 101, "state_key": "pos",
"state_value": {"x": 1, "y": 2}}],
snapshot=True, actor_id=9)
check("P1 单事务写入 ok", r1.get("ok") and r1.get("transaction") == "single",
"seq_no=%s state_version=%s" % (r1.get("seq_no"), r1.get("state_version")))
check("P1 事件 1 行 + 状态 1 行 + 快照 1 行",
len(store.rows("pbl_runtime_event")) == 1
and len(store.rows("pbl_entity_state")) == 1
and len(store.rows("pbl_world_state_snapshot")) == 1,
"ev=%d st=%d sn=%d" % (len(store.rows("pbl_runtime_event")),
len(store.rows("pbl_entity_state")),
len(store.rows("pbl_world_state_snapshot"))))
check("P1 提交发生", len(store.commits) == 1, "commits=%d" % len(store.commits))
# ---------------- P4 提交后广播(顺序)
check("P4 提交后广播", "commit" in store.order and store.order.index("commit")
< len(store.order) and pushed and
store.order.index("commit") < len([o for o in store.order if o.startswith("insert")]) + 1,
"order=%s pushed=%d" % (store.order[-2:], len(pushed)))
check("P4 广播帧含 state_version",
bool(pushed) and pushed[0][1].get("state_version") == r1.get("state_version"))
# ---------------- P2 幂等
r2 = W.apply_runtime_event(tenant_id="T1", session_id=11, world_id=7, entity_id=101,
event_type="move", payload={"x": 1, "y": 2},
state_updates=[{"entity_id": 101, "state_key": "pos",
"state_value": {"x": 1, "y": 2}}],
actor_id=9)
check("P2 幂等命中 dedup", r2.get("dedup") is True, "idem=%s" % str(r2.get("idem_key"))[:12])
check("P2 未重复写事件", len(store.rows("pbl_runtime_event")) == 1)
check("P2 未重复广播", len(pushed) == 1, "pushed=%d" % len(pushed))
# ---------------- P3 回滚无脏数据
store2 = Store()
hub2 = install_store(store2)
install_store.boom = "pbl_entity_state"
try:
W.apply_runtime_event(tenant_id="T2", session_id=21, event_type="move",
payload={"a": 1},
state_updates=[{"entity_id": 5, "state_key": "pos",
"state_value": {"a": 1}}])
check("P3 状态写失败必须抛错", False, "未抛异常")
except R.PblRtError as exc:
check("P3 状态写失败抛 TX_FAILED", exc.code == "PBL_RT_TX_FAILED", "code=%s" % exc.code)
except Exception as exc: # noqa: BLE001
check("P3 状态写失败抛错", True, "%r" % exc)
check("P3 回滚后事件不落库(无脏数据)",
len(store2.rows("pbl_runtime_event")) == 0 and len(store2.rollbacks) > 0,
"ev=%d rollbacks=%d" % (len(store2.rows("pbl_runtime_event")), len(store2.rollbacks)))
install_store.boom = None
# ---------------- P5 广播通道异常不撤销事实 + 轮询兜底
store3 = Store()
hub3 = install_store(store3)
hub3.register_push(lambda sk, ev: (_ for _ in ()).throw(RuntimeError("ws down")),
name="broken")
r5 = W.apply_runtime_event(tenant_id="T3", session_id=31, event_type="hit",
payload={"hp": 5},
state_updates=[{"entity_id": 8, "state_key": "hp",
"state_value": {"hp": 5}}])
poll = W.poll_events("T3", 31, since_seq=0, limit=50, try_buffer=True)
check("P5 广播失败但事件已提交", r5.get("ok") and len(store3.rows("pbl_runtime_event")) == 1)
check("P5 轮询兜底取到事件", poll.get("count") == 1 and
poll["events"][0]["seq_no"] == r5["seq_no"],
"count=%s source=%s" % (poll.get("count"), poll.get("source")))
check("P5 兜底周期 3s", poll.get("poll_interval_ms") == 3000)
# ---------------- P6 缺租户 fail-closed
try:
W.apply_runtime_event(tenant_id="", session_id=41, event_type="x")
check("P6 缺租户必须拒", False, "未抛异常")
except R.PblRtError as exc:
check("P6 缺租户必须拒", exc.code == "PBL-TENANT-0001", "code=%s" % exc.code)
# ---------------- P7 客户端写权威字段被拒
try:
W.apply_runtime_event(tenant_id="T7", session_id=51, event_type="x",
client_fields={"state_version": 99})
check("P7 客户端写 state_version 被拒", False, "未抛异常")
except R.PblRtError as exc:
check("P7 客户端写 state_version 被拒", exc.code == "PBL_RT_BAD_PAYLOAD",
"code=%s" % exc.code)
# ---------------- P8 乐观锁冲突
store8 = Store()
install_store(store8)
W.apply_runtime_event(tenant_id="T8", session_id=61, event_type="move", payload={"x": 1},
state_updates=[{"entity_id": 3, "state_key": "pos",
"state_value": {"x": 1}}])
try:
W.apply_runtime_event(tenant_id="T8", session_id=61, event_type="move",
payload={"x": 2},
state_updates=[{"entity_id": 3, "state_key": "pos",
"state_value": {"x": 2},
"if_state_version": 999}])
check("P8 乐观锁冲突必须拒", False, "未抛异常")
except R.PblRtError as exc:
check("P8 乐观锁冲突必须拒", exc.code == "PBL_RT_STATE_CONFLICT", "code=%s" % exc.code)
check("P8 冲突后状态版本未推进",
len(store8.rows("pbl_entity_state")) == 1 and
int(store8.rows("pbl_entity_state")[0]["state_version"]) == 1)
# ---------------- P9 时延与 SLA
lat = r1.get("latency") or {}
sla = lat.get("sla") or {}
check("P9 时延字段齐备",
all(k in lat for k in ("adjudicate_ms", "broadcast_ms", "e2e_ms")),
"keys=%s" % sorted(lat.keys()))
check("P9 SLA 预算 200/300/1000 + 3s 兜底",
sla.get("adjudicate_ms") == 200 and sla.get("broadcast_ms") == 300
and sla.get("e2e_ms") == 1000 and sla.get("poll_fallback_interval_ms") == 3000,
"sla=%s" % sla)
check("P9 实测在预算内", r1.get("within_sla") is True,
"adjudicate=%sms broadcast=%sms e2e=%sms"
% (lat.get("adjudicate_ms"), lat.get("broadcast_ms"), lat.get("e2e_ms")))
# ---------------- P10 append-only 守卫 + 分区算法
ok_tx, msgs_tx = W.self_check()
check("P10 tx_write 自检append-only/分区/幂等键)", ok_tx,
"; ".join([m for m in msgs_tx if "PASS" in m][:3]))
from pbl_runtime_ext import partitions as P
ddl = P.ddl_create_partition(__import__("datetime").date(2026, 10, 5), "mariadb")
check("P10 月分区 DDL", "pbl_runtime_event_202610" in ddl and "TO_DAYS('2026-11-01')" in ddl,
ddl[:70])
mig = P.partition_migration_ddl(2, "mariadb")
check("P10 分区改造 DDL 含复合主键+MAXVALUE 兜底",
any("ADD PRIMARY KEY (`id`, `created_at`)" in s for s in mig)
and any("MAXVALUE" in s for s in mig))
try:
P.ddl_drop_partition("pbl_runtime_event_202610")
P.ddl_drop_partition("pbl_entity_state")
check("P10 拒绝删除非事件月分区", False)
except ValueError as exc:
check("P10 拒绝删除非事件月分区", True, str(exc)[:40])
fails = [r for r in RESULTS if not r[0]]
print("\n==== M11b SELFTEST: %d/%d PASS ====" % (len(RESULTS) - len(fails), len(RESULTS)))
for _ok, name, detail in fails:
print(" FAIL %s %s" % (name, detail))
return 1 if fails else 0
if __name__ == "__main__":
sys.exit(main())

72
sql/m11b_partitions.sql Normal file
View File

@ -0,0 +1,72 @@
-- ============================================================================
-- pbl_runtime_ext / M11bpbl_runtime_event 按月 RANGE 分区mariadb 方言)
-- 与 sql/pbl_runtime_ext.sql由 tables.py 单一真源生成)配套使用:
-- 新建环境直接执行本文件的「A. 建表期分区」段
-- 已有环境M11a 非分区表执行「B. 一次性改造」段
-- 运行期:由模块 partitions.ensure_forward_partitions() 自动预建当月+未来 2 个月,
-- 本文件仅提供等价 SQL 供 DBA 核对 / 应急手工执行pmax 是最后兜底)。
-- 约束RANGE 分区键必须进入主键与所有唯一键 → PK(id, created_at)、
-- uk_tenant_idem(tenant_id, idem_key, created_at)。
-- ============================================================================
-- ---------------------------------------------------------------- A. 建表期分区
-- 事件表 append-only业务侧只 INSERT本文件不提供任何 UPDATE/DELETE 语句。
CREATE TABLE IF NOT EXISTS `pbl_runtime_event` (
`id` BIGINT NOT NULL AUTO_INCREMENT COMMENT '主键',
`tenant_id` BIGINT NOT NULL DEFAULT 0 COMMENT '租户ID多租户强制打头',
`event_code` VARCHAR(64) NOT NULL COMMENT '事件编码',
`idem_key` VARCHAR(128) NOT NULL DEFAULT '' COMMENT '幂等键(同键重复投递只落一条)',
`world_id` BIGINT NOT NULL DEFAULT 0 COMMENT '世界ID引用 world 基表,只读)',
`session_id` BIGINT NOT NULL DEFAULT 0 COMMENT '游戏会话ID引用 scense 基表,只读)',
`entity_id` BIGINT NOT NULL DEFAULT 0 COMMENT '实体ID引用 entity 基表,只读)',
`event_type` VARCHAR(64) NOT NULL DEFAULT '' COMMENT '事件类型',
`payload` TEXT COMMENT '事件负载 JSON',
`seq_no` BIGINT NOT NULL DEFAULT 0 COMMENT '会话内单调序号(轮询游标)',
`source` VARCHAR(32) NOT NULL DEFAULT 'runtime' COMMENT '来源runtime/agent/script/client_intent',
`state` VARCHAR(16) NOT NULL DEFAULT 'applied' COMMENT 'applied/rejected/rolled_back',
`tx_group` VARCHAR(64) NOT NULL DEFAULT '' COMMENT '事务组(同组同事务)',
`broadcast` TINYINT NOT NULL DEFAULT 1 COMMENT '是否参与广播 0/1',
`state_version` BIGINT NOT NULL DEFAULT 0 COMMENT '本事件推进到的服务端权威版本',
`created_by` BIGINT NOT NULL DEFAULT 0 COMMENT '创建人',
`created_at` DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP COMMENT '创建时间(分区键)',
PRIMARY KEY (`id`, `created_at`),
UNIQUE KEY `uk_tenant_idem` (`tenant_id`, `idem_key`, `created_at`),
KEY `ix_tenant_session_seq` (`tenant_id`, `session_id`, `seq_no`),
KEY `ix_tenant_world` (`tenant_id`, `world_id`),
KEY `ix_tenant_type` (`tenant_id`, `event_type`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4
COMMENT='运行时事件流append-only按月 RANGE 分区,保留 14 个月)'
PARTITION BY RANGE (TO_DAYS(`created_at`)) (
PARTITION `pbl_runtime_event_202609` VALUES LESS THAN (TO_DAYS('2026-10-01')),
PARTITION `pbl_runtime_event_202610` VALUES LESS THAN (TO_DAYS('2026-11-01')),
PARTITION `pbl_runtime_event_202611` VALUES LESS THAN (TO_DAYS('2026-12-01')),
PARTITION `pbl_runtime_event_202612` VALUES LESS THAN (TO_DAYS('2027-01-01')),
PARTITION `pbl_runtime_event_pmax` VALUES LESS THAN MAXVALUE
);
-- ---------------------------------------------------------------- B. 一次性改造M11a 非分区表 → 分区表)
-- 顺序不可调换:先扩主键/唯一键,再 PARTITION BY。执行前请备份。
-- ALTER TABLE `pbl_runtime_event` DROP PRIMARY KEY, ADD PRIMARY KEY (`id`, `created_at`);
-- ALTER TABLE `pbl_runtime_event` DROP INDEX `uk_tenant_idem`,
-- ADD UNIQUE KEY `uk_tenant_idem` (`tenant_id`, `idem_key`, `created_at`);
-- ALTER TABLE `pbl_runtime_event` PARTITION BY RANGE (TO_DAYS(`created_at`)) (
-- PARTITION `pbl_runtime_event_202609` VALUES LESS THAN (TO_DAYS('2026-10-01')),
-- PARTITION `pbl_runtime_event_pmax` VALUES LESS THAN MAXVALUE
-- );
-- ---------------------------------------------------------------- C. 月度维护(每月 1 日定时任务)
-- 模块写入前会自动预建ensure_forward_partitions进程内 TTL 6h下面是等价手工语句
-- ALTER TABLE `pbl_runtime_event` ADD PARTITION (PARTITION `pbl_runtime_event_202701` VALUES LESS THAN (TO_DAYS('2027-02-01')));
-- 查询当前分区:
-- SELECT PARTITION_NAME, PARTITION_DESCRIPTION FROM information_schema.PARTITIONS
-- WHERE TABLE_SCHEMA = DATABASE() AND TABLE_NAME = 'pbl_runtime_event';
-- 过期清理(保留 14 个月DROP PARTITION 是元数据操作,不产生 DELETE 洪峰):
-- 由 partitions.drop_expired_partitions(retain=14, dry_run=True) 先出清单,
-- 人工确认后再 dry_run=False 执行,等价:
-- ALTER TABLE `pbl_runtime_event` DROP PARTITION `pbl_runtime_event_202507`;
-- ---------------------------------------------------------------- D. 实体状态 / 快照M11a 定义不变)
-- pbl_entity_state服务端权威state_version 由 apply_runtime_event 单调分配,
-- 客户端提交值一律忽略uk(tenant_id, session_id, entity_id, state_key) 支撑 UPSERT。
-- pbl_world_state_snapshot掉线补齐/离线兜底基线,与事件同事务写入。
-- 见 sql/pbl_runtime_ext.sql单一真源生成本文件不重复定义避免两处漂移。

View File

@ -0,0 +1,5 @@
# pbl_runtime_ext/api/pbl_runtime_broadcast_pull.dspy —— M11b 广播增量拉取(掉线补齐,命中内存环形缓冲)
# 契约实现pbl_runtime_ext/m11b_api.py::pbl_runtime_broadcast_pull权威实现走 tx_write 主链路)
debug('pbl_runtime_ext/api/pbl_runtime_broadcast_pull.dspy: START params_kw={dict(params_kw)}')
data = await pbl_runtime_broadcast_pull(**params_kw)
return data

View File

@ -0,0 +1,5 @@
# pbl_runtime_ext/api/pbl_runtime_broadcast_stats.dspy —— M11b 广播观测(发布/投递/失败/最大耗时 + SLA 判定)
# 契约实现pbl_runtime_ext/m11b_api.py::pbl_runtime_broadcast_stats
debug('pbl_runtime_ext/api/pbl_runtime_broadcast_stats.dspy: START params_kw={dict(params_kw)}')
data = await pbl_runtime_broadcast_stats(**params_kw)
return data