feat(analytics): 首响时长+回复率数据管道——消息串算法/日汇总扩展列/员工聚合

批1(A1+A2):管理层要看「每个人的响应质量」。

- 迁移 d9e0f1a2b3c4:msg_daily_stats/stat_daily_staff 各加
  cust_threads/threads_answered/fr_sum_ms(NOT NULL server_default 0,
  幂等守卫照 9b0d0c133035 惯例;models.py 双声明同步)
- summary.py _thread_stats:串=客户连续消息不被员工消息打断的段,
  is_send=None 不参与串判定;fr=员工首条回复−串末条客户消息;
  日切近似 v1(日末未闭合串计 threads 不计 answered,口径写进 docstring)
- recompute_days/recompute_all 都先按 ts 排序再算串(乱序防护)
- staff_summary 聚合三列按账号→员工累加(fr_cnt≡threads_answered 不单列)
- 测试 +10:串算法矩阵 9 例+staff 聚合传递;全量 589 绿
- 本地 86 万行回填验证:daily-stats 60s/staff 10s,rollup 总和与
  明细完全一致,随机 5 行 SQL 独立复算全对

Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
浅念
2026-09-07 14:13:09 +08:00
parent 39791b2a49
commit feb3b98665
6 changed files with 359 additions and 5 deletions
+55 -2
View File
@@ -14,6 +14,17 @@
计数语义与旧实时聚合完全一致:is_send=True 计 send、False 计 recv、
NULL 两边不计;timestamp 为 NULL 的消息不参与统计。
首响/回复率三列(2026-09-07 批1,A1+A2)——「消息串」口径(与 8-24
spec 对齐,定案):
- **串**:同会话内客户连续消息(is_send=False)不被员工消息(is_send=True)
打断的段;is_send=None(unknown)不参与串判定(跳过——不打断也不起串)。
- **首响样本**:每个被回串一个样本 fr_ms = 员工首条回复 ts − 串末条客户
消息 ts。fr_cnt ≡ threads_answered(被回串数=样本数),不单列。
- **日切近似**(v1):单日内计算,日末未闭合的串计 threads 不计 answered,
跨夜的串在次日段重新起算——8-24 spec 的 24h 窗口径的简化,读端拼均值
(fr_sum_ms / threads_answered)不受影响,仅跨夜串的归属日有偏差。
"""
from __future__ import annotations
@@ -113,6 +124,36 @@ def _peer_filter(kind: str, peer: str) -> list:
]
def _thread_stats(rows: list) -> tuple[int, int, int]:
"""消息串口径三值:(cust_threads, threads_answered, fr_sum_ms)。
- 串 = 客户连续消息(is_send=False)不被员工消息(is_send=True)打断
的段;is_send=None(unknown)不参与串判定——跳过,不打断也不起串。
- 每个被回串一个首响样本 fr_ms = 员工首条回复 ts − 串**末条**客户消息
ts;fr_sum_ms 为样本时长和(展示用首响时长 = fr_sum_ms/answered)。
- 日切近似 v1:日末未闭合的串计 threads 不计 answered(跨夜的串在次日
段重新起算——8-24 spec 的 24h 窗口径的简化)。
- rows 须先按 ts 升序(调用方 sorted;SQL 无 ORDER BY,乱序会错算串)。
"""
threads = answered = 0
fr_sum = 0
in_thread = False
thread_last_ts: int | None = None
for ts, is_send in rows:
if is_send is False:
if not in_thread:
in_thread = True
threads += 1
thread_last_ts = int(ts)
elif is_send is True:
if in_thread and thread_last_ts is not None:
answered += 1
fr_sum += int(ts) - thread_last_ts
in_thread = False
# is_send=None:跳过(不参与串判定)
return threads, answered, fr_sum
async def recompute_days(
session: AsyncSession,
keys: set[tuple[str, str, str, str, int]] | list[tuple[str, str, str, str, int]],
@@ -140,7 +181,10 @@ async def recompute_days(
send = recv = 0
last_ts: int | None = None
last_is_send: bool | None = None
for ts, is_send in rows:
# 串算法需要严格时序:先排序(SQL 无 ORDER BY,旧代码遍历中 last_ts
# 用 >= 兜乱序;排序后该兜底语义不变——max ts 与并列末条方向一致)。
ordered = sorted(rows, key=lambda r: r[0])
for ts, is_send in ordered:
if is_send is True:
send += 1
elif is_send is False:
@@ -148,6 +192,7 @@ async def recompute_days(
if last_ts is None or ts >= last_ts:
last_ts = int(ts)
last_is_send = is_send
threads, answered, fr_sum = _thread_stats(ordered)
await session.execute(
sa.delete(MsgDailyStats).where(
MsgDailyStats.channel == channel,
@@ -169,6 +214,9 @@ async def recompute_days(
recv_cnt=recv,
last_ts=last_ts,
last_is_send=last_is_send,
cust_threads=threads,
threads_answered=answered,
fr_sum_ms=fr_sum,
)
)
n += 1
@@ -224,7 +272,8 @@ async def recompute_all(session: AsyncSession) -> int:
send = recv = 0
last_ts = None
last_is_send = None
for ts, is_send in day_rows:
ordered = sorted(day_rows, key=lambda r: r[0]) # 串算法需严格时序
for ts, is_send in ordered:
if is_send is True:
send += 1
elif is_send is False:
@@ -232,6 +281,7 @@ async def recompute_all(session: AsyncSession) -> int:
if last_ts is None or ts >= last_ts:
last_ts = int(ts)
last_is_send = is_send
threads, answered, fr_sum = _thread_stats(ordered)
session.add(
MsgDailyStats(
channel=channel,
@@ -243,6 +293,9 @@ async def recompute_all(session: AsyncSession) -> int:
recv_cnt=recv,
last_ts=last_ts,
last_is_send=last_is_send,
cust_threads=threads,
threads_answered=answered,
fr_sum_ms=fr_sum,
)
)
total_pairs += 1