feat: 软踢预警线提前至 1.95M 用量墙与 10万余额墙

- should_soft_evict: usage ≥ min(effective_limit, 1.95M) 或
  (balance_at_sync < 10万 且 usage>0)——usage>0 防小额初始池(B0<10万)
  模型误杀永久死锁
- should_restore_model 恢复对称: usage < min(额度,1.95M) 且
  (余额≥10万 或 usage==0),否则踢了立即被同步恢复成抖动
- _check_request_block 对已软踢条目(*-limit-reached)恒 429,
  防踢后 1.95M~2M 区间小请求放行→路由原名缺失→400
- startup_audit / restore_blocked_models / restore_model_if_healthy
  统一接入判定; SOFT_EVICT_USAGE_LIMIT/SOFT_EVICT_BALANCE_LIMIT env 可调
- docs §7.6 新增软踢预警线小节(两墙条件/恢复对称/前置联动/2.5% 采集代价)
- 数字用例 9/9 验证(踢/不踢/死锁防护/防抖动/动态墙取 min)
This commit is contained in:
vibe-code
2026-08-13 08:56:26 +00:00
parent f15b64e31f
commit 795cc99faa
2 changed files with 124 additions and 21 deletions
+24
View File
@@ -386,6 +386,28 @@ reserve = 18输入 + 16声明 = 34,实际消耗 +167/+150(正文 ~16 + 思
修复:输出侧取 `max(声明 max_tokens, THINKING_RESERVE_FLOOR=4096)` 兜底思考开销。 修复:输出侧取 `max(声明 max_tokens, THINKING_RESERVE_FLOOR=4096)` 兜底思考开销。
大输入请求仍偏保守(cl100k_base 中文 2-3 token/字高估),方向安全。 大输入请求仍偏保守(cl100k_base 中文 2-3 token/字高估),方向安全。
#### 软踢预警线(2026-08-13)
后置 soft_evict 触发点从"撞有效墙"提前为两个预警线(`volc_limit_sync.py` `should_soft_evict`,
env `SOFT_EVICT_USAGE_LIMIT=1950000` / `SOFT_EVICT_BALANCE_LIMIT=100000` 可调):
| 预警线 | 触发条件 | 目的 |
|---|---|---|
| 用量墙 | usage ≥ min(effective_limit, 1.95M) | 避开贴 2M 边的灰色地带(大请求 429/小请求放行 + reserve 低估/并发击穿余量) |
| 余额墙 | balance_at_sync < 10万 且 usage > 0 | 免费额度见底提前摘除;usage>0 防小额初始池(B0<10万)模型误杀永久死锁 |
**恢复对称(`should_restore_model`)**:usage < min(effective_limit, 1.95M) 且
(余额 ≥ 10万 或 usage == 0) 才恢复——否则踢了立即被下一轮同步恢复成踢/恢复抖动。
恢复函数读 balance_at_sync 与 usage 同源(快照口径)。`restore_blocked_models` /
`restore_model_if_healthy` / `startup_audit` 同口径。
**前置联动**:`_check_request_block` 对已软踢条目(`*-limit-reached`)恒 429,防踢后
1.95M~2M 区间小请求放行 → 路由原名缺失 → 400 路由错误。
**代价**:用量墙让 1.95M~2M 的 5 万采集放弃(明日返还少 5 万 ≈2.5%),换取贴边期容错;
余额墙对正常大池模型无影响(余额 ≥10万 不触发)。数字用例验证 2026-08-13:9/9 场景
(踢/不踢/死锁防护/防抖动/动态墙取 min)全部符合预期。
--- ---
## 8. 排障速查 ## 8. 排障速查
@@ -426,6 +448,8 @@ reserve = 18输入 + 16声明 = 34,实际消耗 +167/+150(正文 ~16 + 思
工具(未公开公式),FAQ 只说"图片计入输入 token"。闸门调用未传 use_default_image_token_count → 工具(未公开公式),FAQ 只说"图片计入输入 token"。闸门调用未传 use_default_image_token_count →
high-detail 图会在 hook 内同步下载(阻塞)。偏差量级 = 每图几十~几百 token,相对 high-detail 图会在 hook 内同步下载(阻塞)。偏差量级 = 每图几十~几百 token,相对
THINKING_RESERVE_FLOOR(4096) 小;方向待实测(Token 计算器对照或发图看 usage),有图流量时关注 THINKING_RESERVE_FLOOR(4096) 小;方向待实测(Token 计算器对照或发图看 usage),有图流量时关注
- [x] 软踢预警线(2026-08-13 实施,§7.6):后置踢出提前为 用量墙 1.95M / 余额墙 10万(usage>0),
恢复对称 + 前置 limit-reached 恒 429;9/9 数字用例验证
- [ ] 空窗期(0:00~返还到账)行为观察:v6 下 B0 自动收紧,确认无异常误杀 - [ ] 空窗期(0:00~返还到账)行为观察:v6 下 B0 自动收紧,确认无异常误杀
- [ ] 数据库泄漏:08-12 一天 deepseek-v4 消耗记录量级约 5 万条导致 Redis 内存膨胀;触发监控/清理机制 - [ ] 数据库泄漏:08-12 一天 deepseek-v4 消耗记录量级约 5 万条导致 Redis 内存膨胀;触发监控/清理机制
- [ ] v6 无静态 2M 兜底时的安全:08-13 幸由 B0(2,242,178) > 2M 静态墙兜底;若某日 B0 < 2M, - [ ] v6 无静态 2M 兜底时的安全:08-13 幸由 B0(2,242,178) > 2M 静态墙兜底;若某日 B0 < 2M,
+100 -21
View File
@@ -173,6 +173,19 @@ HARD_DAILY_LIMITS = {
"__DEFAULT__": 2_000_000, "__DEFAULT__": 2_000_000,
} }
# ==========================================
# 🚨 Soft Evict Thresholds(软踢预警线)
# ==========================================
# 提前于硬限额/动态墙踢出模型的预警线,留出缓冲:
# 1. 用量墙:usage ≥ SOFT_EVICT_USAGE_LIMIT(默认 1.95M) 即踢——避免贴 2M 边时
# 大请求 429/小请求放行的灰色地带(reserve 低估 / 并发击穿余量)
# 2. 余额墙:balance_at_sync < SOFT_EVICT_BALANCE_LIMIT(10万) 且 usage > 0 即踢——
# usage>0 限定防误杀小额初始池模型(B0<10 万未使用时被永久踢出死锁)
# 恢复条件对称(should_restore_model):usage < min(effective_limit, 用量墙)
# 且 (余额 ≥ 10万 或 usage == 0),否则踢了立即被恢复成踢/恢复抖动
SOFT_EVICT_USAGE_LIMIT = float(os.getenv("SOFT_EVICT_USAGE_LIMIT", "1950000"))
SOFT_EVICT_BALANCE_LIMIT = float(os.getenv("SOFT_EVICT_BALANCE_LIMIT", "100000"))
# ========================================== # ==========================================
# 🔌 Redis Connection # 🔌 Redis Connection
# ========================================== # ==========================================
@@ -477,6 +490,67 @@ def get_effective_limit(clean_model, api_base, current_usage=None):
# ========================================== # ==========================================
# 👻 Handler Logic (DailyLimitHandler) # 👻 Handler Logic (DailyLimitHandler)
# ==========================================
# 🚨 Soft Evict / Restore Decision(软踢与恢复对称判定)
# ==========================================
def soft_evict_wall(effective_limit):
"""软踢用量墙 = min(有效限额, SOFT_EVICT_USAGE_LIMIT);None 表示无限额。"""
if effective_limit is None:
return None
return min(float(effective_limit), SOFT_EVICT_USAGE_LIMIT)
def should_soft_evict(clean_model, api_base, current_usage, effective_limit):
"""
软踢判定:任一满足即触发
- 用量墙:current_usage >= min(effective_limit, SOFT_EVICT_USAGE_LIMIT)
- 余额墙:balance_at_sync < SOFT_EVICT_BALANCE_LIMIT 且 current_usage > 0
(balance_at_sync 缺失视为充足;usage>0 防小额初始池模型误杀死锁)
返回 (triggered, reason)。
"""
wall = soft_evict_wall(effective_limit)
if wall is not None and float(current_usage) >= wall:
return True, f"usage {int(float(current_usage)):,} >= soft wall {int(wall):,}"
try:
bal_key = f"limit:dynamic:volcengine:balance_at_sync:{clean_model}"
bal_val = redis_client.get(bal_key) if redis_client else None
if (
bal_val is not None
and float(bal_val) < SOFT_EVICT_BALANCE_LIMIT
and float(current_usage) > 0
):
return (
True,
f"balance {int(float(bal_val)):,} < {SOFT_EVICT_BALANCE_LIMIT:.0f} (usage>0)",
)
except Exception:
pass
return False, ""
def should_restore_model(clean_model, api_base, current_usage, effective_limit):
"""
恢复判定(与 should_soft_evict 对称,防踢/恢复抖动):
usage < min(effective_limit, SOFT_EVICT_USAGE_LIMIT) 且 (余额 ≥ 10万 或 usage == 0)
effective_limit None 时仅按余额条件判定(原逻辑 None → 可恢复)。
"""
wall = soft_evict_wall(effective_limit)
if wall is not None and float(current_usage) >= wall:
return False
try:
bal_key = f"limit:dynamic:volcengine:balance_at_sync:{clean_model}"
bal_val = redis_client.get(bal_key) if redis_client else None
if (
bal_val is not None
and float(bal_val) < SOFT_EVICT_BALANCE_LIMIT
and float(current_usage) > 0
):
return False
except Exception:
pass
return True
# ========================================== # ==========================================
def soft_evict_from_router(target_model, target_api_base): def soft_evict_from_router(target_model, target_api_base):
""" """
@@ -596,14 +670,11 @@ def restore_blocked_models():
# get_effective_limit handles fetching usage if None passed. # get_effective_limit handles fetching usage if None passed.
effective_limit = get_effective_limit(clean_model, original_api_base) effective_limit = get_effective_limit(clean_model, original_api_base)
should_restore = False usage_key = get_usage_key(clean_model, original_api_base)
if effective_limit is None: current_val = float(redis_client.get(usage_key) or 0)
should_restore = True should_restore = should_restore_model(
else: clean_model, original_api_base, current_val, effective_limit
usage_key = get_usage_key(clean_model, original_api_base) )
current_val = float(redis_client.get(usage_key) or 0)
if current_val < float(effective_limit):
should_restore = True
if should_restore: if should_restore:
dprint(f"♻️ [RESTORE] Restoring {backup['model_name']} (Quota OK)") dprint(f"♻️ [RESTORE] Restoring {backup['model_name']} (Quota OK)")
@@ -655,14 +726,11 @@ def restore_model_if_healthy(clean_model, allowed_hashes=None):
continue continue
effective_limit = get_effective_limit(clean_model, m_base) effective_limit = get_effective_limit(clean_model, m_base)
should_restore = False usage_key = get_usage_key(clean_model, m_base)
if effective_limit is None: current_val = float(redis_client.get(usage_key) or 0)
should_restore = True should_restore = should_restore_model(
else: clean_model, m_base, current_val, effective_limit
usage_key = get_usage_key(clean_model, m_base) )
current_val = float(redis_client.get(usage_key) or 0)
if current_val < float(effective_limit):
should_restore = True
if should_restore: if should_restore:
dprint(f"♻️ [RESTORE] Restoring {backup['model_name']} (Quota OK)") dprint(f"♻️ [RESTORE] Restoring {backup['model_name']} (Quota OK)")
@@ -732,11 +800,14 @@ def startup_audit():
continue continue
usage_key = get_usage_key(clean_model, api_base) usage_key = get_usage_key(clean_model, api_base)
current_val = redis_client.get(usage_key) current_val = float(redis_client.get(usage_key) or 0)
if current_val and float(current_val) >= float(effective_limit): evict, evict_reason = should_soft_evict(
clean_model, api_base, current_val, effective_limit
)
if evict:
dprint( dprint(
f"🚫 [STARTUP BLOCK] Found {raw_model} over limit ({int(float(current_val))}/{int(effective_limit)}). Evicting now." f"🚫 [STARTUP BLOCK] {raw_model} soft-evict: {evict_reason}. Evicting now."
) )
soft_evict_from_router(raw_model, api_base) soft_evict_from_router(raw_model, api_base)
blocked_count += 1 blocked_count += 1
@@ -798,6 +869,11 @@ def _check_request_block(model_name, reserve=0.0):
val = redis_client.get(usage_key) val = redis_client.get(usage_key)
usage = float(val or 0) usage = float(val or 0)
# 已被软踢的条目恒 429(无论 reserve),避免踢后灰色区间(如 1.95M~2M)
# 小请求被放行 → 路由原名不存在 → 400 路由错误
if entry_name == f"{model_name}-limit-reached" and "_daily_limit_backup" in entry:
return (clean, usage, effective_limit)
if usage + float(reserve) >= float(effective_limit): if usage + float(reserve) >= float(effective_limit):
return (clean, usage, effective_limit) return (clean, usage, effective_limit)
@@ -964,9 +1040,12 @@ class DailyLimitHandler(CustomLogger):
level="INFO", level="INFO",
) )
if effective_limit is not None and float(current_total) >= float(effective_limit): evict, evict_reason = should_soft_evict(
clean_model, api_base, current_total, effective_limit
)
if evict:
log( log(
f"🚫 [BLOCK] {clean_model} Limit Hit: {current_total}/{effective_limit}", f"🚫 [BLOCK] {clean_model} soft-evict: {evict_reason} (limit: {effective_limit})",
level="WARNING", level="WARNING",
) )
soft_evict_from_router(raw_model, api_base) soft_evict_from_router(raw_model, api_base)