M4(文风): style-auditor 双轨(提取指纹/漂移第四审)+ jobs 长任务框架(zombie reaper) + 回炉 refine + GET /style read-back。 M5(生成+扩展): worldbuilder/character-gen(入库 continuity 409 gate + partition_writes 白名单 + schema→JSONB 形变); 网关多 provider 回退链/熔断/能力降级(Anthropic/Gemini 适配器);Skill registry + 表权限沙箱 + 规则; 前端 角色生成器/世界观/Codex/规则页/技能库/⌘K 命令面板。 K1(Kimi Code 订阅接入): OAuth device-flow(kimi-code)+ 静态 Console key(kimi-code-key)两路径; coding 端点 KimiCLI 伪造头(实测 UA allow-list 门禁,缺则 403)+ JSON 模式结构化(thinking ⊥ tool_choice)。 本地联调修复: CORS 中间件;assemble 注入 premise+「写第N章」指令(修空 prompt 400); GET /outline·/draft read-back + 大纲/工作台/审稿页重载;写页 client/server 常量边界 + notFound 健壮化; 字数 toLocaleString locale 水合;审稿页终稿从已存草稿 seed(修 accept 422)。 门禁: backend ruff/mypy(157)/alembic 无漂移/pytest 451 · frontend lint/tsc/vitest/build。 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
175 lines
6.3 KiB
Python
175 lines
6.3 KiB
Python
"""T5.4 回退链 + 重试 + 熔断(ARCH §4.5)。
|
||
|
||
主模型 transient 失败 → 退避重试 → 仍失败切回退链下一个;回退服务时标
|
||
`served_by.fell_back=True`;记账落实际服务方;链耗尽抛 LLM_UNAVAILABLE。
|
||
熔断:某 provider 连续失败超阈值后短时熔断、直接跳过走回退。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import pytest
|
||
from fakes_resilience import FakeLedger, ScriptedAdapter, chain, chain_resolver, transient
|
||
from ww_llm_gateway.gateway import CircuitBreaker, Gateway
|
||
from ww_llm_gateway.types import LlmRequest
|
||
from ww_shared import AppError, ErrorCode
|
||
|
||
|
||
async def test_primary_success_no_fallback(req: LlmRequest) -> None:
|
||
primary = ScriptedAdapter("deepseek", text="主模型")
|
||
backup = ScriptedAdapter("openai", text="备用")
|
||
ledger = FakeLedger()
|
||
gw = Gateway(
|
||
{"deepseek": primary, "openai": backup},
|
||
ledger,
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.text == "主模型"
|
||
assert resp.served_by.provider == "deepseek"
|
||
assert resp.served_by.fell_back is False
|
||
assert backup.complete_calls == 0
|
||
assert ledger.records[0].provider == "deepseek"
|
||
|
||
|
||
async def test_falls_through_to_next_provider_on_transient(req: LlmRequest) -> None:
|
||
# 主模型每次 complete 都 transient 失败(足够耗尽重试) → 切回退。
|
||
primary = ScriptedAdapter("deepseek", failures=[transient() for _ in range(10)])
|
||
backup = ScriptedAdapter("openai", text="备用结果")
|
||
ledger = FakeLedger()
|
||
gw = Gateway(
|
||
{"deepseek": primary, "openai": backup},
|
||
ledger,
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
max_retries=2,
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.text == "备用结果"
|
||
assert resp.served_by.provider == "openai"
|
||
assert resp.served_by.fell_back is True
|
||
# 记账记实际服务方 openai,不是失败的 deepseek。
|
||
assert len(ledger.records) == 1
|
||
assert ledger.records[0].provider == "openai"
|
||
|
||
|
||
async def test_retries_then_succeeds_on_same_provider(req: LlmRequest) -> None:
|
||
# 前两次 transient,第三次成功 → 不应切回退(max_retries=2 即最多 3 次尝试)。
|
||
primary = ScriptedAdapter("deepseek", text="重试后成功", failures=[transient(), transient()])
|
||
backup = ScriptedAdapter("openai", text="备用")
|
||
ledger = FakeLedger()
|
||
gw = Gateway(
|
||
{"deepseek": primary, "openai": backup},
|
||
ledger,
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
max_retries=2,
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.text == "重试后成功"
|
||
assert resp.served_by.provider == "deepseek"
|
||
assert resp.served_by.fell_back is False
|
||
assert backup.complete_calls == 0
|
||
|
||
|
||
async def test_rate_limited_triggers_fallback(req: LlmRequest) -> None:
|
||
rate_limited = AppError(ErrorCode.RATE_LIMITED, "429")
|
||
primary = ScriptedAdapter("deepseek", failures=[rate_limited for _ in range(10)])
|
||
backup = ScriptedAdapter("openai", text="降级备用")
|
||
ledger = FakeLedger()
|
||
gw = Gateway(
|
||
{"deepseek": primary, "openai": backup},
|
||
ledger,
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
max_retries=1,
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.served_by.provider == "openai"
|
||
assert resp.served_by.fell_back is True
|
||
|
||
|
||
async def test_chain_exhausted_raises_llm_unavailable(req: LlmRequest) -> None:
|
||
p1 = ScriptedAdapter("deepseek", failures=[transient() for _ in range(10)])
|
||
p2 = ScriptedAdapter("openai", failures=[transient() for _ in range(10)])
|
||
gw = Gateway(
|
||
{"deepseek": p1, "openai": p2},
|
||
FakeLedger(),
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
max_retries=1,
|
||
)
|
||
|
||
with pytest.raises(AppError) as exc:
|
||
await gw.run(req)
|
||
|
||
assert exc.value.code == ErrorCode.LLM_UNAVAILABLE
|
||
|
||
|
||
async def test_missing_adapter_in_chain_skipped(req: LlmRequest) -> None:
|
||
# 链上首个 provider 没注册适配器 → 跳过、走下一个(不硬失败)。
|
||
backup = ScriptedAdapter("openai", text="可用")
|
||
gw = Gateway(
|
||
{"openai": backup},
|
||
FakeLedger(),
|
||
chain_resolver=chain_resolver(chain(("missing", "m"), ("openai", "gpt-4o"))),
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.served_by.provider == "openai"
|
||
assert resp.served_by.fell_back is True
|
||
|
||
|
||
# ---- 熔断器 ----
|
||
|
||
|
||
def test_circuit_breaker_trips_after_threshold() -> None:
|
||
cb = CircuitBreaker(threshold=3, reset_seconds=60.0)
|
||
assert cb.is_open("deepseek") is False
|
||
cb.record_failure("deepseek")
|
||
cb.record_failure("deepseek")
|
||
assert cb.is_open("deepseek") is False # 未到阈值
|
||
cb.record_failure("deepseek")
|
||
assert cb.is_open("deepseek") is True # 第 3 次 → 熔断
|
||
|
||
|
||
def test_circuit_breaker_success_resets() -> None:
|
||
cb = CircuitBreaker(threshold=2, reset_seconds=60.0)
|
||
cb.record_failure("deepseek")
|
||
cb.record_success("deepseek")
|
||
cb.record_failure("deepseek")
|
||
assert cb.is_open("deepseek") is False # 成功清零计数
|
||
|
||
|
||
def test_circuit_breaker_reopens_after_cooldown() -> None:
|
||
now = [1000.0]
|
||
cb = CircuitBreaker(threshold=1, reset_seconds=30.0, clock=lambda: now[0])
|
||
cb.record_failure("deepseek")
|
||
assert cb.is_open("deepseek") is True
|
||
now[0] += 31.0 # 冷却窗口过 → 半开(放行试探)
|
||
assert cb.is_open("deepseek") is False
|
||
|
||
|
||
async def test_open_circuit_skips_provider(req: LlmRequest) -> None:
|
||
# 熔断已打开的主 provider 被直接跳过,连 complete 都不调,直接走回退。
|
||
primary = ScriptedAdapter("deepseek", text="不该被调")
|
||
backup = ScriptedAdapter("openai", text="回退服务")
|
||
cb = CircuitBreaker(threshold=1, reset_seconds=60.0)
|
||
cb.record_failure("deepseek") # 预先熔断
|
||
gw = Gateway(
|
||
{"deepseek": primary, "openai": backup},
|
||
FakeLedger(),
|
||
chain_resolver=chain_resolver(chain(("deepseek", "deepseek-chat"), ("openai", "gpt-4o"))),
|
||
breaker=cb,
|
||
)
|
||
|
||
resp = await gw.run(req)
|
||
|
||
assert resp.served_by.provider == "openai"
|
||
assert resp.served_by.fell_back is True
|
||
assert primary.complete_calls == 0
|