chore: normalize line endings to LF via .gitattributes
Root-cause fix for mixed CRLF/LF in the repo (no .gitattributes previously). Adds '* text=auto eol=lf' and explicit binary rules so all future checkouts are LF. Renormalizes all tracked text files (no logical content change). Backup tags: image qihuaiflow:backup-lineend-20260725-121632, git backup-lineend-pre-20260725-121632.
This commit is contained in:
QihuAI Design committed
1 parent
2a7bb875f9
commit
771c8456b8
39 files changed
+34810
-34777
No files matched your search
@@ -1,395 +1,395 @@
|
||||
"""app/core/orchestrator.py — Phase 3.3 多 Agent 编排图引擎。
|
||||
|
||||
设计目标：
|
||||
- 以轻量 DAG 描述多 Agent 协作：节点(llm/local/external/reduce) + 有向边(数据流)。
|
||||
- 拓扑分层后按层并行执行(asyncio.gather)，互相独立的节点真正并发。
|
||||
- 失败隔离：每个节点异常不影响其它分支；per-node on_error=halt|continue。
|
||||
- 节点执行复用既有原语(_build_system_prompt / _execute_skill_tool / _call_external_agent /
|
||||
3.1 工具沙箱)，完全向后兼容——agent.orchestration 为 null 时 chat_service 走原逻辑。
|
||||
- 并发安全：每个节点用独立 SessionLocal 会话做 DB 读，避免共享请求会话跨协程竞态。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
import logging
|
||||
from typing import Any, Optional
|
||||
|
||||
import httpx
|
||||
|
||||
logger = logging.getLogger("aiflow.orchestrator")
|
||||
|
||||
VALID_NODE_TYPES = ("llm", "local", "external", "reduce")
|
||||
|
||||
|
||||
class OrchestrationError(ValueError):
|
||||
"""图结构非法或执行被 halt。"""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 校验 + 拓扑
|
||||
# ---------------------------------------------------------------------------
|
||||
def validate_graph(graph: Any) -> None:
|
||||
if not isinstance(graph, dict):
|
||||
raise OrchestrationError("orchestration 必须为对象")
|
||||
nodes = graph.get("nodes")
|
||||
edges = graph.get("edges")
|
||||
if not isinstance(nodes, list) or not nodes:
|
||||
raise OrchestrationError("nodes 必须为非空数组")
|
||||
if not isinstance(edges, list):
|
||||
raise OrchestrationError("edges 必须为数组")
|
||||
|
||||
ids = set()
|
||||
for i, n in enumerate(nodes):
|
||||
if not isinstance(n, dict):
|
||||
raise OrchestrationError(f"节点 #{i} 必须为对象")
|
||||
nid = n.get("id")
|
||||
if not nid or not isinstance(nid, str):
|
||||
raise OrchestrationError(f"节点 #{i} 缺少字符串 id")
|
||||
if nid in ids:
|
||||
raise OrchestrationError(f"节点 id 重复: {nid}")
|
||||
ids.add(nid)
|
||||
ntype = n.get("type")
|
||||
if ntype not in VALID_NODE_TYPES:
|
||||
raise OrchestrationError(
|
||||
f"节点 {nid} 类型非法: {ntype} (允许 {', '.join(VALID_NODE_TYPES)})"
|
||||
)
|
||||
if ntype in ("local", "external") and not n.get("agent_id"):
|
||||
raise OrchestrationError(f"节点 {nid} 类型 {ntype} 必须指定 agent_id")
|
||||
|
||||
for i, e in enumerate(edges):
|
||||
if not isinstance(e, dict):
|
||||
raise OrchestrationError(f"边 #{i} 必须为对象")
|
||||
f = e.get("from")
|
||||
t = e.get("to")
|
||||
if f not in ids:
|
||||
raise OrchestrationError(f"边 #{i} 的 from={f} 指向不存在的节点")
|
||||
if t not in ids:
|
||||
raise OrchestrationError(f"边 #{i} 的 to={t} 指向不存在的节点")
|
||||
if f == t:
|
||||
raise OrchestrationError(f"边 #{i} 存在自环: {f}")
|
||||
|
||||
|
||||
def topological_layers(graph: dict) -> list:
|
||||
"""Kahn 算法返回分层(每层内节点可并行)；含环则抛错。"""
|
||||
nodes = graph["nodes"]
|
||||
edges = graph.get("edges", [])
|
||||
indeg = {n["id"]: 0 for n in nodes}
|
||||
adj = {n["id"]: [] for n in nodes}
|
||||
for e in edges:
|
||||
f, t = e["from"], e["to"]
|
||||
adj[f].append(t)
|
||||
indeg[t] += 1
|
||||
order = {n["id"]: i for i, n in enumerate(nodes)}
|
||||
layers: list = []
|
||||
remaining = set(indeg)
|
||||
while remaining:
|
||||
layer = sorted([nid for nid in remaining if indeg[nid] == 0], key=lambda x: order[x])
|
||||
if not layer:
|
||||
cycle = sorted(remaining, key=lambda x: order[x])
|
||||
raise OrchestrationError(f"检测到环路，无法拓扑排序: {cycle}")
|
||||
layers.append(layer)
|
||||
for nid in layer:
|
||||
remaining.discard(nid)
|
||||
for nxt in adj[nid]:
|
||||
indeg[nxt] -= 1
|
||||
return layers
|
||||
|
||||
|
||||
def _node_by_id(graph: dict, nid: str) -> dict:
|
||||
for n in graph["nodes"]:
|
||||
if n["id"] == nid:
|
||||
return n
|
||||
return {}
|
||||
|
||||
|
||||
def _predecessors(graph: dict, nid: str) -> list:
|
||||
return [e["from"] for e in graph.get("edges", []) if e["to"] == nid]
|
||||
|
||||
|
||||
def _successors(graph: dict, nid: str) -> list:
|
||||
return [e["to"] for e in graph.get("edges", []) if e["from"] == nid]
|
||||
|
||||
|
||||
def _find_output_node(graph: dict) -> str:
|
||||
if graph.get("output_node"):
|
||||
return graph["output_node"]
|
||||
sinks = [n["id"] for n in graph["nodes"] if not _successors(graph, n["id"])]
|
||||
if len(sinks) == 1:
|
||||
return sinks[0]
|
||||
for s in sinks:
|
||||
if _node_by_id(graph, s).get("type") == "reduce":
|
||||
return s
|
||||
if sinks:
|
||||
return sinks[0]
|
||||
return graph["nodes"][-1]["id"]
|
||||
|
||||
|
||||
def _assemble_node_input(query: str, upstream: str, is_entry: bool) -> str:
|
||||
"""入口节点直接吃用户原问题；下游节点综合原问题 + 上游输出。"""
|
||||
if is_entry:
|
||||
return query or ""
|
||||
if upstream:
|
||||
return f"原始用户问题:\n{query}\n\n上游节点输出:\n{upstream}"
|
||||
return query or ""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 模型解析 + LLM 调用原语
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _resolve_model_and_endpoint(model_id, db):
|
||||
from app.models import AIModel, ModelProvider
|
||||
from ..capability_registry import resolve_auto_model
|
||||
|
||||
ai_model = None
|
||||
if model_id and model_id != "auto":
|
||||
ai_model = db.query(AIModel).filter(AIModel.id == model_id).first()
|
||||
if not ai_model:
|
||||
ai_model = resolve_auto_model("text", db)
|
||||
if not ai_model:
|
||||
raise OrchestrationError("无法解析到可用模型")
|
||||
provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
|
||||
api_key = ai_model.api_key or (provider.api_key if provider else None)
|
||||
api_url = ai_model.endpoint
|
||||
if not api_url and provider and provider.api_base_url:
|
||||
api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
|
||||
if not api_url or not api_key:
|
||||
raise OrchestrationError(f"节点模型 {ai_model.name} 缺少 endpoint 或 api_key")
|
||||
return ai_model, api_url, api_key
|
||||
|
||||
|
||||
async def _llm_complete(system_prompt: str, user_content: str, model_id, db, user_id) -> str:
|
||||
from app.api.chat_utils import _proxy_for_url
|
||||
from app.core.observability import span
|
||||
|
||||
ai_model, api_url, api_key = await _resolve_model_and_endpoint(model_id, db)
|
||||
messages = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_content},
|
||||
]
|
||||
request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
|
||||
async with span("orchestrator.llm", model=ai_model.id):
|
||||
async with httpx.AsyncClient(
|
||||
timeout=ai_model.timeout or 120,
|
||||
proxy=_proxy_for_url(api_url),
|
||||
trust_env=False,
|
||||
) as client:
|
||||
resp = await client.post(
|
||||
api_url,
|
||||
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
|
||||
json=request_body,
|
||||
)
|
||||
if resp.status_code != 200:
|
||||
raise OrchestrationError(f"LLM 调用失败 HTTP {resp.status_code}: {resp.text[:200]}")
|
||||
data = resp.json()
|
||||
return data["choices"][0]["message"].get("content") or ""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# local 节点：复用单 Agent system prompt + 受限 tool round + 3.1 沙箱
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _run_local_agent_turn(agent: dict, node_input: str, db, user_id, conv_id) -> dict:
|
||||
from app.api.chat_skill import _resolve_market_skills, _execute_skill_tool
|
||||
from app.api.chat_context import _build_system_prompt, _retrieve_kb_context
|
||||
from app.api.chat_service import _resolve_agent_tool_policy, _check_tool_policy
|
||||
from app.api.chat_utils import _proxy_for_url
|
||||
|
||||
skills = _resolve_market_skills(db, agent.get("skill_ids") or [], user_id)
|
||||
kb_context = await _retrieve_kb_context(db, agent.get("knowledge_base_ids") or [], node_input)
|
||||
system_prompt, function_definitions = _build_system_prompt(
|
||||
skills=skills,
|
||||
agents=[agent],
|
||||
kb_context=kb_context,
|
||||
user_content=node_input,
|
||||
auto_discovered=False,
|
||||
)
|
||||
ai_model, api_url, api_key = await _resolve_model_and_endpoint(agent.get("model_id"), db)
|
||||
tool_policy = _resolve_agent_tool_policy([agent])
|
||||
|
||||
messages = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": node_input},
|
||||
]
|
||||
called_skills: list = []
|
||||
tool_call_count = 0
|
||||
max_tool_rounds = 5
|
||||
last_content = ""
|
||||
async with httpx.AsyncClient(
|
||||
timeout=ai_model.timeout or 120,
|
||||
proxy=_proxy_for_url(api_url),
|
||||
trust_env=False,
|
||||
) as client:
|
||||
for round_num in range(max_tool_rounds):
|
||||
request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
|
||||
if function_definitions:
|
||||
request_body["tools"] = [
|
||||
{"type": "function", "function": fd} for fd in function_definitions
|
||||
]
|
||||
request_body["tool_choice"] = "auto"
|
||||
resp = await client.post(
|
||||
api_url,
|
||||
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
|
||||
json=request_body,
|
||||
)
|
||||
if resp.status_code != 200:
|
||||
raise OrchestrationError(f"local 节点 LLM 失败 HTTP {resp.status_code}: {resp.text[:200]}")
|
||||
data = resp.json()
|
||||
message = data["choices"][0]["message"]
|
||||
tool_calls = message.get("tool_calls") or []
|
||||
content = message.get("content") or ""
|
||||
last_content = content
|
||||
if not tool_calls or not function_definitions:
|
||||
return {"text": content, "used_tools": called_skills}
|
||||
messages.append({"role": "assistant", "content": content, "tool_calls": tool_calls})
|
||||
round_tool_count = 0
|
||||
for tc in tool_calls:
|
||||
func_name = tc.get("function", {}).get("name", "")
|
||||
check = _check_tool_policy(
|
||||
func_name, None, skills, tool_policy, tool_call_count, round_tool_count
|
||||
)
|
||||
if not check["allowed"]:
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", f"call_{round_num}"),
|
||||
"content": f"[工具调用已被策略拦截] {check.get('message', check.get('reason'))}",
|
||||
})
|
||||
tool_call_count += 1
|
||||
round_tool_count += 1
|
||||
continue
|
||||
tool_call_count += 1
|
||||
round_tool_count += 1
|
||||
try:
|
||||
func_args = json.loads(tc.get("function", {}).get("arguments", "{}") or "{}")
|
||||
except Exception:
|
||||
func_args = {}
|
||||
res = _execute_skill_tool(db, func_name, func_args, skills, node_input)
|
||||
called_skills.append(func_name)
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", f"call_{round_num}"),
|
||||
"content": res,
|
||||
})
|
||||
return {"text": last_content, "used_tools": called_skills}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 单节点执行（含独立 DB 会话，保证并发安全）
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _run_node(node, graph, blackboard, query, user_id, conv_id, trace_id) -> dict:
|
||||
from app.api.chat_skill import _resolve_market_agents
|
||||
from app.database import SessionLocal
|
||||
|
||||
nid = node["id"]
|
||||
ntype = node["type"]
|
||||
preds = _predecessors(graph, nid)
|
||||
pred_results = [blackboard[p] for p in preds]
|
||||
is_entry = not preds
|
||||
|
||||
if pred_results:
|
||||
parts = []
|
||||
for b in pred_results:
|
||||
label = f"节点 {b.get('node_id')}/{b.get('node_type')}"
|
||||
if b.get("ok"):
|
||||
parts.append(f"【{label}】\n{b.get('output', '')}")
|
||||
else:
|
||||
parts.append(f"【{label} 执行失败】\n错误: {b.get('error', '')}")
|
||||
upstream = "\n\n".join(parts)
|
||||
else:
|
||||
upstream = ""
|
||||
|
||||
with SessionLocal() as db:
|
||||
if ntype in ("llm", "reduce"):
|
||||
if ntype == "llm":
|
||||
sys_prompt = node.get("system_prompt") or (
|
||||
"你是一个处理节点，请根据上游输入完成指定的文本变换、摘要或路由判断，"
|
||||
"只输出处理后的结果，不要多余解释。"
|
||||
)
|
||||
else:
|
||||
sys_prompt = node.get("system_prompt") or (
|
||||
"你是一个结果汇总器，请将多个子任务的输出综合成连贯、完整、面向用户的最终回答。"
|
||||
)
|
||||
uc = _assemble_node_input(query, upstream, is_entry)
|
||||
text = await _llm_complete(sys_prompt, uc, node.get("model_id"), db, user_id)
|
||||
return {"ok": True, "node_id": nid, "node_type": ntype, "output": text}
|
||||
|
||||
if ntype == "local":
|
||||
agents = _resolve_market_agents(db, [node["agent_id"]], user_id)
|
||||
if not agents:
|
||||
raise OrchestrationError(f"local 节点 {nid} 引用的 agent_id={node['agent_id']} 不存在")
|
||||
ni = _assemble_node_input(query, upstream, is_entry)
|
||||
res = await _run_local_agent_turn(agents[0], ni, db, user_id, conv_id)
|
||||
return {
|
||||
"ok": True,
|
||||
"node_id": nid,
|
||||
"node_type": "local",
|
||||
"output": res["text"],
|
||||
"used_tools": res.get("used_tools"),
|
||||
}
|
||||
|
||||
if ntype == "external":
|
||||
from app.api.chat_agents import _call_external_agent
|
||||
|
||||
agents = _resolve_market_agents(db, [node["agent_id"]], user_id)
|
||||
if not agents:
|
||||
raise OrchestrationError(f"external 节点 {nid} 引用的 agent_id={node['agent_id']} 不存在")
|
||||
ni = _assemble_node_input(query, upstream, is_entry)
|
||||
answer, _ = await _call_external_agent(
|
||||
db, agents[0], query=ni, user_id=user_id, ext_conversation_id=None, history=[]
|
||||
)
|
||||
return {"ok": True, "node_id": nid, "node_type": "external", "output": answer or ""}
|
||||
|
||||
raise OrchestrationError(f"未知节点类型 {ntype}")
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 入口
|
||||
# ---------------------------------------------------------------------------
|
||||
async def run_graph(graph, query, db, user_id, conv_id, trace_id=None):
|
||||
"""执行编排图，返回 (answer: str, trace: dict)。
|
||||
|
||||
- 拓扑分层后逐层 asyncio.gather 并行执行。
|
||||
- 节点异常按 on_error=halt|continue 处理（默认 continue，失败时下游拿到错误作为输入）。
|
||||
"""
|
||||
validate_graph(graph)
|
||||
layers = topological_layers(graph)
|
||||
blackboard: dict = {}
|
||||
trace = {"nodes": [], "trace_id": trace_id, "layers": layers}
|
||||
|
||||
for layer in layers:
|
||||
tasks = [
|
||||
_run_node(_node_by_id(graph, nid), graph, blackboard, query, user_id, conv_id, trace_id)
|
||||
for nid in layer
|
||||
]
|
||||
results = await asyncio.gather(*tasks, return_exceptions=True)
|
||||
for nid, res in zip(layer, results):
|
||||
node = _node_by_id(graph, nid)
|
||||
if isinstance(res, Exception):
|
||||
on_error = node.get("on_error", "continue")
|
||||
err_msg = str(res)
|
||||
blackboard[nid] = {
|
||||
"ok": False,
|
||||
"node_id": nid,
|
||||
"node_type": node["type"],
|
||||
"error": err_msg,
|
||||
}
|
||||
trace["nodes"].append({"id": nid, "type": node["type"], "ok": False, "error": err_msg})
|
||||
logger.warning(f"[orchestrator] 节点 {nid} 失败 on_error={on_error}: {err_msg}")
|
||||
if on_error == "halt":
|
||||
raise OrchestrationError(f"节点 {nid} 执行失败(halt): {err_msg}")
|
||||
else:
|
||||
blackboard[nid] = res
|
||||
trace["nodes"].append({
|
||||
"id": nid,
|
||||
"type": res.get("node_type"),
|
||||
"ok": True,
|
||||
"used_tools": res.get("used_tools"),
|
||||
})
|
||||
|
||||
output_node = _find_output_node(graph)
|
||||
final = blackboard.get(output_node, {})
|
||||
if final.get("ok"):
|
||||
answer = final.get("output", "")
|
||||
else:
|
||||
answer = f"⚠️ 编排输出节点 {output_node} 执行失败: {final.get('error', '未知错误')}"
|
||||
trace["output_node"] = output_node
|
||||
return answer, trace
|
||||
"""app/core/orchestrator.py — Phase 3.3 多 Agent 编排图引擎。
|
||||
|
||||
设计目标：
|
||||
- 以轻量 DAG 描述多 Agent 协作：节点(llm/local/external/reduce) + 有向边(数据流)。
|
||||
- 拓扑分层后按层并行执行(asyncio.gather)，互相独立的节点真正并发。
|
||||
- 失败隔离：每个节点异常不影响其它分支；per-node on_error=halt|continue。
|
||||
- 节点执行复用既有原语(_build_system_prompt / _execute_skill_tool / _call_external_agent /
|
||||
3.1 工具沙箱)，完全向后兼容——agent.orchestration 为 null 时 chat_service 走原逻辑。
|
||||
- 并发安全：每个节点用独立 SessionLocal 会话做 DB 读，避免共享请求会话跨协程竞态。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
import logging
|
||||
from typing import Any, Optional
|
||||
|
||||
import httpx
|
||||
|
||||
logger = logging.getLogger("aiflow.orchestrator")
|
||||
|
||||
VALID_NODE_TYPES = ("llm", "local", "external", "reduce")
|
||||
|
||||
|
||||
class OrchestrationError(ValueError):
|
||||
"""图结构非法或执行被 halt。"""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 校验 + 拓扑
|
||||
# ---------------------------------------------------------------------------
|
||||
def validate_graph(graph: Any) -> None:
|
||||
if not isinstance(graph, dict):
|
||||
raise OrchestrationError("orchestration 必须为对象")
|
||||
nodes = graph.get("nodes")
|
||||
edges = graph.get("edges")
|
||||
if not isinstance(nodes, list) or not nodes:
|
||||
raise OrchestrationError("nodes 必须为非空数组")
|
||||
if not isinstance(edges, list):
|
||||
raise OrchestrationError("edges 必须为数组")
|
||||
|
||||
ids = set()
|
||||
for i, n in enumerate(nodes):
|
||||
if not isinstance(n, dict):
|
||||
raise OrchestrationError(f"节点 #{i} 必须为对象")
|
||||
nid = n.get("id")
|
||||
if not nid or not isinstance(nid, str):
|
||||
raise OrchestrationError(f"节点 #{i} 缺少字符串 id")
|
||||
if nid in ids:
|
||||
raise OrchestrationError(f"节点 id 重复: {nid}")
|
||||
ids.add(nid)
|
||||
ntype = n.get("type")
|
||||
if ntype not in VALID_NODE_TYPES:
|
||||
raise OrchestrationError(
|
||||
f"节点 {nid} 类型非法: {ntype} (允许 {', '.join(VALID_NODE_TYPES)})"
|
||||
)
|
||||
if ntype in ("local", "external") and not n.get("agent_id"):
|
||||
raise OrchestrationError(f"节点 {nid} 类型 {ntype} 必须指定 agent_id")
|
||||
|
||||
for i, e in enumerate(edges):
|
||||
if not isinstance(e, dict):
|
||||
raise OrchestrationError(f"边 #{i} 必须为对象")
|
||||
f = e.get("from")
|
||||
t = e.get("to")
|
||||
if f not in ids:
|
||||
raise OrchestrationError(f"边 #{i} 的 from={f} 指向不存在的节点")
|
||||
if t not in ids:
|
||||
raise OrchestrationError(f"边 #{i} 的 to={t} 指向不存在的节点")
|
||||
if f == t:
|
||||
raise OrchestrationError(f"边 #{i} 存在自环: {f}")
|
||||
|
||||
|
||||
def topological_layers(graph: dict) -> list:
|
||||
"""Kahn 算法返回分层(每层内节点可并行)；含环则抛错。"""
|
||||
nodes = graph["nodes"]
|
||||
edges = graph.get("edges", [])
|
||||
indeg = {n["id"]: 0 for n in nodes}
|
||||
adj = {n["id"]: [] for n in nodes}
|
||||
for e in edges:
|
||||
f, t = e["from"], e["to"]
|
||||
adj[f].append(t)
|
||||
indeg[t] += 1
|
||||
order = {n["id"]: i for i, n in enumerate(nodes)}
|
||||
layers: list = []
|
||||
remaining = set(indeg)
|
||||
while remaining:
|
||||
layer = sorted([nid for nid in remaining if indeg[nid] == 0], key=lambda x: order[x])
|
||||
if not layer:
|
||||
cycle = sorted(remaining, key=lambda x: order[x])
|
||||
raise OrchestrationError(f"检测到环路，无法拓扑排序: {cycle}")
|
||||
layers.append(layer)
|
||||
for nid in layer:
|
||||
remaining.discard(nid)
|
||||
for nxt in adj[nid]:
|
||||
indeg[nxt] -= 1
|
||||
return layers
|
||||
|
||||
|
||||
def _node_by_id(graph: dict, nid: str) -> dict:
|
||||
for n in graph["nodes"]:
|
||||
if n["id"] == nid:
|
||||
return n
|
||||
return {}
|
||||
|
||||
|
||||
def _predecessors(graph: dict, nid: str) -> list:
|
||||
return [e["from"] for e in graph.get("edges", []) if e["to"] == nid]
|
||||
|
||||
|
||||
def _successors(graph: dict, nid: str) -> list:
|
||||
return [e["to"] for e in graph.get("edges", []) if e["from"] == nid]
|
||||
|
||||
|
||||
def _find_output_node(graph: dict) -> str:
|
||||
if graph.get("output_node"):
|
||||
return graph["output_node"]
|
||||
sinks = [n["id"] for n in graph["nodes"] if not _successors(graph, n["id"])]
|
||||
if len(sinks) == 1:
|
||||
return sinks[0]
|
||||
for s in sinks:
|
||||
if _node_by_id(graph, s).get("type") == "reduce":
|
||||
return s
|
||||
if sinks:
|
||||
return sinks[0]
|
||||
return graph["nodes"][-1]["id"]
|
||||
|
||||
|
||||
def _assemble_node_input(query: str, upstream: str, is_entry: bool) -> str:
|
||||
"""入口节点直接吃用户原问题；下游节点综合原问题 + 上游输出。"""
|
||||
if is_entry:
|
||||
return query or ""
|
||||
if upstream:
|
||||
return f"原始用户问题:\n{query}\n\n上游节点输出:\n{upstream}"
|
||||
return query or ""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 模型解析 + LLM 调用原语
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _resolve_model_and_endpoint(model_id, db):
|
||||
from app.models import AIModel, ModelProvider
|
||||
from ..capability_registry import resolve_auto_model
|
||||
|
||||
ai_model = None
|
||||
if model_id and model_id != "auto":
|
||||
ai_model = db.query(AIModel).filter(AIModel.id == model_id).first()
|
||||
if not ai_model:
|
||||
ai_model = resolve_auto_model("text", db)
|
||||
if not ai_model:
|
||||
raise OrchestrationError("无法解析到可用模型")
|
||||
provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
|
||||
api_key = ai_model.api_key or (provider.api_key if provider else None)
|
||||
api_url = ai_model.endpoint
|
||||
if not api_url and provider and provider.api_base_url:
|
||||
api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
|
||||
if not api_url or not api_key:
|
||||
raise OrchestrationError(f"节点模型 {ai_model.name} 缺少 endpoint 或 api_key")
|
||||
return ai_model, api_url, api_key
|
||||
|
||||
|
||||
async def _llm_complete(system_prompt: str, user_content: str, model_id, db, user_id) -> str:
|
||||
from app.api.chat_utils import _proxy_for_url
|
||||
from app.core.observability import span
|
||||
|
||||
ai_model, api_url, api_key = await _resolve_model_and_endpoint(model_id, db)
|
||||
messages = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_content},
|
||||
]
|
||||
request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
|
||||
async with span("orchestrator.llm", model=ai_model.id):
|
||||
async with httpx.AsyncClient(
|
||||
timeout=ai_model.timeout or 120,
|
||||
proxy=_proxy_for_url(api_url),
|
||||
trust_env=False,
|
||||
) as client:
|
||||
resp = await client.post(
|
||||
api_url,
|
||||
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
|
||||
json=request_body,
|
||||
)
|
||||
if resp.status_code != 200:
|
||||
raise OrchestrationError(f"LLM 调用失败 HTTP {resp.status_code}: {resp.text[:200]}")
|
||||
data = resp.json()
|
||||
return data["choices"][0]["message"].get("content") or ""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# local 节点：复用单 Agent system prompt + 受限 tool round + 3.1 沙箱
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _run_local_agent_turn(agent: dict, node_input: str, db, user_id, conv_id) -> dict:
|
||||
from app.api.chat_skill import _resolve_market_skills, _execute_skill_tool
|
||||
from app.api.chat_context import _build_system_prompt, _retrieve_kb_context
|
||||
from app.api.chat_service import _resolve_agent_tool_policy, _check_tool_policy
|
||||
from app.api.chat_utils import _proxy_for_url
|
||||
|
||||
skills = _resolve_market_skills(db, agent.get("skill_ids") or [], user_id)
|
||||
kb_context = await _retrieve_kb_context(db, agent.get("knowledge_base_ids") or [], node_input)
|
||||
system_prompt, function_definitions = _build_system_prompt(
|
||||
skills=skills,
|
||||
agents=[agent],
|
||||
kb_context=kb_context,
|
||||
user_content=node_input,
|
||||
auto_discovered=False,
|
||||
)
|
||||
ai_model, api_url, api_key = await _resolve_model_and_endpoint(agent.get("model_id"), db)
|
||||
tool_policy = _resolve_agent_tool_policy([agent])
|
||||
|
||||
messages = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": node_input},
|
||||
]
|
||||
called_skills: list = []
|
||||
tool_call_count = 0
|
||||
max_tool_rounds = 5
|
||||
last_content = ""
|
||||
async with httpx.AsyncClient(
|
||||
timeout=ai_model.timeout or 120,
|
||||
proxy=_proxy_for_url(api_url),
|
||||
trust_env=False,
|
||||
) as client:
|
||||
for round_num in range(max_tool_rounds):
|
||||
request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
|
||||
if function_definitions:
|
||||
request_body["tools"] = [
|
||||
{"type": "function", "function": fd} for fd in function_definitions
|
||||
]
|
||||
request_body["tool_choice"] = "auto"
|
||||
resp = await client.post(
|
||||
api_url,
|
||||
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
|
||||
json=request_body,
|
||||
)
|
||||
if resp.status_code != 200:
|
||||
raise OrchestrationError(f"local 节点 LLM 失败 HTTP {resp.status_code}: {resp.text[:200]}")
|
||||
data = resp.json()
|
||||
message = data["choices"][0]["message"]
|
||||
tool_calls = message.get("tool_calls") or []
|
||||
content = message.get("content") or ""
|
||||
last_content = content
|
||||
if not tool_calls or not function_definitions:
|
||||
return {"text": content, "used_tools": called_skills}
|
||||
messages.append({"role": "assistant", "content": content, "tool_calls": tool_calls})
|
||||
round_tool_count = 0
|
||||
for tc in tool_calls:
|
||||
func_name = tc.get("function", {}).get("name", "")
|
||||
check = _check_tool_policy(
|
||||
func_name, None, skills, tool_policy, tool_call_count, round_tool_count
|
||||
)
|
||||
if not check["allowed"]:
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", f"call_{round_num}"),
|
||||
"content": f"[工具调用已被策略拦截] {check.get('message', check.get('reason'))}",
|
||||
})
|
||||
tool_call_count += 1
|
||||
round_tool_count += 1
|
||||
continue
|
||||
tool_call_count += 1
|
||||
round_tool_count += 1
|
||||
try:
|
||||
func_args = json.loads(tc.get("function", {}).get("arguments", "{}") or "{}")
|
||||
except Exception:
|
||||
func_args = {}
|
||||
res = _execute_skill_tool(db, func_name, func_args, skills, node_input)
|
||||
called_skills.append(func_name)
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", f"call_{round_num}"),
|
||||
"content": res,
|
||||
})
|
||||
return {"text": last_content, "used_tools": called_skills}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 单节点执行（含独立 DB 会话，保证并发安全）
|
||||
# ---------------------------------------------------------------------------
|
||||
async def _run_node(node, graph, blackboard, query, user_id, conv_id, trace_id) -> dict:
|
||||
from app.api.chat_skill import _resolve_market_agents
|
||||
from app.database import SessionLocal
|
||||
|
||||
nid = node["id"]
|
||||
ntype = node["type"]
|
||||
preds = _predecessors(graph, nid)
|
||||
pred_results = [blackboard[p] for p in preds]
|
||||
is_entry = not preds
|
||||
|
||||
if pred_results:
|
||||
parts = []
|
||||
for b in pred_results:
|
||||
label = f"节点 {b.get('node_id')}/{b.get('node_type')}"
|
||||
if b.get("ok"):
|
||||
parts.append(f"【{label}】\n{b.get('output', '')}")
|
||||
else:
|
||||
parts.append(f"【{label} 执行失败】\n错误: {b.get('error', '')}")
|
||||
upstream = "\n\n".join(parts)
|
||||
else:
|
||||
upstream = ""
|
||||
|
||||
with SessionLocal() as db:
|
||||
if ntype in ("llm", "reduce"):
|
||||
if ntype == "llm":
|
||||
sys_prompt = node.get("system_prompt") or (
|
||||
"你是一个处理节点，请根据上游输入完成指定的文本变换、摘要或路由判断，"
|
||||
"只输出处理后的结果，不要多余解释。"
|
||||
)
|
||||
else:
|
||||
sys_prompt = node.get("system_prompt") or (
|
||||
"你是一个结果汇总器，请将多个子任务的输出综合成连贯、完整、面向用户的最终回答。"
|
||||
)
|
||||
uc = _assemble_node_input(query, upstream, is_entry)
|
||||
text = await _llm_complete(sys_prompt, uc, node.get("model_id"), db, user_id)
|
||||
return {"ok": True, "node_id": nid, "node_type": ntype, "output": text}
|
||||
|
||||
if ntype == "local":
|
||||
agents = _resolve_market_agents(db, [node["agent_id"]], user_id)
|
||||
if not agents:
|
||||
raise OrchestrationError(f"local 节点 {nid} 引用的 agent_id={node['agent_id']} 不存在")
|
||||
ni = _assemble_node_input(query, upstream, is_entry)
|
||||
res = await _run_local_agent_turn(agents[0], ni, db, user_id, conv_id)
|
||||
return {
|
||||
"ok": True,
|
||||
"node_id": nid,
|
||||
"node_type": "local",
|
||||
"output": res["text"],
|
||||
"used_tools": res.get("used_tools"),
|
||||
}
|
||||
|
||||
if ntype == "external":
|
||||
from app.api.chat_agents import _call_external_agent
|
||||
|
||||
agents = _resolve_market_agents(db, [node["agent_id"]], user_id)
|
||||
if not agents:
|
||||
raise OrchestrationError(f"external 节点 {nid} 引用的 agent_id={node['agent_id']} 不存在")
|
||||
ni = _assemble_node_input(query, upstream, is_entry)
|
||||
answer, _ = await _call_external_agent(
|
||||
db, agents[0], query=ni, user_id=user_id, ext_conversation_id=None, history=[]
|
||||
)
|
||||
return {"ok": True, "node_id": nid, "node_type": "external", "output": answer or ""}
|
||||
|
||||
raise OrchestrationError(f"未知节点类型 {ntype}")
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 入口
|
||||
# ---------------------------------------------------------------------------
|
||||
async def run_graph(graph, query, db, user_id, conv_id, trace_id=None):
|
||||
"""执行编排图，返回 (answer: str, trace: dict)。
|
||||
|
||||
- 拓扑分层后逐层 asyncio.gather 并行执行。
|
||||
- 节点异常按 on_error=halt|continue 处理（默认 continue，失败时下游拿到错误作为输入）。
|
||||
"""
|
||||
validate_graph(graph)
|
||||
layers = topological_layers(graph)
|
||||
blackboard: dict = {}
|
||||
trace = {"nodes": [], "trace_id": trace_id, "layers": layers}
|
||||
|
||||
for layer in layers:
|
||||
tasks = [
|
||||
_run_node(_node_by_id(graph, nid), graph, blackboard, query, user_id, conv_id, trace_id)
|
||||
for nid in layer
|
||||
]
|
||||
results = await asyncio.gather(*tasks, return_exceptions=True)
|
||||
for nid, res in zip(layer, results):
|
||||
node = _node_by_id(graph, nid)
|
||||
if isinstance(res, Exception):
|
||||
on_error = node.get("on_error", "continue")
|
||||
err_msg = str(res)
|
||||
blackboard[nid] = {
|
||||
"ok": False,
|
||||
"node_id": nid,
|
||||
"node_type": node["type"],
|
||||
"error": err_msg,
|
||||
}
|
||||
trace["nodes"].append({"id": nid, "type": node["type"], "ok": False, "error": err_msg})
|
||||
logger.warning(f"[orchestrator] 节点 {nid} 失败 on_error={on_error}: {err_msg}")
|
||||
if on_error == "halt":
|
||||
raise OrchestrationError(f"节点 {nid} 执行失败(halt): {err_msg}")
|
||||
else:
|
||||
blackboard[nid] = res
|
||||
trace["nodes"].append({
|
||||
"id": nid,
|
||||
"type": res.get("node_type"),
|
||||
"ok": True,
|
||||
"used_tools": res.get("used_tools"),
|
||||
})
|
||||
|
||||
output_node = _find_output_node(graph)
|
||||
final = blackboard.get(output_node, {})
|
||||
if final.get("ok"):
|
||||
answer = final.get("output", "")
|
||||
else:
|
||||
answer = f"⚠️ 编排输出节点 {output_node} 执行失败: {final.get('error', '未知错误')}"
|
||||
trace["output_node"] = output_node
|
||||
return answer, trace
|
||||
+131
-131
@@ -1,131 +1,131 @@
|
||||
"""
|
||||
OpenTelemetry 桥接 (Phase 1.I 标准化)
|
||||
=====================================
|
||||
设计基准（标准对照表见 STANDARDS_AUDIT_2026-07-24.md §I）：
|
||||
- 标准：OpenTelemetry SDK + OTLP gRPC exporter —— 云原生可观测事实标准。
|
||||
- traces：span 经 observability.span() 嵌套导出（见 observability.py）。
|
||||
- metrics：MeterProvider 经 OTLP 导出 span 耗时直方图 + 调用计数，供 Grafana 出图。
|
||||
- 失败安全（生产红线）：若 OTLP 端点不可达 / 初始化异常，自动降级为 no-op，
|
||||
绝不阻断业务请求、不产生未捕获异常。
|
||||
- 门控：OTEL_ENABLED=1 默认开启；置 0 则零开销、不走 OTel。
|
||||
- 发射目标：OTEL_EXPORTER_OTLP_ENDPOINT，默认 http://qihu-aiflow-alloy:4317
|
||||
（经 Grafana Alloy 桥接 OTLP→Prometheus，再由 Prometheus 喂给 Grafana 出图）。
|
||||
|
||||
对外 API：
|
||||
init_otel() 首次调用时构建 TracerProvider + MeterProvider（幂等）
|
||||
get_tracer() 返回 tracer（未启用/失败则 None）
|
||||
get_meter() 返回 meter（未启用/失败则 None）
|
||||
current_span_cm(name) 返回一个 OTel span 的上下文管理器（供 observability 嵌套）
|
||||
record_span_metrics(name, elapsed_ms, status) 记录 span 指标（耗时直方图 + 计数）
|
||||
"""
|
||||
|
||||
import os
|
||||
import logging
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
_ENABLED = os.getenv("OTEL_ENABLED", "1") not in ("0", "false", "False", "")
|
||||
_ENDPOINT = os.getenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://qihu-aiflow-alloy:4317").rstrip("/")
|
||||
if not _ENDPOINT.startswith("http://") and not _ENDPOINT.startswith("https://"):
|
||||
_ENDPOINT = "http://" + _ENDPOINT
|
||||
|
||||
_tracer = None
|
||||
_meter = None
|
||||
_initialized = False
|
||||
_duration_hist = None
|
||||
_span_counter = None
|
||||
|
||||
|
||||
def init_otel() -> None:
|
||||
"""构建 TracerProvider + MeterProvider（幂等）。失败则降级为 no-op。"""
|
||||
global _tracer, _meter, _initialized
|
||||
if _initialized:
|
||||
return
|
||||
_initialized = True
|
||||
if not _ENABLED:
|
||||
logger.info("[otel] disabled (OTEL_ENABLED=0)")
|
||||
return
|
||||
try:
|
||||
from opentelemetry import trace, metrics
|
||||
from opentelemetry.sdk.resources import Resource
|
||||
from opentelemetry.semconv.resource import ResourceAttributes
|
||||
from opentelemetry.sdk.trace import TracerProvider
|
||||
from opentelemetry.sdk.trace.export import BatchSpanProcessor
|
||||
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
|
||||
OTLPSpanExporter,
|
||||
)
|
||||
from opentelemetry.sdk.metrics import MeterProvider
|
||||
from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
|
||||
from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import (
|
||||
OTLPMetricExporter,
|
||||
)
|
||||
|
||||
resource = Resource.create({ResourceAttributes.SERVICE_NAME: "qihu-aiflow"})
|
||||
|
||||
# --- traces ---
|
||||
trace_provider = TracerProvider(resource=resource)
|
||||
trace_provider.add_span_processor(
|
||||
BatchSpanProcessor(OTLPSpanExporter(endpoint=_ENDPOINT, insecure=True))
|
||||
)
|
||||
trace.set_tracer_provider(trace_provider)
|
||||
_tracer = trace.get_tracer("qihu-aiflow")
|
||||
|
||||
# --- metrics ---
|
||||
metric_reader = PeriodicExportingMetricReader(
|
||||
OTLPMetricExporter(endpoint=_ENDPOINT, insecure=True)
|
||||
)
|
||||
meter_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
|
||||
metrics.set_meter_provider(meter_provider)
|
||||
_meter = metrics.get_meter("qihu-aiflow")
|
||||
|
||||
logger.info(f"[otel] tracer+meter ready -> {_ENDPOINT}")
|
||||
except Exception as e: # pragma: no cover - 降级保护
|
||||
logger.warning(f"[otel] init failed, degraded to no-op: {e}")
|
||||
_tracer = None
|
||||
_meter = None
|
||||
|
||||
|
||||
def get_tracer():
|
||||
"""返回 tracer；首次调用触发 init。未启用/失败返回 None。"""
|
||||
if _tracer is None and _ENABLED:
|
||||
init_otel()
|
||||
return _tracer
|
||||
|
||||
|
||||
def get_meter():
|
||||
"""返回 meter；首次调用触发 init。未启用/失败返回 None。"""
|
||||
if _meter is None and _ENABLED:
|
||||
init_otel()
|
||||
return _meter
|
||||
|
||||
|
||||
def current_span_cm(name: str, attributes: dict | None = None):
|
||||
"""返回一个 OTel span 的上下文管理器（同步/异步通用）。失败时返回 None。"""
|
||||
tr = get_tracer()
|
||||
if tr is None:
|
||||
return None
|
||||
try:
|
||||
return tr.start_as_current_span(name, attributes=attributes or {})
|
||||
except Exception as e: # pragma: no cover
|
||||
logger.warning(f"[otel] start_span failed: {e}")
|
||||
return None
|
||||
|
||||
|
||||
def record_span_metrics(name: str, elapsed_ms: float, status: str) -> None:
|
||||
"""记录 span 指标：耗时直方图 + 调用计数（供 Grafana 出图）。失败安全。"""
|
||||
m = get_meter()
|
||||
if m is None:
|
||||
return
|
||||
global _duration_hist, _span_counter
|
||||
try:
|
||||
if _duration_hist is None:
|
||||
_duration_hist = m.create_histogram(
|
||||
"aiflow_span_duration_ms", "Span duration in milliseconds"
|
||||
)
|
||||
_span_counter = m.create_counter("aiflow_span_count", "Span invocation count")
|
||||
attrs = {"span": name, "status": status}
|
||||
_duration_hist.record(elapsed_ms, attrs)
|
||||
_span_counter.add(1, attrs)
|
||||
except Exception: # pragma: no cover
|
||||
pass
|
||||
|
||||
"""
|
||||
OpenTelemetry 桥接 (Phase 1.I 标准化)
|
||||
=====================================
|
||||
设计基准（标准对照表见 STANDARDS_AUDIT_2026-07-24.md §I）：
|
||||
- 标准：OpenTelemetry SDK + OTLP gRPC exporter —— 云原生可观测事实标准。
|
||||
- traces：span 经 observability.span() 嵌套导出（见 observability.py）。
|
||||
- metrics：MeterProvider 经 OTLP 导出 span 耗时直方图 + 调用计数，供 Grafana 出图。
|
||||
- 失败安全（生产红线）：若 OTLP 端点不可达 / 初始化异常，自动降级为 no-op，
|
||||
绝不阻断业务请求、不产生未捕获异常。
|
||||
- 门控：OTEL_ENABLED=1 默认开启；置 0 则零开销、不走 OTel。
|
||||
- 发射目标：OTEL_EXPORTER_OTLP_ENDPOINT，默认 http://qihu-aiflow-alloy:4317
|
||||
（经 Grafana Alloy 桥接 OTLP→Prometheus，再由 Prometheus 喂给 Grafana 出图）。
|
||||
|
||||
对外 API：
|
||||
init_otel() 首次调用时构建 TracerProvider + MeterProvider（幂等）
|
||||
get_tracer() 返回 tracer（未启用/失败则 None）
|
||||
get_meter() 返回 meter（未启用/失败则 None）
|
||||
current_span_cm(name) 返回一个 OTel span 的上下文管理器（供 observability 嵌套）
|
||||
record_span_metrics(name, elapsed_ms, status) 记录 span 指标（耗时直方图 + 计数）
|
||||
"""
|
||||
|
||||
import os
|
||||
import logging
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
_ENABLED = os.getenv("OTEL_ENABLED", "1") not in ("0", "false", "False", "")
|
||||
_ENDPOINT = os.getenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://qihu-aiflow-alloy:4317").rstrip("/")
|
||||
if not _ENDPOINT.startswith("http://") and not _ENDPOINT.startswith("https://"):
|
||||
_ENDPOINT = "http://" + _ENDPOINT
|
||||
|
||||
_tracer = None
|
||||
_meter = None
|
||||
_initialized = False
|
||||
_duration_hist = None
|
||||
_span_counter = None
|
||||
|
||||
|
||||
def init_otel() -> None:
|
||||
"""构建 TracerProvider + MeterProvider（幂等）。失败则降级为 no-op。"""
|
||||
global _tracer, _meter, _initialized
|
||||
if _initialized:
|
||||
return
|
||||
_initialized = True
|
||||
if not _ENABLED:
|
||||
logger.info("[otel] disabled (OTEL_ENABLED=0)")
|
||||
return
|
||||
try:
|
||||
from opentelemetry import trace, metrics
|
||||
from opentelemetry.sdk.resources import Resource
|
||||
from opentelemetry.semconv.resource import ResourceAttributes
|
||||
from opentelemetry.sdk.trace import TracerProvider
|
||||
from opentelemetry.sdk.trace.export import BatchSpanProcessor
|
||||
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
|
||||
OTLPSpanExporter,
|
||||
)
|
||||
from opentelemetry.sdk.metrics import MeterProvider
|
||||
from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
|
||||
from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import (
|
||||
OTLPMetricExporter,
|
||||
)
|
||||
|
||||
resource = Resource.create({ResourceAttributes.SERVICE_NAME: "qihu-aiflow"})
|
||||
|
||||
# --- traces ---
|
||||
trace_provider = TracerProvider(resource=resource)
|
||||
trace_provider.add_span_processor(
|
||||
BatchSpanProcessor(OTLPSpanExporter(endpoint=_ENDPOINT, insecure=True))
|
||||
)
|
||||
trace.set_tracer_provider(trace_provider)
|
||||
_tracer = trace.get_tracer("qihu-aiflow")
|
||||
|
||||
# --- metrics ---
|
||||
metric_reader = PeriodicExportingMetricReader(
|
||||
OTLPMetricExporter(endpoint=_ENDPOINT, insecure=True)
|
||||
)
|
||||
meter_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
|
||||
metrics.set_meter_provider(meter_provider)
|
||||
_meter = metrics.get_meter("qihu-aiflow")
|
||||
|
||||
logger.info(f"[otel] tracer+meter ready -> {_ENDPOINT}")
|
||||
except Exception as e: # pragma: no cover - 降级保护
|
||||
logger.warning(f"[otel] init failed, degraded to no-op: {e}")
|
||||
_tracer = None
|
||||
_meter = None
|
||||
|
||||
|
||||
def get_tracer():
|
||||
"""返回 tracer；首次调用触发 init。未启用/失败返回 None。"""
|
||||
if _tracer is None and _ENABLED:
|
||||
init_otel()
|
||||
return _tracer
|
||||
|
||||
|
||||
def get_meter():
|
||||
"""返回 meter；首次调用触发 init。未启用/失败返回 None。"""
|
||||
if _meter is None and _ENABLED:
|
||||
init_otel()
|
||||
return _meter
|
||||
|
||||
|
||||
def current_span_cm(name: str, attributes: dict | None = None):
|
||||
"""返回一个 OTel span 的上下文管理器（同步/异步通用）。失败时返回 None。"""
|
||||
tr = get_tracer()
|
||||
if tr is None:
|
||||
return None
|
||||
try:
|
||||
return tr.start_as_current_span(name, attributes=attributes or {})
|
||||
except Exception as e: # pragma: no cover
|
||||
logger.warning(f"[otel] start_span failed: {e}")
|
||||
return None
|
||||
|
||||
|
||||
def record_span_metrics(name: str, elapsed_ms: float, status: str) -> None:
|
||||
"""记录 span 指标：耗时直方图 + 调用计数（供 Grafana 出图）。失败安全。"""
|
||||
m = get_meter()
|
||||
if m is None:
|
||||
return
|
||||
global _duration_hist, _span_counter
|
||||
try:
|
||||
if _duration_hist is None:
|
||||
_duration_hist = m.create_histogram(
|
||||
"aiflow_span_duration_ms", "Span duration in milliseconds"
|
||||
)
|
||||
_span_counter = m.create_counter("aiflow_span_count", "Span invocation count")
|
||||
attrs = {"span": name, "status": status}
|
||||
_duration_hist.record(elapsed_ms, attrs)
|
||||
_span_counter.add(1, attrs)
|
||||
except Exception: # pragma: no cover
|
||||
pass
|
||||
|
||||
Reference in new issue
Block a user