Files
ragflow/api/db/services/canvas_service.py
tmimmanuel 6ce014c23b fix: offload blocking DB/Redis calls to thread pool for high-concurrency support (#13825) (#13941)
### What problem does this PR solve?

Addresses event-loop blocking under high concurrency reported in #13825.
When multiple requests hit the API simultaneously, synchronous DB/Redis
calls block the async event loop, preventing Quart from handling other
requests and causing cascading 502/504 timeouts.

This PR wraps all remaining blocking DB/Redis calls in `canvas_app.py`,
`chat_api.py`, `session.py`, and `canvas_service.py` with `await
thread_pool_exec()`
- Offload all synchronous `Service.*`, `REDIS_CONN.*`, and
`APIToken.query` calls to the thread pool
- Convert sync endpoint handlers (`list_chats`, `get_chat`, `templates`,
`sessions`, etc.) to `async def`
- Convert sync helper functions (`_ensure_owned_chat`,
`_validate_llm_id`, `_validate_dataset_ids`, etc.) to async - no
duplicate sync/async pairs
- Wrap `CanvasReplicaService` Redis IO calls (`bootstrap`,
`replace_for_set`, `commit_after_run`)
- Use `asyncio.gather()` for concurrent file uploads and chat response
building

**Note:** This fixes the code-level event-loop blocking, which is a
prerequisite for handling concurrent requests. For the full "30
concurrent requests without 502/504" goal described in the issue, users
should also tune deployment config:
- `WS=4` or higher (HTTP worker processes, default 1)
- `MAX_CONCURRENT_CHATS=50` (default 10)
- `SANDBOX_EXECUTOR_MANAGER_POOL_SIZE` for workflow-heavy workloads

### Performance verification

Reviewer asked for a before-vs-after comparison
([comment](https://github.com/infiniflow/ragflow/pull/13941#issuecomment-4393667231)).
I built a self-contained microbenchmark that reproduces the exact
failure mode this PR targets: an async handler that performs blocking
DB/Redis-style calls (50 ms each, 3 per request, 30 concurrent requests)
is run twice — once with the pre-PR pattern (sync call directly inside
the async handler) and once with the post-PR pattern (`await
thread_pool_exec(...)`). The benchmark imports nothing from RAGFlow
except `thread_pool_exec` itself, so it is hermetic and reproducible
(`THREAD_POOL_MAX_WORKERS=128`, Python 3.13.12).

**Throughput — wall-clock for 30 concurrent requests (lower is better)**

| flavour | wall(s) | p50(s) | p95(s) | max(s) |
|---|---:|---:|---:|---:|
| before | 4.986 | 0.158 | 0.207 | 0.269 |
| after  | 0.248 | 0.181 | 0.230 | 0.231 |

The pre-PR handler serializes the entire load on the event-loop thread,
so 30 × 3 × 50 ms ≈ 4.5 s shows up as the wall time. The post-PR handler
parallelizes the blocking work across the thread pool and finishes the
same load in 248 ms — a **~20× speedup** on this workload.

**Event-loop responsiveness — latency of an unrelated probe coroutine
while the 30 slow requests are running (lower is better)**

| flavour | samples | probe p50 (ms) | probe p95 (ms) | probe max (ms) |
|---|---:|---:|---:|---:|
| before | 1 | 5442.26 | 5442.26 | 5442.26 |
| after  | 28 | 0.88 | 11.53 | 98.02 |

This is the metric that maps directly to "the API still answers other
requests while one is busy". A 5 ms-interval probe was scheduled while
the 30 slow handlers ran. With the pre-PR code the event loop was frozen
for the entire duration of the blocking work, so only one probe sample
was ever picked up and it waited **5,442 ms**. After the PR, 28 probe
samples landed with **p50 0.88 ms / p95 11.53 ms**, meaning unrelated
requests are no longer starved by the slow ones. That is the regression
mode behind the cascading 502/504s reported in #13825.

<details>
<summary>Raw benchmark output</summary>

```
config: 30 concurrent requests, 3 blocking calls of 50ms each per request, THREAD_POOL_MAX_WORKERS=128

=== Throughput (lower wall is better) ===
flavour       wall(s)     p50(s)     p95(s)     max(s)
before          4.986      0.158      0.207      0.269
after           0.248      0.181      0.230      0.231

=== Event-loop responsiveness (lower probe latency is better) ===
flavour       samples    probe p50(ms)    probe p95(ms)    probe max(ms)
before              1          5442.26          5442.26          5442.26
after              28             0.88            11.53            98.02
```

</details>

The benchmark script is included as a comment on the PR for
reproducibility.

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] Performance Improvement

Closes [#13825](https://github.com/infiniflow/ragflow/issues/13825)

---------

Co-authored-by: tmimmanuel <tmimmanuel@users.noreply.github.com>
Co-authored-by: Kevin Hu <kevinhu.sh@gmail.com>
2026-05-11 15:08:55 +08:00

407 lines
15 KiB
Python

#
# Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import json
import logging
import time
from uuid import uuid4
from agent.canvas import Canvas
from api.db import CanvasCategory, TenantPermission
from api.db.db_models import DB, CanvasTemplate, User, UserCanvas, API4Conversation, UserCanvasVersion
from api.db.services.api_service import API4ConversationService
from api.db.services.common_service import CommonService
from api.db.services.user_canvas_version import UserCanvasVersionService
from common.misc_utils import get_uuid, thread_pool_exec
from api.utils.api_utils import get_data_openai
import tiktoken
from peewee import fn
class CanvasTemplateService(CommonService):
model = CanvasTemplate
class DataFlowTemplateService(CommonService):
"""
Alias of CanvasTemplateService
"""
model = CanvasTemplate
class UserCanvasService(CommonService):
model = UserCanvas
@classmethod
@DB.connection_context()
def get_list(cls, tenant_id,
page_number, items_per_page, orderby, desc, id, title, canvas_category=CanvasCategory.Agent):
agents = cls.model.select()
if id:
agents = agents.where(cls.model.id == id)
if title:
agents = agents.where(cls.model.title == title)
agents = agents.where(cls.model.user_id == tenant_id)
agents = agents.where(cls.model.canvas_category == canvas_category)
if desc:
agents = agents.order_by(cls.model.getter_by(orderby).desc())
else:
agents = agents.order_by(cls.model.getter_by(orderby).asc())
agents = agents.paginate(page_number, items_per_page)
return list(agents.dicts())
@classmethod
@DB.connection_context()
def get_all_agents_by_tenant_ids(cls, tenant_ids, user_id):
# will get all permitted agents, be cautious
fields = [
cls.model.id,
cls.model.avatar,
cls.model.title,
cls.model.permission,
cls.model.canvas_type,
cls.model.canvas_category
]
# find team agents and owned agents
agents = cls.model.select(*fields).where(
(cls.model.user_id.in_(tenant_ids) & (cls.model.permission == TenantPermission.TEAM.value)) | (
cls.model.user_id == user_id
)
)
# sort by create_time, asc
agents.order_by(cls.model.create_time.asc())
# maybe cause slow query by deep paginate, optimize later
offset, limit = 0, 50
res = []
while True:
ag_batch = agents.offset(offset).limit(limit)
_temp = list(ag_batch.dicts())
if not _temp:
break
res.extend(_temp)
offset += limit
return res
@classmethod
@DB.connection_context()
def get_by_canvas_id(cls, pid):
try:
fields = [
cls.model.id,
cls.model.avatar,
cls.model.title,
cls.model.dsl,
cls.model.description,
cls.model.permission,
cls.model.update_time,
cls.model.user_id,
cls.model.create_time,
cls.model.create_date,
cls.model.update_date,
cls.model.canvas_category,
User.nickname,
User.avatar.alias('tenant_avatar'),
]
agents = cls.model.select(*fields) \
.join(User, on=(cls.model.user_id == User.id)) \
.where(cls.model.id == pid)
# obj = cls.model.query(id=pid)[0]
return True, agents.dicts()[0]
except Exception as e:
logging.exception(e)
return False, None
@classmethod
@DB.connection_context()
def get_basic_info_by_canvas_ids(cls, canvas_id):
fields = [
cls.model.id,
cls.model.avatar,
cls.model.user_id,
cls.model.title,
cls.model.permission,
cls.model.canvas_category
]
return cls.model.select(*fields).where(cls.model.id.in_(canvas_id)).dicts()
@classmethod
@DB.connection_context()
def get_by_tenant_ids(
cls,
joined_tenant_ids,
user_id,
page_number,
items_per_page,
orderby,
desc,
keywords,
canvas_category=None,
):
fields = [
cls.model.id,
cls.model.avatar,
cls.model.title,
cls.model.description,
cls.model.permission,
cls.model.user_id.alias("tenant_id"),
User.nickname,
User.avatar.alias('tenant_avatar'),
cls.model.update_time,
cls.model.canvas_category,
]
if keywords:
agents = cls.model.select(*fields).join(User, on=(cls.model.user_id == User.id)).where(
(((cls.model.user_id.in_(joined_tenant_ids)) & (cls.model.permission == TenantPermission.TEAM.value)) | (cls.model.user_id == user_id)),
(fn.LOWER(cls.model.title).contains(keywords.lower()))
)
else:
agents = cls.model.select(*fields).join(User, on=(cls.model.user_id == User.id)).where(
(((cls.model.user_id.in_(joined_tenant_ids)) & (cls.model.permission == TenantPermission.TEAM.value)) | (cls.model.user_id == user_id))
)
if canvas_category:
agents = agents.where(cls.model.canvas_category == canvas_category)
if desc:
agents = agents.order_by(cls.model.getter_by(orderby).desc())
else:
agents = agents.order_by(cls.model.getter_by(orderby).asc())
count = agents.count()
if page_number and items_per_page:
agents = agents.paginate(page_number, items_per_page)
agents_list = list(agents.dicts())
# Get latest release time for each canvas
if agents_list:
canvas_ids = [a['id'] for a in agents_list]
release_times = (
UserCanvasVersion.select(UserCanvasVersion.user_canvas_id, fn.MAX(UserCanvasVersion.create_time).alias("release_time"))
.where((UserCanvasVersion.user_canvas_id.in_(canvas_ids)) & (UserCanvasVersion.release))
.group_by(UserCanvasVersion.user_canvas_id)
)
release_time_map = {r.user_canvas_id: r.release_time for r in release_times}
for agent in agents_list:
agent['release_time'] = release_time_map.get(agent['id'])
return agents_list, count
@classmethod
@DB.connection_context()
def accessible(cls, canvas_id, tenant_id):
from api.db.services.user_service import UserTenantService
e, c = UserCanvasService.get_by_canvas_id(canvas_id)
if not e:
return False
tids = [t.tenant_id for t in UserTenantService.query(user_id=tenant_id)]
if c["user_id"] == tenant_id:
return True
if c["user_id"] not in tids:
return False
if c["permission"] != TenantPermission.TEAM.value:
return False
return True
@classmethod
def get_agent_dsl_with_release(cls, agent_id, release_mode=False, tenant_id=None):
e, cvs = cls.get_by_id(agent_id)
if not e:
raise LookupError("Agent not found.")
if release_mode:
released_version = UserCanvasVersionService.get_latest_released(agent_id)
if not released_version:
raise PermissionError("No available published version")
dsl = released_version.dsl
else:
dsl = cvs.dsl
if not isinstance(dsl, str):
dsl = json.dumps(dsl, ensure_ascii=False)
return cvs, dsl
async def completion(tenant_id, agent_id, session_id=None, **kwargs):
query = kwargs.get("query", "") or kwargs.get("question", "")
files = kwargs.get("files", [])
inputs = kwargs.get("inputs", {})
user_id = kwargs.get("user_id", "")
custom_header = kwargs.get("custom_header", "")
release_mode = str(kwargs.get("release", "")).strip().lower()
if session_id:
e, conv = await thread_pool_exec(API4ConversationService.get_by_id, session_id)
if not e:
raise LookupError("Session not found!")
if not conv.message:
conv.message = []
if not isinstance(conv.dsl, str):
conv.dsl = json.dumps(conv.dsl, ensure_ascii=False)
canvas = Canvas(conv.dsl, tenant_id, agent_id, canvas_id=agent_id, custom_header=custom_header)
else:
cvs, dsl = await thread_pool_exec(UserCanvasService.get_agent_dsl_with_release, agent_id, release_mode=release_mode == "true", tenant_id=tenant_id)
session_id = get_uuid()
canvas = Canvas(dsl, tenant_id, agent_id, canvas_id=cvs.id, custom_header=custom_header)
canvas.reset()
# Get the version title based on release_mode
version_title = await thread_pool_exec(UserCanvasVersionService.get_latest_version_title, cvs.id, release_mode=release_mode == "true")
conv = {"id": session_id, "dialog_id": cvs.id, "user_id": user_id, "message": [], "source": "agent", "dsl": dsl, "reference": [], "version_title": version_title}
await thread_pool_exec(API4ConversationService.save, **conv)
conv = API4Conversation(**conv)
message_id = str(uuid4())
conv.message.append({
"role": "user",
"content": query,
"id": message_id,
"files": files
})
txt = ""
async for ans in canvas.run(query=query, files=files, user_id=user_id, inputs=inputs):
ans["session_id"] = session_id
if ans["event"] == "message":
txt += ans["data"]["content"]
if ans["data"].get("start_to_think", False):
txt += "<think>"
elif ans["data"].get("end_to_think", False):
txt += "</think>"
yield "data:" + json.dumps(ans, ensure_ascii=False) + "\n\n"
conv.message.append({"role": "assistant", "content": txt, "created_at": time.time(), "id": message_id})
conv.reference = canvas.get_reference()
conv.errors = canvas.error
conv.dsl = str(canvas)
conv = conv.to_dict()
await thread_pool_exec(API4ConversationService.append_message, conv["id"], conv)
async def completion_openai(tenant_id, agent_id, question, session_id=None, stream=True, **kwargs):
tiktoken_encoder = tiktoken.get_encoding("cl100k_base")
prompt_tokens = len(tiktoken_encoder.encode(str(question)))
user_id = kwargs.get("user_id", "")
if stream:
completion_tokens = 0
try:
async for ans in completion(
tenant_id=tenant_id,
agent_id=agent_id,
session_id=session_id,
query=question,
user_id=user_id,
**kwargs
):
if isinstance(ans, str):
try:
ans = json.loads(ans[5:]) # remove "data:"
except Exception as e:
logging.exception(f"Agent OpenAI-Compatible completion_openai parse answer failed: {e}")
continue
if ans.get("event") not in ["message", "message_end"]:
continue
content_piece = ""
if ans["event"] == "message":
content_piece = ans["data"]["content"]
completion_tokens += len(tiktoken_encoder.encode(content_piece))
openai_data = get_data_openai(
id=session_id or str(uuid4()),
model=agent_id,
content=content_piece,
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
stream=True
)
if ans.get("data", {}).get("reference", None):
openai_data["choices"][0]["delta"]["reference"] = ans["data"]["reference"]
yield "data: " + json.dumps(openai_data, ensure_ascii=False) + "\n\n"
yield "data: [DONE]\n\n"
except Exception as e:
logging.exception(e)
yield "data: " + json.dumps(
get_data_openai(
id=session_id or str(uuid4()),
model=agent_id,
content=f"**ERROR**: {str(e)}",
finish_reason="stop",
prompt_tokens=prompt_tokens,
completion_tokens=len(tiktoken_encoder.encode(f"**ERROR**: {str(e)}")),
stream=True
),
ensure_ascii=False
) + "\n\n"
yield "data: [DONE]\n\n"
else:
try:
all_content = ""
reference = {}
async for ans in completion(
tenant_id=tenant_id,
agent_id=agent_id,
session_id=session_id,
query=question,
user_id=user_id,
**kwargs
):
if isinstance(ans, str):
ans = json.loads(ans[5:])
if ans.get("event") not in ["message", "message_end"]:
continue
if ans["event"] == "message":
all_content += ans["data"]["content"]
if ans.get("data", {}).get("reference", None):
reference.update(ans["data"]["reference"])
completion_tokens = len(tiktoken_encoder.encode(all_content))
openai_data = get_data_openai(
id=session_id or str(uuid4()),
model=agent_id,
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
content=all_content,
finish_reason="stop",
param=None
)
if reference:
openai_data["choices"][0]["message"]["reference"] = reference
yield openai_data
except Exception as e:
logging.exception(e)
yield get_data_openai(
id=session_id or str(uuid4()),
model=agent_id,
prompt_tokens=prompt_tokens,
completion_tokens=len(tiktoken_encoder.encode(f"**ERROR**: {str(e)}")),
content=f"**ERROR**: {str(e)}",
finish_reason="stop",
param=None
)