unexpected connection_lost() call when cancelling a direct_tls connection with an active request can lead to a connection leak
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 35/100
调研方向
首先,针对 direct_tls PostgreSQL 设置运行提供的 asyncio 复现程序,并在取消期间跟踪 asyncpg.connect()、Connection.close() 和 terminate()。完成的标准是:取消一个活动请求会关闭 direct_tls 连接,且不会意外调用 connection_lost(),同时 PostgreSQL 不再保留这些会话。
由索引模型根据 Issue 内容生成。
描述
- asyncpg version: 0.30.0
- PostgreSQL version: 16
- Do you use a PostgreSQL SaaS? If so, which? Can you reproduce
the issue with a local PostgreSQL install?: CloudSQL with cloud sql connector, but it is possible to reproduce it locally with a direct_tls setup. - Python version: 3.12
- Platform: linux / macos
- Do you use pgbouncer?: no
- Did you install asyncpg with pip?: poetry
- If you built asyncpg locally, which version of Cython did you use?:
- Can the issue be reproduced under both asyncio and
uvloop?: the reproducer uses asyncio, but I hit the problem "in prod" with uvicorn / uvloop
Hi, I reported the issue first to sqlalchemy (https://github.com/sqlalchemy/sqlalchemy/issues/12099) but managed to reproduce it using only asyncpg.
From what I could observe, if a direct_tls connection is waiting for postgres to reply and the task within which this connection lives is cancelled, calling connection.close() in an exception handler will fail with unexpected connection_lost() call, leaving the connection open. (a call to connection.terminate() after that doesn't close the connection either)
The problem only occurs with direct_tls=True.
The issue was first observed with sqlalchemy with a connection pool of asyncpg connections.
SA thought the connections were closed and would then open new ones, this lead to using up all the available slots on the pg side.
This script can reproduce the issue, however, it needs a postgres setup where direct_tls can be used.
There a docker-compose file at https://github.com/brittlesoft/repro-starlette-sa-conn-leak that can be used to get a working setup quickly.
import asyncio
import asyncpg
import logging
logging.basicConfig(level=logging.DEBUG)
async def do(i):
try:
# connect using direct_tls leads to `unexpected connection_lost() call` when calling conn.close()
conn = await asyncpg.connect('postgresql://postgres:postgres@localhost:5443/postgres', direct_tls=True)
# connect using default params works fine
#conn = await asyncpg.connect('postgresql://postgres:postgres@localhost:5432/postgres')
# lock and simulate work (or block if lock already taken) -- using select pg_sleep(10) would also work
await conn.execute("select pg_advisory_lock(1234)")
await asyncio.sleep(10)
except BaseException as e:
print(i,"got exc:", e, type(e))
try:
await conn.close(timeout=2)
except BaseException as e:
print(i, "close got exc: ",e)
# NOTE: aborting transport here seems to release the connection
#conn._transport.abort()
try:
print(i, "calling terminate")
conn.terminate()
except BaseException as e:
print(i, "terminate got exc: ",e)
async def main():
ts = []
for i in range(10):
ts.append(asyncio.create_task(do(i)))
async def timeouter():
await asyncio.sleep(1)
for t in ts:
t.cancel()
asyncio.create_task(timeouter())
try:
await asyncio.gather(*ts)
except asyncio.CancelledError:
print("cancelled")
# Sleep so we can observe the state of the connections
await asyncio.sleep(30)
if __name__ == '__main__':
asyncio.run(main())
output:
DEBUG:asyncio:Using selector: KqueueSelector
0 got exc: <class 'asyncio.exceptions.CancelledError'>
1 got exc: <class 'asyncio.exceptions.CancelledError'>
2 got exc: <class 'asyncio.exceptions.CancelledError'>
3 got exc: <class 'asyncio.exceptions.CancelledError'>
4 got exc: <class 'asyncio.exceptions.CancelledError'>
5 got exc: <class 'asyncio.exceptions.CancelledError'>
6 got exc: <class 'asyncio.exceptions.CancelledError'>
7 got exc: <class 'asyncio.exceptions.CancelledError'>
8 got exc: <class 'asyncio.exceptions.CancelledError'>
9 got exc: <class 'asyncio.exceptions.CancelledError'>
1 close got exc: unexpected connection_lost() call
1 calling terminate
7 close got exc: unexpected connection_lost() call
7 calling terminate
0 close got exc: unexpected connection_lost() call
0 calling terminate
4 close got exc: unexpected connection_lost() call
4 calling terminate
5 close got exc: unexpected connection_lost() call
5 calling terminate
3 close got exc: unexpected connection_lost() call
3 calling terminate
9 close got exc: unexpected connection_lost() call
9 calling terminate
6 close got exc: unexpected connection_lost() call
6 calling terminate
8 close got exc: unexpected connection_lost() call
8 calling terminate
And it postgres we see this:
2024-11-15 13:29:02.164078+00 | 2024-11-15 13:29:02.164078+00 | 2024-11-15 13:29:02.164078+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.188851+00 | 2024-11-15 13:29:02.188851+00 | 2024-11-15 13:29:02.188852+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
| 2024-11-15 13:29:02.163992+00 | 2024-11-15 13:29:03.04934+00 | Client | ClientRead | idle | | | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.163992+00 | 2024-11-15 13:29:02.163992+00 | 2024-11-15 13:29:02.163994+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.189558+00 | 2024-11-15 13:29:02.189558+00 | 2024-11-15 13:29:02.189558+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.188884+00 | 2024-11-15 13:29:02.188884+00 | 2024-11-15 13:29:02.188885+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.188826+00 | 2024-11-15 13:29:02.188826+00 | 2024-11-15 13:29:02.188829+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.189478+00 | 2024-11-15 13:29:02.189478+00 | 2024-11-15 13:29:02.189479+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
2024-11-15 13:29:02.19116+00 | 2024-11-15 13:29:02.19116+00 | 2024-11-15 13:29:02.191161+00 | Lock | advisory | active | | 1584 | | select pg_advisory_lock(1234) | client backend
- 主要语言
- Python
- 星标
- 8.1k
- 派生
- 474
- 平均合并
- 3 天 14 小时
- 30 天内合并 PR
- 17
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
MagicStack/asyncpg 的其他 Issue
-
难度 3/5 1-2 天 新手友好度 72/100
MagicStack/asyncpg#1342 ·
维护者通常 1 天内回复
-
难度 3/5 1-2 天 新手友好度 56/100
MagicStack/asyncpg#1340 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 28/100
MagicStack/asyncpg#1337 ·
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 42/100
MagicStack/asyncpg#1330 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 38/100
MagicStack/asyncpg#1322 ·
维护者通常 1 天内回复
查看 MagicStack/asyncpg 的全部 Issue
相似的 Issue
-
comp/tools duplicate P2 sweeper:risk-compatibility tool/mcp type/bug
难度 1/5 1 小时以内 新手友好度 88/100
NousResearch/hermes-agent#132042 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
deepset-ai/haystack#13092 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 86/100
EverMind-AI/Raven#845 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
AstrBotDevs/AstrBot#10340 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
BasedHardware/omi#20401 · 1 条评论 ·
维护者通常 1 天内回复