Segmentation Fault MacOS with Python 3.13

[MCubek]

1. What versions are you using?

Oracle ATP 19.c

platform.platform: macOS-15.4-arm64-arm-64bit-Mach-O
sys.maxsize > 2**32: True
platform.python_version: 3.13.3

oracledb.version: 3.1.0

2. Is it an error or a hang or a crash?

Segmentation Fault Crash

3. What error(s) or behavior you are seeing?

https://gist.github.com/MCubek/ffa2806d8bf0a8e73fe3ac4436d888ed

This issue does not happen on Python 3.12.

4. Does your application call init_oracle_client()?

Yes, thick mode is used.

5. Include a runnable Python script that shows the problem.

Test script below couldn't reproduce the issue but represents how the pool is created...

# Minimal settings (replace placeholders)
username = "your_user"
password = "your_password"
dsn = "your_dsn" 
wallet_path = "/path/to/your/oracle/wallet"
instant_client_path = "/path/to/your/instantclient"


if __name__ == "__main__":
    # Initialize Oracle client (thick mode)
    print("Initializing Oracle Client...")
    oracledb.init_oracle_client(lib_dir=instant_client_path, config_dir=wallet_path)

    # Attempt creating a connection pool (crashes here on Python 3.13)
    print("Creating Oracle connection pool...")
    pool = oracledb.create_pool(
        user=username,
        password=password,
        config_dir=wallet_path,
        dsn=dsn,
        wallet_location=wallet_path,
        wallet_password="",
        min=1,
        max=2,
        increment=1,
    )

    print("Acquiring connection from pool...")
    with pool.acquire() as connection:
        print("Oracle DB version:", connection.version)

    print("Test completed successfully (if no crash).")

In production FastAPI is used to create pool as shown in snippets:

@asynccontextmanager
async def lifespan(app: FastAPI):
    # Create the connection pool
    app.state.pool = create_db_connection_pool()
    yield
    # Close the connection pool on shutdown
    app.state.pool.close()
    logger.debug("Closed Oracle database connection pool.")

if os.getenv("ENV_FOR_DYNACONF") != "test":
    # noinspection PyUnresolvedReferences
    oracledb.init_oracle_client(
        lib_dir=settings.db.instant_client_library, config_dir=wallet_path
    )


def create_db_connection_pool() -> oracledb.ConnectionPool:
    logger.debug("Creating Oracle database connection pool...")
    pool = oracledb.create_pool(
        user=username,
        password=password,
        config_dir=wallet_path,
        dsn=dsn,
        wallet_location=wallet_path,
        wallet_password="",
        min=1,  # Minimum number of connections in the pool
        max=5,  # Maximum number of connections in the pool
        increment=1,  # Number of connections to add when more are needed
    )
    logger.debug("Created Oracle database connection pool.")
    return cast("oracledb.ConnectionPool", pool)

[cjbj]

We'll take a look. I fear this will be tricky to do anything with unless you can get a reproducible case, but maybe there is something obvious in the driver that can be tweaked. Do you know if it also happens on other platforms (e.g. Linux)? Is there any other info available about when it does / doesn't reproduce, e.g with different pool settings?

[MCubek]

I have forgotten to note that this occurs about half to 2/3 of the time when starting the FastAPI webserver, not each time.
I will try test a bit without connection pooling and with different environments (eg. docker linux/amd64).

[MCubek]

Can't reproduce the issue under docker linux/amd64 environment, and under macos seems to work more often now then before for some reason but still fails 1/4 of the time with seg fault.
No luck reproducing the issue in a smaller snippet yet.

[MCubek]

Issue seems to originate from import of xgboost package.
from xgboost import XGBRegressor

When import exists in code and when running with debuger in pycharm more then half of the time the provided segmentation fault will occur. Other times runs just fine...

Xgboost library at version 3.0.0

[MCubek]

from contextlib import asynccontextmanager
from http.client import HTTPException
from typing import Annotated, AsyncGenerator, cast

from xgboost import XGBRegressor
import oracledb
from fastapi import Depends, FastAPI, Request, status

username = "username"
password = "password"
dsn = "dns"
wallet_path = "/opt/oracle/wallet"
instant_client_path = "/opt/oracle/instantclient_23_3"

oracledb.init_oracle_client(lib_dir=instant_client_path, config_dir=wallet_path)


def create_db_connection_pool() -> oracledb.ConnectionPool:
    print("Creating Oracle database connection pool...")
    pool = oracledb.create_pool(
        user=username,
        password=password,
        config_dir=wallet_path,
        dsn=dsn,
        wallet_location=wallet_path,
        wallet_password="",
        min=1,  # Minimum number of connections in the pool
        max=5,  # Maximum number of connections in the pool
        increment=1,  # Number of connections to add when more are needed
    )
    print("Created Oracle database connection pool.")
    return cast("oracledb.ConnectionPool", pool)


async def get_tenant_schema(
    request: Request,
) -> str:
    # Try to get tenant_schema from headers
    tenant_schema = request.headers.get("tenant_schema")

    # If not in headers, try to get from query params
    if not tenant_schema:
        tenant_schema = request.query_params.get("tenant_schema")

    # If not in headers, try to get from query params
    if not tenant_schema:
        tenant_schema = request.query_params.get("tenantSchema")

    # If still not found and body may be present
    if not tenant_schema and request.method in ("POST", "PUT", "PATCH"):
        body = await request.json()
        tenant_schema = body.get("tenant_schema")

    # If still not found and body may be present
    if not tenant_schema and request.method in ("POST", "PUT", "PATCH"):
        body = await request.json()
        tenant_schema = body.get("tenantSchema")

    if not tenant_schema:
        raise HTTPException(
            status_code=status.HTTP_400_BAD_REQUEST, detail="Tenant schema not provided"
        )

    return tenant_schema


async def get_db_connection_from_pool(
    request: Request, tenant_schema: str = Depends(get_tenant_schema)
) -> AsyncGenerator[oracledb.Connection]:
    pool = request.app.state.pool
    conn = pool.acquire()
    conn.current_schema = tenant_schema
    try:
        yield conn
    finally:
        conn.close()


@asynccontextmanager
async def lifespan(app: FastAPI):
    # Create the connection pool
    app.state.pool = create_db_connection_pool()

    print("Acquiring connection from pool...")
    with app.state.pool.acquire() as connection:
        print("Oracle DB version:" + connection.version)

    print("Test completed successfully (if no crash).")
    yield
    # Close the connection pool on shutdown
    app.state.pool.close()


app = FastAPI(
    title="test",
    description="test",
    version="1.0.0",
    lifespan=lifespan,
)


@app.get(
    "/database",
)
async def test_database_connection(
    tenant_schema: str | None,  # noqa: ARG001
    connection: Annotated[oracledb.Connection, Depends(get_db_connection_from_pool)],
):
    with connection.cursor() as cursor:
        cursor.execute("SELECT * FROM all_tables")
        result = cursor.fetchone()

        if result:
            print("works")


if __name__ == "__main__":
    import uvicorn

    uvicorn.run(app, host="127.0.0.1", port=8000, log_config=None, log_level=None)

This chunk of code can be used to reproduce issue when running in debugger.
Failure rate is about 50% of the time.
Image shown shows a successful run with debugger command shown.

[anthony-tuininga]

Are you able to use thin mode? Thick mode with version 23ai on macOS is not as stable as on other platforms. If you can avoid use of the Oracle Client libraries by using thin mode you may be able to workaround the issue.

[cjbj]

A "-2" update for the 23.3 Instant Client for macOS landed on https://www.oracle.com/database/technologies/instant-client/macos-arm64-downloads.html recently. It has a few bug fixes and might be worth trying.

[MCubek]

I'm afraid that for one of the databases mTLS is required so I can't avoid using the thick mode.
I will try with the new instant client and report back.

[cjbj]

@MCubek you might have meant NNE. But if you did mean mTLS you can use Thin mode it.

[MCubek]

Suppose so, it's an autonomous database which I have not provisioned myself.
Which ever way, unfortunately I can't connect to it using a thin client.

[MCubek]

I’ve been testing both my application and the provided example using the new Instant Client, and I haven’t encountered any crashes so far.

I’ll continue testing to monitor for any further issues, but based on the current stability, I believe this issue can be considered resolved and closed for now.

[anthony-tuininga]

That's great news! Will close for now.