connection: fix logging of non-IP sockets

avelanarius commented 10 months ago

Before this fix, the debug log would crash _connect_socket for UNIX domain sockets. getsockname() for UNIX domain sockets returns a single string instead of a tuple (as is the case for IPv4/IPv6). Therefore the code could crash as it tried to get the second element of a non-tuple (empty string):

Traceback (most recent call last):
  File "/home/margdoc/Workspace/scylla/maintenance_mode_testing.py", line 5, in <module>
    s = c.connect()
      ^^^^^^^^^^^
  File "cassandra/cluster.py", line 1750, in cassandra.cluster.Cluster.connect
  File "cassandra/cluster.py", line 1776, in cassandra.cluster.Cluster.connect
  File "cassandra/cluster.py", line 1763, in cassandra.cluster.Cluster.connect
  File "cassandra/cluster.py", line 3581, in cassandra.cluster.ControlConnection.connect
  File "cassandra/cluster.py", line 3642, in cassandra.cluster.ControlConnection._reconnect_internal
cassandra.cluster.NoHostAvailable: ('Unable to connect to any servers', {'test_socket': IndexError('string index out of range')})

Fix the issue by not unpacking those values and just printing them as-is, relying on %s formatter to print all elements of a tuple (host, port) for IP sockets and string for UNIX domain sockets. The printed log is not formatted as nice as before, however this is a DEBUG print so few users will ever see it. The new approach should work with any format of getsockname().

Fixes #278

mykaul commented 10 months ago

Do we have tests for Unix sockets?

avelanarius commented 10 months ago

There is this test: https://github.com/scylladb/python-driver/blob/64e7fad42ec88dfc72c7f12c389c3ef6c3f392fb/tests/integration/advanced/test_unixsocketendpoint.py

but it's not enabled. I'll check if this test works.

scylladb / python-driver

connection: fix logging of non-IP sockets #279