Ensure a monitor thread is restarted by Topology.open if it exited with
an exception, set Topology._opened = False in Topology.close, and extend
the timeout in test_client_disconnect.
Don't log out the global test client either; keep it logged in as root and
use it to create additional non-root users to test Database.add_user,
authenticate, logout, and so on.
Connect the MongoClient to the replica set if one is running,
otherwise connect it to the standalone. Only continue to use
MongoReplicaSetClient when testing the specific behavior of
that deprecated class.
Also suppress some warnings when running tests, and better
determine whether to test IPv6.
Replace MongoClient with an implementation that relies on Cluster and Server. The new MongoClient takes over MongoReplicaSetClient's responsibilities.
Authentication, secondary-pinning, and Mongos high-availability are broken and will be reimplemented in a future commit. RS tests are temporarily disabled.
This change does a few things:
1. In anticipation of the server defaulting to a "disabled" mode for
auth, only run authentication related tests if auth is enabled.
2. Attempt to clean up users in a finally block, even if tests failed.
3. Look for auth settings in "parsed" from getCmdLineOpts in case the
cluster was started with config files.
4. Remove useless remove_all_users calls at the beginning of tests - if
users were added in a previous test and we forgot to clean them up
these are going to fail no matter what when auth is enabled.
* RSState, Member, and MovingAverage are now immutable
* In refresh(), try up members before down ones
* A test_ha fixup (clear process-list after killing them in each test, so we don't re-kill previous tests' processes)
This change adds two new top level classes, MongoClient
and MongoReplicaSetClient. They work almost exactly the
same as Connection and ReplicaSetConnection with a few
differences:
1. These connection objects awknowledge write operations by default.
2. The network_timeout parameter is gone, use socketTimeoutMS instead.
3. auto_start_request defaults to False.
Default acknowledged write operation will be the biggest change for
most users. Using MongoClient and MongoReplicaSetClient you no longer
have to pass safe=True to a write method (save, update, insert, remove)
to request an acknowledgement from the server.
Connection and ReplicaSetConnection are now deprecated but retain their
previous behavior. This should not be a backward breaking change.
Replace the 'mongo' dict with a Member object everywhere in ReplicaSetConnection.
A handful of commands obey read preferences; most are always sent to primary.
Track a 5-sample moving average of each replica set member's ping time.
Connection detects whether it's connected to primary, secondary, or mongos.
Note: In CPython the monitor is no longer a daemon thread. This means that
care must be taken in the interactive shell and simple scripts to set
ReplicaSetConnection references to None so the monitor dies.
The monitor is still a daemon thread in Jython and PyPy.
Big change to PyMongo connection pooling:
* While we still allocate a socket per thread by default, this is now optional. It is also possible to share sockets among threads, safely, and thus using fewer total sockets, by creating a Connection or ReplicaSetConnection with auto_start_request=False.
* In the past, when a thread died without calling end_request() its socket was closed. We now reclaim such sockets for the pool, which should reduce connection churn.
* start_request() now returns a context manager so you can do "with connection.start_request():"
* ReplicaSetConnection now supports start_request, although its semantics aren't consistent for ReadPreferences other than PRIMARY
* Refactoring: Connection and ReplicaSetConnection had different pool implementations, now they share one.
