Stop sending $readPreference to standalone servers.
Stop sending $readPreference primary because it's the server default.
Remove outdated secondary_ok flag.
Tweak spec test because pymongo unpins cursors eagerly after errors.
Tweak spec test for PoolClearedEvent ordering when MongoDB handshake fails (see DRIVERS-1785).
Only skip killCursors for some error codes.
Rely on SDAM error handling to close the connection after a state change error.
Add service_id to various events.
Retain reference to pinned sockets to prevent premptive closure by CPython's cyclic GC.
Disable SRV Polling, SDAM compatibility check, logicalSessionTimeoutMinutes check.
server session pool pruning, server selection, and server monitoring.
A ServerType of LoadBalancer MUST be considered a data-bearing server.
"drivers MUST emit the following series of SDAM events" section.
Send loadBalanced:True with handshakes, validate serviceId.
Add topologyVersion fallback when serviceId is missing.
Don't mark load balancers unknown.
Mark server unknown and clear the pool when background connections fail.
Eagerly evict threads from the wait queue when pool is paused. Evicted
threads will raise the following error:
AutoReconnect('localhost:27017: connection pool paused')
Introduce PoolClearedEvent and ConnectionPoolListener.pool_ready.
CMAP spec test changes:
- CMAP unit tests should not use real monitors
- Assert that CMAP threads complete all scheduled operations
Add retryReads URI option that defaults to True.
Supported read operations will be retried once after transient
network, election, and shutdown errors on MongoDB 3.6+.
Supported operations are:
listCollections, listIndexes, and listDatabases
distinct
count, estimated_document_count, count_documents
aggregate (not including $out)
find (only for the initial find command, getMore commands are not
retried).
ChangeStreams: watch (initial aggregate command).
GridFS read APIs.
Test changes:
Add retryable reads spec test runner.
Disable retryable reads in network error tests.
Move exhaust getMore out of Cursor.
Move cursor command response decoding into Server so that all command
listener events can be published in the same location.
Use _OpReply class instead of passing bytes around.
Remove unnecessary operation argument to receive message.
Move _first_batch to message.py to avoid circular import.
This commit add the pymongo.collation.Collation class and associated enums. A
collation may be sent with individual write operations when connected to MongoDB
3.4+. Unlike read and write concern, a collation may not be attached to a
database or collection instance.
This commit implements the following simplifications and improvements:
- The Subscriber ABC has been renamed to CommandListener
- The subscribe function has been renamed to register
- The COMMAND constant has been removed
- The get_subscribers function has been removed
- Command listeners can now be registered per MongoClient instance
instead of just globally
This commit adds support for monitoring queries and getMore
operations, including getMore operations for command cursors
(e.g. aggregate cursor getMore). It also fixes a bug in duration
calculation.
SocketInfo and Pool are now responsible for catching all socket.errors and
gaierrors and translating them to ConnectionFailure. Server and MongoClient
need no longer worry about anything but ConnectionFailure. Functions in pool.py
and network.py still throw socket.errors into SocketInfo and Pool.
Call ismaster on each new connection and store the results on the SocketInfo
instance.
The upcoming Authentication Spec says: "If credentials exist, upon opening a
socket, drivers MUST send an isMaster command immediately. This allows a driver
to determine whether the server is an Arbiter. Calling ismaster additionally
allows the driver to know if the default authentication method for each socket
is MONGODB-CR or SCRAM-SHA-1, avoiding races when the driver repopulates the
pool after a disconnect."
In theory we could choose not to call ismaster if there are no credentials, but
it's simpler always to call ismaster, and paves the way for future breaking
changes to the wire protocol besides the current breaking change to
authentication.
Remove MongoClient's methods start_request(), in_request(), and end_request().
The purpose of requests was to provide read-your-writes consistency when using
w=0 write concern. Starting a request pins a socket to a thread so any
operations on that thread end up in the same queue on the server side.
Justification for removing: mongos 2.6+ doesn't support socket pinning by
default, and mongos 2.8+ doesn't support it at all (SERVER-12273), so whatever
weak consistency guarantees a request was supposed to provide are not provided
with sharding.
It's unnecessary with MongoDB 2.6+ since write commands always block and send a
response.
It's a confusing feature that should rarely be used, if ever, yet people
incorrectly use it all the time. It makes our connection pool unmaintainable to
any but the most expert developers.
The Server Discovery And Monitoring Spec has replaced the term "cluster"
with "topology", since "cluster" in MongoDB terminology should be reserved
for sharded clusters. "Topology" describes a single server, replica set, or a
set of mongoses.
The previous commit replaced "cluster" with "topology" in file names. Here
I replace the term in class and variable names.
