Commit 30758a3

1.0.0b7

Nov 12, 2013

This release makes many stability improvements, especially around

prepared statements and node failure handling. In particular,

several cases where a request would never be completed (and as a

result, leave the application hanging) have been resolved.

Features

* Add `timeout` kwarg to ``ResponseFuture.result()``

* Create connection pools to all hosts in parallel when initializing

new Sesssions.

Bug Fixes

* Properly set exception on ResponseFuture when a query fails

against all hosts

* Improved cleanup and reconnection efforts when reconnection fails

on a node that has recently come up

* Use correct consistency level on when retrying failed operations

against a different host. (An invalid consistency level was being

used, causing the retry to fail.)

* Better error messages for failed prepare() opertaions

* Prepare new statements against all hosts in parallel (formerly

sequential)

* Fix failure to save the new current keyspace on connections. (This

could cause problems for prepared statements and lead to extra

operations to continuously re-set the keyspace)

* Avoid sharing LoadBalancingPolicies across Cluster instances. (When a second

Cluster was connected, it effectively mark nodes down for the first Cluster.)

* Better handling of failures during the re-preparation sequence for unrecognized

prepared statements

* Throttle trashing of underutilized connections to avoid trashing newly

created connections

* Fix race condition which could result in trashed connections being closed

before the last operations had completed

* Avoid preparing statements on the event loop thread (which could lead to

deadlock)

* Correctly mark up non-contact point nodes discovered by the control

connection. (This lead to prepared statements not being prepared

against those hosts, generating extra traffic later when the

statements were executed and unrecognized.)

* Correctly handle large messages through libev

* Add timeout to schema agreement check queries

* More complete (and less contended) locking around manipulation of the

pending message deque for libev connections

Other

* Prepare statements in batches of 10. (When many prepared statements

are in use, this allows the driver to start utilizing nodes that

were restarted more quickly.)

* Better debug logging around connection management

* Don't retain unreferenced prepared statements in the local cache.

(If many different prepared statements were created, this would

increase memory usage and greatly increase the amount of time

required to being utilizing a node that was added or marked

up.)