Upgrading¶
Upgrading to 3.0¶
Version 3.0 of the DataStax Python driver for Apache Cassandra adds support for Cassandra 3.0 while maintaining support for previously supported versions. In addition to substantial internal rework, there are several updates to the API that integrators will need to consider:
Default consistency is now LOCAL_ONE
¶
Previous value was ONE
. The new value is introduced to mesh with the default
DC-aware load balancing policy and to match other drivers.
Execution API Updates¶
Result return normalization¶
Previously results would be returned as a list
of rows for result rows
up to fetch_size
, and PagedResult
afterward. This could break
application code that assumed one type and got another.
Now, all results are returned as an iterable ResultSet
.
The preferred way to consume results of unknown size is to iterate through them, letting automatic paging occur as they are consumed.
results = session.execute("SELECT * FROM system.local")
for row in results:
process(row)
If the expected size of the results is known, it is still possible to materialize a list using the iterator:
results = session.execute("SELECT * FROM system.local")
row_list = list(results)
For backward compatability, ResultSet
supports indexing. If
the result is paged, all pages will be materialized. A warning will
be logged if a paged query is implicitly materialized.
Trace information is not attached to executed Statements¶
Previously trace data was attached to Statements if tracing was enabled. This could lead to confusion if the same statement was used for multiple executions.
Now, trace data is associated with the ResponseFuture
and ResultSet
returned for each query:
ResponseFuture.get_query_trace()
Binding named parameters now ignores extra names¶
Previously, BoundStatement.bind()
would raise if a mapping
was passed with extra names not found in the prepared statement.
Behavior in 3.0+ is to ignore extra names.
blist removed as soft dependency¶
Previously the driver had a soft dependency on blist sortedset
, using
that where available and using an internal fallback where possible.
Now, the driver never chooses the blist
variant, instead returning the
internal util.SortedSet
for all set
results. The class implements
all standard set operations, so no integration code should need to change unless
it explicitly checks for sortedset
type.
Metadata API Updates¶
PYTHON-276, PYTHON-408, PYTHON-400, PYTHON-422
Cassandra 3.0 brought a substantial overhaul to the internal schema metadata representation. This version of the driver supports that metadata in addition to the legacy version. Doing so also brought some changes to the metadata model.
The present API is documented: cassandra.metadata
. Changes highlighted below:
- All types are now exposed as CQL types instead of types derived from the internal server implementation
- Some metadata attributes have changed names to match current nomenclature (for example,
Index.kind
in place ofIndex.type
). - Some metadata attributes removed
TableMetadata.keyspace
reference replaced withTableMetadata.keyspace_name
ColumnMetadata.index
is removed table- and keyspace-level mappings are still maintained
Several deprecated features are removed¶
ResponseFuture.result
timeout parameter is removed, useSession.execute
timeout instead (031ebb0)Cluster.refresh_schema
removed, useCluster.refresh_*_metadata
instead (419fcdf)Cluster.submit_schema_refresh
removed (574266d)cqltypes
time/date functions removed, useutil
entry points instead (bb984ee)decoder
module removed (e16a073)TableMetadata.keyspace
attribute replaced withkeyspace_name
(cc94073)cqlengine.columns.TimeUUID.from_datetime
removed, useutil
variant instead (96489cc)cqlengine.columns.Float(double_precision)
parameter removed, usecolumns.Double
instead (a2d3a98)cqlengine
keyspace management functions are removed in favor of the strategy-specific entry points (4bd5909)cqlengine.Model.__polymorphic_*__
attributes removed, use__discriminator*
attributes instead (9d98c8e)cqlengine.statements
will no longer warn about list list prepend behavior (79efe97)
Upgrading to 2.1 from 2.0¶
Version 2.1 of the DataStax Python driver for Apache Cassandra adds support for Cassandra 2.1 and version 3 of the native protocol.
Cassandra 1.2, 2.0, and 2.1 are all supported. However, 1.2 only supports protocol version 1, and 2.0 only supports versions 1 and 2, so some features may not be available.
Using the v3 Native Protocol¶
By default, the driver will attempt to use version 2 of the
native protocol. To use version 3, you must explicitly
set the protocol_version
:
from cassandra.cluster import Cluster
cluster = Cluster(protocol_version=3)
Note that protocol version 3 is only supported by Cassandra 2.1+.
In future releases, the driver may default to using protocol version 3.
Working with User-Defined Types¶
Cassandra 2.1 introduced the ability to define new types:
USE KEYSPACE mykeyspace;
CREATE TYPE address (street text, city text, zip int);
The driver generally expects you to use instances of a specific
class to represent column values of this type. You can let the
driver know what class to use with Cluster.register_user_type()
:
cluster = Cluster()
class Address(object):
def __init__(self, street, city, zipcode):
self.street = street
self.city = text
self.zipcode = zipcode
cluster.register_user_type('mykeyspace', 'address', Address)
When inserting data for address
columns, you should pass in
instances of Address
. When querying data, address
column
values will be instances of Address
.
If no class is registered for a user-defined type, query results
will use a namedtuple
class and data may only be inserted
though prepared statements.
See User Defined Types for more details.
Customizing Encoders for Non-prepared Statements¶
Starting with version 2.1 of the driver, it is possible to customize
how Python types are converted to CQL literals when working with
non-prepared statements. This is done on a per-Session
basis through Session.encoder
:
cluster = Cluster()
session = cluster.connect()
session.encoder.mapping[tuple] = session.encoder.cql_encode_tuple
See Type Conversions for the table of default CQL literal conversions.
Using Client-Side Protocol-Level Timestamps¶
With version 3 of the native protocol, timestamps may be supplied by the client at the protocol level. (Normally, if they are not specified within the CQL query itself, a timestamp is generated server-side.)
When protocol_version
is set to 3 or higher, the driver
will automatically use client-side timestamps with microsecond precision
unless Session.use_client_timestamp
is changed to False
.
If a timestamp is specified within the CQL query, it will override the
timestamp generated by the driver.
Upgrading to 2.0 from 1.x¶
Version 2.0 of the DataStax Python driver for Apache Cassandra includes some notable improvements over version 1.x. This version of the driver supports Cassandra 1.2, 2.0, and 2.1. However, not all features may be used with Cassandra 1.2, and some new features in 2.1 are not yet supported.
Using the v2 Native Protocol¶
By default, the driver will attempt to use version 2 of Cassandra’s native protocol. You can explicitly set the protocol version to 2, though:
from cassandra.cluster import Cluster
cluster = Cluster(protocol_version=2)
When working with Cassandra 1.2, you will need to
explicitly set the protocol_version
to 1:
from cassandra.cluster import Cluster
cluster = Cluster(protocol_version=1)
Automatic Query Paging¶
Version 2 of the native protocol adds support for automatic query paging, which can make dealing with large result sets much simpler.
See Paging Large Queries for full details.
Protocol-Level Batch Statements¶
With version 1 of the native protocol, batching of statements required using a BATCH cql query. With version 2 of the native protocol, you can now batch statements at the protocol level. This allows you to use many different prepared statements within a single batch.
See BatchStatement
for details and usage examples.
SASL-based Authentication¶
Also new in version 2 of the native protocol is SASL-based authentication. See the section on Security for details and examples.
Lightweight Transactions¶
Lightweight transactions are another new feature. To use lightweight transactions, add IF
clauses
to your CQL queries and set the serial_consistency_level
on your statements.
Calling Cluster.shutdown()¶
In order to fix some issues around garbage collection and unclean interpreter
shutdowns, version 2.0 of the driver requires you to call Cluster.shutdown()
on your Cluster
objects when you are through with them.
This helps to guarantee a clean shutdown.
Deprecations¶
The following functions have moved from cassandra.decoder
to cassandra.query
.
The original functions have been left in place with a DeprecationWarning
for
now:
cassandra.decoder.tuple_factory
has moved tocassandra.query.tuple_factory
cassandra.decoder.named_tuple_factory
has moved tocassandra.query.named_tuple_factory
cassandra.decoder.dict_factory
has moved tocassandra.query.dict_factory
cassandra.decoder.ordered_dict_factory
has moved tocassandra.query.ordered_dict_factory
Dependency Changes¶
The following dependencies have officially been made optional:
scales
blist
And one new dependency has been added (to enable Python 3 support):
six