Upgrade guide


Cross-datacenter failover

JAVA-2899 re-introduced the ability to perform cross-datacenter failover using the driver’s built-in load balancing policies. See Load balancing in the manual for details.

Cross-datacenter failover is disabled by default, therefore existing applications should not experience any disruption.

New RetryVerdict API

JAVA-2900 introduced RetryVerdict, a new interface that allows custom retry policies to customize the request before it is retried.

For this reason, the following methods in the RetryPolicy interface were added; they all return a RetryVerdict instance:

  1. onReadTimeoutVerdict
  2. onWriteTimeoutVerdict
  3. onUnavailableVerdict
  4. onRequestAbortedVerdict
  5. onErrorResponseVerdict

The following methods were deprecated and will be removed in the next major version:

  1. onReadTimeout
  2. onWriteTimeout
  3. onUnavailable
  4. onRequestAborted
  5. onErrorResponse

Driver 4.10.0 also re-introduced a retry policy whose behavior is equivalent to the DowngradingConsistencyRetryPolicy from driver 3.x. See this FAQ entry for more information.

Enhancements to the Uuids utility class

JAVA-2449 modified the implementation of Uuids.random(): this method does not delegate anymore to the JDK’s java.util.UUID.randomUUID() implementation, but instead re-implements random UUID generation using the non-cryptographic random number generator java.util.Random.

For most users, non-cryptographic strength is enough and this change should translate into better performance when generating UUIDs for database insertion. However, in the unlikely case where your application requires cryptographic strength for UUID generation, you should update your code to use java.util.UUID.randomUUID() instead of com.datastax.oss.driver.api.core.uuid.Uuids.random() from now on.

This release also introduces two new methods for random UUID generation:

  1. Uuids.random(Random): similar to Uuids.random() but allows to pass a custom instance of java.util.Random and/or re-use the same instance across calls.
  2. Uuids.random(SplittableRandom): similar to Uuids.random() but uses a java.util.SplittableRandom instead.

System and DSE keyspaces automatically excluded from metadata and token map computation

JAVA-2871 now allows for a more fine-grained control over which keyspaces should qualify for metadata and token map computation, including the ability to exclude keyspaces based on their names.

From now on, the following keyspaces are automatically excluded:

  1. The system keyspace;
  2. All keyspaces starting with system_;
  3. DSE-specific keyspaces:
    1. All keyspaces starting with dse_;
    2. The solr_admin keyspace;
    3. The OpsCenter keyspace.

This means that they won’t show up anymore in Metadata.getKeyspaces(), and TokenMap will return empty replicas and token ranges for them. If you need the driver to keep computing metadata and token map for these keyspaces, you now must modify the following configuration option: datastax-java-driver.advanced.metadata.schema.refreshed-keyspaces.

DSE Graph dependencies are now optional

Until driver 4.9.0, the driver declared a mandatory dependency to Apache TinkerPop, a library required only when connecting to DSE Graph. The vast majority of Apache Cassandra users did not need that library, but were paying the price of having that heavy-weight library in their application’s classpath.

Starting with driver 4.10.0, TinkerPop is now considered an optional dependency.

Regular users of Apache Cassandra that do not use DSE Graph will not notice any disruption.

DSE Graph users, however, will now have to explicitly declare a dependency to Apache TinkerPop. This can be achieved with Maven by adding the following dependencies to the <dependencies> section of your POM file:


See the integration section in the manual for more details as well as a driver vs. TinkerPop version compatibility matrix.

4.5.x - 4.6.0

These versions are subject to JAVA-2676, a bug that causes performance degradations in certain scenarios. We strongly recommend upgrading to at least 4.6.1.


Datastax Enterprise support is now available directly in the main driver. There is no longer a separate DSE driver.

For Apache Cassandra® users

The great news is that reactive execution is now available for everyone. See the CqlSession.executeReactive methods.

Apart from that, the only visible change is that DSE-specific features are now exposed in the API:

  • new execution methods: CqlSession.executeGraph, CqlSession.executeContinuously*. They all have default implementations so this doesn’t break binary compatibility. You can just ignore them.
  • new driver dependencies: TinkerPop, ESRI, Reactive Streams. If you want to keep your classpath lean, you can exclude some dependencies when you don’t use the corresponding DSE features; see the Integration>Driver dependencies section.

For Datastax Enterprise users

Adjust your Maven coordinates to use the unified artifact:

<!-- Replace: -->

<!-- By: -->

<!-- Do the same for the other modules: query builder, mapper... -->

The new driver is a drop-in replacement for the DSE driver. Note however that we’ve deprecated a few DSE-specific types in favor of their OSS equivalents. They still work, so you don’t need to make the changes right away; but you will get deprecation warnings:

  • DseSession: use CqlSession instead, it can now do everything that a DSE session does. This also applies to the builder:

    // Replace:
    DseSession session = DseSession.builder().build()  
    // By:
    CqlSession session = CqlSession.builder().build()
  • DseDriverConfigLoader: the driver no longer needs DSE-specific config loaders. All the factory methods in this class now redirect to DriverConfigLoader. On that note, dse-reference.conf does not exist anymore, all the driver defaults are now in reference.conf.

  • plain-text authentication: there is now a single implementation that works with both Cassandra and DSE. If you used DseProgrammaticPlainTextAuthProvider, replace it by PlainTextProgrammaticAuthProvider. Similarly, if you wrote a custom implementation by subclassing DsePlainTextAuthProviderBase, extend PlainTextAuthProviderBase instead.

  • DseLoadBalancingPolicy: DSE-specific features (the slow replica avoidance mechanism) have been merged into DefaultLoadBalancingPolicy. DseLoadBalancingPolicy still exists for backward compatibility, but it is now identical to the default policy.

Class Loader

The default class loader used by the driver when instantiating classes by reflection changed. Unless specified by the user, the driver will now use the same class loader that was used to load the driver classes themselves, in order to ensure that implemented interfaces and implementing classes are fully compatible.

This should ensure a more streamlined experience for OSGi users, who do not need anymore to define a specific class loader to use.

However if you are developing a web application and your setup corresponds to the following scenario, then you will now be required to explicitly define another class loader to use: if in your application the driver jar is loaded by the web server’s system class loader (for example, because the driver jar was placed in the “/lib” folder of the web server), then the default class loader will be the server’s system class loader. Then if the application tries to load, say, a custom load balancing policy declared in the web app’s “WEB-INF/lib” folder, then the default class loader will not be able to locate that class. Instead, you must use the web app’s class loader, that you can obtain in most web environments by calling Thread.getContextClassLoader():


See the javadocs of SessionBuilder.withClassLoader for more information.


Object mapper

4.1.0 marks the introduction of the new object mapper in the 4.x series.

Like driver 3, it relies on annotations to configure mapped entities and queries. However, there are a few notable differences:

  • it uses compile-time annotation processing instead of runtime reflection;
  • the “mapper” and “accessor” concepts have been unified into a single “DAO” component, that handles both pre-defined CRUD patterns, and user-provided queries.

Refer to the mapper manual for all the details.

Internal API

NettyOptions#afterBootstrapInitialized is now responsible for setting socket options on driver connections (see advanced.socket in the configuration). If you had written a custom NettyOptions for 4.0, you’ll have to copy over – and possibly adapt – the contents of DefaultNettyOptions#afterBootstrapInitialized (if you didn’t override NettyOptions, you don’t have to change anything).


Version 4 is major redesign of the internal architecture. As such, it is not binary compatible with previous versions. However, most of the concepts remain unchanged, and the new API will look very familiar to 2.x and 3.x users.

New Maven coordinates

The core driver is available from:


Runtime requirements

The driver now requires Java 8 or above. It does not depend on Guava anymore (we still use it internally but it’s shaded).

We have dropped support for legacy protocol versions v1 and v2. As a result, the driver is compatible with:

  • Apache Cassandra®: 2.1 and above;
  • Datastax Enterprise: 4.7 and above.


We’ve adopted new API conventions to better organize the driver code and make it more modular. As a result, package names have changed. However most public API types have the same names; you can use the auto-import or “find class” features of your IDE to discover the new locations.

Here’s a side-by-side comparison with the legacy driver for a basic example:

// Driver 3:
import com.datastax.driver.core.ResultSet;
import com.datastax.driver.core.Row;
import com.datastax.driver.core.SimpleStatement;

SimpleStatement statement =
  new SimpleStatement("SELECT release_version FROM system.local");
ResultSet resultSet = session.execute(statement);
Row row = resultSet.one();

// Driver 4:
import com.datastax.oss.driver.api.core.cql.ResultSet;
import com.datastax.oss.driver.api.core.cql.Row;
import com.datastax.oss.driver.api.core.cql.SimpleStatement;

SimpleStatement statement =
  SimpleStatement.newInstance("SELECT release_version FROM system.local");
ResultSet resultSet = session.execute(statement);
Row row = resultSet.one();

Notable changes:

  • the imports;
  • simple statement instances are now created with the newInstance static factory method. This is because SimpleStatement is now an interface (as most public API types).


The configuration has been completely revamped. Instead of ad-hoc configuration classes, the default mechanism is now file-based, using the Typesafe Config library. This is a better choice for most deployments, since it allows configuration changes without recompiling the client application (note that there are still programmatic setters for things that are likely to be injected dynamically, such as contact points).

The driver JAR contains a reference.conf file that defines the options with their defaults:

datastax-java-driver {
  basic.request {
    timeout = 2 seconds
    consistency = LOCAL_ONE
    page-size = 5000
  // ... and many more (~10 basic options, 70 advanced ones)

You can place an application.conf in your application’s classpath to override options selectively:

datastax-java-driver {
  basic.request.consistency = ONE

Options can also be overridden with system properties when launching your application:

java -Ddatastax-java-driver.basic.request.consistency=ONE MyApp

The configuration also supports execution profiles, that allow you to capture and reuse common sets of options:

// application.conf:
datastax-java-driver {
  profiles {
    profile1 { basic.request.consistency = QUORUM }
    profile2 { basic.request.consistency = ONE }

// Application code:
SimpleStatement statement1 =
SimpleStatement statement2 =

The configuration can be reloaded periodically at runtime:

datastax-java-driver {
  basic.config-reload-interval = 5 minutes

This is fully customizable: the configuration is exposed to the rest of the driver as an abstract DriverConfig interface; if the default implementation doesn’t work for you, you can write your own.

For more details, refer to the manual.


Cluster does not exist anymore; the session is now the main component, initialized in a single step:

CqlSession session = CqlSession.builder().build();

Protocol negotiation in mixed clusters has been improved: you no longer need to force the protocol version during a rolling upgrade. The driver will detect that there are older nodes, and downgrade to the best common denominator (see JAVA-1295).

Reconnection is now possible at startup: if no contact point is reachable, the driver will retry at periodic intervals (controlled by the reconnection policy) instead of throwing an error. To turn this on, set the following configuration option:

datastax-java-driver {
  advanced.reconnect-on-init = true

The session now has a built-in throttler to limit how many requests can execute concurrently. Here’s an example based on the number of requests (a rate-based variant is also available):

datastax-java-driver {
  advanced.throttler {
    class = ConcurrencyLimitingRequestThrottler
    max-concurrent-requests = 10000
    max-queue-size = 100000

Load balancing policy

Previous driver versions came with multiple load balancing policies that could be nested into each other. In our experience, this was one of the most complicated aspects of the configuration.

In driver 4, we are taking a more opinionated approach: we provide a single default policy, with what we consider as the best practices:

  • local only: we believe that failover should be handled at infrastructure level, not by application code.
  • token-aware.
  • optionally filtering nodes with a custom predicate.

You can still provide your own policy by implementing the LoadBalancingPolicy interface.


Simple, bound and batch statements are now exposed in the public API as interfaces. The internal implementations are immutable. This makes them automatically thread-safe: you don’t need to worry anymore about sharing them or reusing them between asynchronous executions.

Note that all mutating methods return a new instance, so make sure you don’t accidentally ignore their result:

BoundStatement boundSelect = preparedSelect.bind();

// This doesn't work: setInt doesn't modify boundSelect in place:
boundSelect.setInt("k", key);

// Instead, reassign the statement every time:
boundSelect = boundSelect.setInt("k", key);

These methods are annotated with @CheckReturnValue. Some code analysis tools – such as ErrorProne – can check correct usage at build time, and report mistakes as compiler errors.

Unlike 3.x, the request timeout now spans the entire request. In other words, it’s the maximum amount of time that session.execute will take, including any retry, speculative execution, etc. You can set it with Statement.setTimeout, or globally in the configuration with the basic.request.timeout option.

Prepared statements are now cached client-side: if you call session.prepare() twice with the same query string, it will no longer log a warning. The second call will return the same statement instance, without sending anything to the server:

PreparedStatement ps1 = session.prepare("SELECT * FROM product WHERE sku = ?");
PreparedStatement ps2 = session.prepare("SELECT * FROM product WHERE sku = ?");
assert ps1 == ps2;

This cache takes into account all execution parameters. For example, if you prepare the same query string with different consistency levels, you will get two distinct prepared statements, each propagating its own consistency level to its bound statements:

PreparedStatement ps1 =
      SimpleStatement.newInstance("SELECT * FROM product WHERE sku = ?")
PreparedStatement ps2 =
      SimpleStatement.newInstance("SELECT * FROM product WHERE sku = ?")

assert ps1 != ps2;

BoundStatement bs1 = ps1.bind();
assert bs1.getConsistencyLevel() == DefaultConsistencyLevel.ONE;

BoundStatement bs2 = ps2.bind();
assert bs2.getConsistencyLevel() == DefaultConsistencyLevel.TWO;

DDL statements are now debounced; see Why do DDL queries have a higher latency than driver 3? in the FAQ.

Dual result set APIs

In 3.x, both synchronous and asynchronous execution models shared a common result set implementation. This made asynchronous usage notably error-prone, because of the risk of accidentally triggering background synchronous fetches.

There are now two separate APIs: synchronous queries return a ResultSet; asynchronous queries return a future of AsyncResultSet.

ResultSet behaves much like its 3.x counterpart, except that background pre-fetching (fetchMoreResults) was deliberately removed, in order to keep this interface simple and intuitive. If you were using synchronous iterations with background pre-fetching, you should now switch to fully asynchronous iterations (see below).

AsyncResultSet is a simplified type that only contains the rows of the current page. When iterating asynchronously, you no longer need to stop the iteration manually: just consume all the rows in currentPage(), and then call fetchNextPage to retrieve the next page asynchronously. You will find more information about asynchronous iterations in the manual pages about asynchronous programming and paging.

CQL to Java type mappings

Since the driver now has access to Java 8 types, some of the CQL to Java type mappings have changed when it comes to temporal types such as date and timestamp:

The corresponding setter methods were also changed to expect these new types as inputs.


Metrics are now divided into two categories: session-wide and per-node. Each metric can be enabled or disabled individually in the configuration:

datastax-java-driver {
  advanced.metrics {
    // more are available, see reference.conf for the full list
    session.enabled = [ bytes-sent, bytes-received, cql-requests ]
    node.enabled = [ bytes-sent, bytes-received, pool.in-flight ]

Note that unlike 3.x, JMX is not supported out of the box. You’ll need to add the dependency explicitly:



Session.getMetadata() is now immutable and updated atomically. The node list, schema metadata and token map exposed by a given Metadata instance are guaranteed to be in sync. This is convenient for analytics clients that need a consistent view of the cluster at a given point in time; for example, a keyspace in metadata.getKeyspaces() will always have a corresponding entry in metadata.getTokenMap().

On the other hand, this means you have to call getMetadata() again each time you need a fresh copy; do not cache the result:

Metadata metadata = session.getMetadata();
Optional<KeyspaceMetadata> ks = metadata.getKeyspace("test");
assert !ks.isPresent();

      + "WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 1}");

// This is still the same metadata from before the CREATE
ks = metadata.getKeyspace("test");
assert !ks.isPresent();

// You need to fetch the whole metadata again
metadata = session.getMetadata();
ks = metadata.getKeyspace("test");
assert ks.isPresent();

Refreshing the metadata can be CPU-intensive, in particular the token map. To help alleviate that, it can now be filtered to a subset of keyspaces. This is useful if your application connects to a shared cluster, but does not use the whole schema:

datastax-java-driver {
  // defaults to empty (= all keyspaces)
  advanced.metadata.schema.refreshed-keyspaces = [ "users", "products" ]

See the manual for all the details.

Query builder

The query builder is now distributed as a separate artifact:


It is more cleanly separated from the core driver, and only focuses on query string generation. Built queries are no longer directly executable, you need to convert them into a string or a statement:

import static com.datastax.oss.driver.api.querybuilder.QueryBuilder.*;

BuildableQuery query =
        .value("id", bindMarker())
        .value("first_name", bindMarker())
        .value("last_name", bindMarker());

String cql = query.asCql();
// INSERT INTO user (id,first_name,last_name) VALUES (?,?,?)

SimpleStatement statement = query
    .addNamedValue("id", 0)
    .addNamedValue("first_name", "Jane")
    .addNamedValue("last_name", "Doe")

All query builder types are immutable, making them inherently thread-safe and share-safe.

The query builder has its own manual chapter, where the syntax is covered in detail.

Dedicated type for CQL identifiers

Instead of raw strings, the names of schema objects (keyspaces, tables, columns, etc.) are now wrapped in a dedicated CqlIdentifier type. This avoids ambiguities with regard to case sensitivity.

Pluggable request execution logic

Session is now a high-level abstraction capable of executing arbitrary requests. Out of the box, the driver exposes a more familiar subtype CqlSession, that provides familiar signatures for CQL queries (execute(Statement), prepare(String), etc).

However, the request execution logic is completely pluggable, and supports arbitrary request types (as long as you write the boilerplate to convert them to protocol messages).

We use that in our DSE driver to implement a reactive API and support for DSE graph. You can also take advantage of it to plug your own request types (if you’re interested, take a look at RequestProcessor in the internal API).