Load balancing

Load balancing controls how queries are distributed to nodes in a Cassandra cluster.

Without additional configuration the C/C++ driver defaults to using Datacenter-aware load balancing with token-aware routing. This means that driver will only send queries to nodes in the local datacenter (for local consistency levels) and it will use the primary key of queries to route them directly to the nodes where the corresponding data is located.

Round-robin Load Balancing

This load balancing policy equally distributes queries across cluster without consideration of datacenter locality. This should only be used with Cassandra clusters where all nodes are located in the same datacenter.

Datacenter-aware Load Balancing

This load balancing policy equally distributes queries to nodes in the local datacenter. Nodes in remote datacenters are only used when all local nodes are unavailable. Additionally, remote nodes are only considered when non-local consistency levels are used or if the driver is configured to use remote nodes with the allow_remote_dcs_for_local_cl setting.

CassCluster* cluster = cass_cluster_new();

const char* local_dc = "dc1"; /* Local datacenter name */

 * Use up to 2 remote datacenter nodes for remote consistency levels
 * or when `allow_remote_dcs_for_local_cl` is enabled.
unsigned used_hosts_per_remote_dc = 2;

/* Don't use remote datacenter nodes for local consistency levels */
cass_bool_t allow_remote_dcs_for_local_cl = cass_false;


/* ... */


Token-aware Routing

Token-aware routing uses the primary key of queries to route requests directly to the Cassandra nodes where the data is located. Using this policy avoids having to route requests through an extra coordinator node in the Cassandra cluster. This can improve query latency and reduce load on the Cassandra nodes. It can be used in conjunction with other load balancing and routing policies.

CassCluster* cluster = cass_cluster_new();

/* Enable token-aware routing (this is the default setting) */
cass_cluster_set_token_aware_routing(cluster, cass_true);

/* Disable token-aware routing */
cass_cluster_set_token_aware_routing(cluster, cass_false);

/* ... */


Latency-aware Routing

Latency-aware routing tracks the latency of queries to avoid sending new queries to poorly performing Cassandra nodes. It can be used in conjunction with other load balancing and routing policies.

CassCluster* cluster = cass_cluster_new();

/* Disable latency-aware routing (this is the default setting) */
cass_cluster_set_latency_aware_routing(cluster, cass_false);

/* Enable latency-aware routing */
cass_cluster_set_latency_aware_routing(cluster, cass_true);

 * Configure latency-aware routing settings

/* Up to 2 times the best performing latency is okay */
cass_double_t exclusion_threshold = 2.0;

 /* Use the default scale */
cass_uint64_t scale_ms = 100;

/* Retry a node after 10 seconds even if it was performing poorly before */
cass_uint64_t retry_period_ms = 10000;

/* Find the best performing latency every 100 milliseconds */
cass_uint64_t update_rate_ms = 100;

/* Only consider the average latency of a node after it's been queried 50 times */
cass_uint64_t min_measured = 50;


/* ... */


Filtering policies


This policy ensures that only hosts from the provided whitelist filter will ever be used. Any host that is not contained within the whitelist will be considered ignored and a connection will not be established. It can be used in conjunction with other load balancing and routing policies.

NOTE: Using this policy to limit the connections of the driver to a predefined set of hosts will defeat the auto-detection features of the driver. If the goal is to limit connections to hosts in a local datacenter use DC aware in conjunction with the round robin load balancing policy.

CassCluster* cluster = cass_cluster_new();

/* Set the list of predefined hosts the driver is allowed to connect to */

/* The whitelist can be cleared (and disabled) by using an empty string */
cass_cluster_set_whitelist_filtering(cluster, "");

/* ... */



This policy is the inverse of the whitelist policy where hosts provided in the blacklist filter will be ignored and a connection will not be established.

CassCluster* cluster = cass_cluster_new();

/* Set the list of predefined hosts the driver is NOT allowed to connect to */

/* The blacklist can be cleared (and disabled) by using an empty string */
cass_cluster_set_blacklist_filtering(cluster, "");

/* ... */



Filtering can also be performed on all hosts in a datacenter or multiple datacenters when using the whitelist/blacklist datacenter filtering polices.

CassCluster* cluster = cass_cluster_new();

/* Set the list of predefined datacenters the driver is allowed to connect to */
cass_cluster_set_whitelist_dc_filtering(cluster, "dc2, dc4");

/* The datacenter whitelist can be cleared/disabled by using an empty string */
cass_cluster_set_whitelist_dc_filtering(cluster, "");

/* ... */

CassCluster* cluster = cass_cluster_new();

/* Set the list of predefined datacenters the driver is NOT allowed to connect to */
cass_cluster_set_blacklist_dc_filtering(cluster, "dc2, dc4");

/* The datacenter blacklist can be cleared/disabled by using an empty string */
cass_cluster_set_blacklist_dc_filtering(cluster, "");

/* ... */


Speculative Execution

For certain applications it is of the utmost importance to minimize latency. Speculative execution is a way to minimize latency by preemptively executing several instances of the same query against different nodes. The fastest response is then returned to the client application and the other requests are cancelled. Speculative execution is disabled by default.

Query Idempotence

Speculative execution will result in executing the same query several times. Therefore, it is important that queries are idempotent i.e. a query can be applied multiple times without changing the result beyond the initial application. Queries that are not explicitly marked as idempotent will not be scheduled for speculative executions.

The following types of queries are not idempotent:

  • Mutation of counter columns
  • Prepending or appending to a list column
  • Use of non-idempotent CQL function e.g. now() or uuid()

The driver is unable to determine if a query is idempotent therefore it is up to an application to explicitly mark a statement as being idempotent.

CassStatement* statement = cass_statement_new( "SELECT * FROM table1", 0);

/* Make the statement idempotent */
cass_statement_set_is_idempotent(statement, cass_true);


Enabling speculative execution

Speculative execution is enabled by connecting a CassSession with a CassCluster that has a speculative execution policy enabled. The driver currently only supports a constant policy, but may support more in the future.

Constant speculative execution policy

The following will start up to 2 more executions after the initial execution with the subsequent executions being created 500 milliseconds apart.

CassCluster* cluster = cass_cluster_new();

cass_int64_t constant_delay_ms = 500; /* Delay before a new execution is created */
int max_speculative_executions = 2;   /* Number of executions */


/* ... */


Connection Heartbeats

To prevent intermediate network devices (routers, switches, etc.) from disconnecting pooled connections the driver sends a lightweight heartbeat request (using an OPTIONS protocol request) periodically. By default the driver sends a heartbeat every 30 seconds. This can be changed or disabled (0 second interval) using the following:

CassCluster* cluster = cass_cluster_new();

/* Change the heartbeat interval to 1 minute */
cass_cluster_set_connection_heartbeat_interval(cluster, 60);

/* Disable heartbeat requests */
cass_cluster_set_connection_heartbeat_interval(cluster, 0);

/* ... */


Heartbeats are also used to detect unresponsive connections. An idle timeout setting controls the amount of time a connection is allowed to be without a successful heartbeat before being terminated and scheduled for reconnection. This interval can be changed from the default of 60 seconds:

CassCluster* cluster = cass_cluster_new();

/* Change the idle timeout to 2 minute */
cass_cluster_set_connection_idle_timeout(cluster, 120);

/* ... */


It can be disabled by setting the value to a very long timeout or by disabling heartbeats.

Host State Changes

The status and membership of a node can change within the life-cycle of the cluster. A host listener callback can be used to detect these changes.

Important: The driver runs the host listener callback on a thread that is different from the application. Any data accessed in the callback must be immutable or synchronized with a mutex, semaphore, etc.

void on_host_listener(CassHostListenerEvent event, CassInet inet, void* data) {
  /* Get the string representation of the inet address */
  char address[CASS_INET_STRING_LENGTH];
  cass_inet_string(inet, address);

  /* Perform application logic for host listener event */
    printf("Host %s has been ADDED\n", address);
   } else if (event == CASS_HOST_LISTENER_EVENT_REMOVE) {
    printf("Host %s has been REMOVED\n", address);
   } else if (event == CASS_HOST_LISTENER_EVENT_UP) {
    printf("Host %s is UP\n", address);
   } else if (event == CASS_HOST_LISTENER_EVENT_DOWN) {
    printf("Host %s is DOWN\n", address);

int main() {
  CassCluster* cluster = cass_cluster_new();

  /* Register the host listener callback */
  cass_cluster_set_host_listener_callback(cluster, on_host_listener, NULL);

  /* ... */


Note: Expensive (e.g. slow) operations should not be performed in host listener callbacks. Performing expensive operations in a callback will block or slow the driver’s normal operation.

Reconnection Policy

The reconnection policy controls the interval between each attempt for a given connection.

Exponential Reconnection Policy

The exponential reconnection policy is the default reconnection policy. It starts by using a base delay in milliseconds which is then exponentially increased (doubled) during each reconnection attempt; up to the defined maximum delay.

Note: Once the connection is re-established, this policy will restart using base delay if a reconnection occurs.

Constant Reconnection Policy

The constant reconnection policy is a fixed delay for each reconnection attempt.

Performance Tips

Use a single persistent session

Sessions are expensive objects to create in both time and resources because they maintain a pool of connections to your Cassandra cluster. An application should create a minimal number of sessions and maintain them for the lifetime of an application.

Use token-aware and latency-aware policies

The token-aware load balancing can reduce the latency of requests by avoiding an extra network hop through a coordinator node. When using the token-aware policy requests are sent to one of the nodes which will retrieved or stored instead of routing the request through a proxy node (coordinator node).

The latency-aware load balancing policy can also reduce the latency of requests by routing requests to nodes that historical performing with the lowest latency. This can prevent requests from being sent to nodes that are underperforming.

Both latency-aware and token-aware can be use together to obtain the benefits of both.

Use paging when retrieving large result sets

Using a large page size or a very high LIMIT clause can cause your application to delay for each individual request. The driver’s paging mechanism can be used to decrease the latency of individual requests.

Choose a lower consistency level

Ultimately, choosing a consistency level is a trade-off between consistency and availability. Performance should not be a large deciding factor when choosing a consistency level. However, it can affect high-percentile latency numbers because requests with consistency levels greater than ONE can cause requests to wait for one or more nodes to respond back to the coordinator node before a request can complete. In multi-datacenter configurations, consistency levels such as EACH_QUORUM can cause a request to wait for replication across a slower cross datacenter network link. More information about setting the consistency level can be found here.

Driver Tuning

Beyond the performance tips and best practices considered in the previous section your application might consider tuning the more fine-grain driver settings in this section to achieve optimal performance for your application’s specific workload.

Increasing core connections

In some workloads, throughput can be increased by increasing the number of core connections. By default, the driver uses a single core connection per host. It’s recommended that you try increasing the core connections to two and slowly increase this number while doing performance testing. Two core connections is often a good setting and increasing the core connections too high will decrease performance because having multiple connections to a single host inhibits the driver’s ability to coalesce multiple requests into a fewer number of system calls.

Coalesce delay

The coalesce delay is an optimization to reduce the number of system calls required to process requests. This setting controls how long the driver’s I/O threads wait for requests to accumulate before flushing them on to the wire. Larger values for coalesce delay are preferred for throughput-based workloads as it can significantly reduce the number of system calls required to process requests.

In general, the coalesce delay should be increased for throughput-based workloads and can be decreased for latency-based workloads. Most importantly, the delay should consider the responsiveness guarantees of your application.

Note: Single, sporadic requests are not generally affected by this delay and are processed immediately.

New request ratio

The new request ratio controls how much time an I/O thread spends processing new requests versus handling outstanding requests. This value is a percentage (with a value from 1 to 100), where larger values will dedicate more time to processing new requests and less time on outstanding requests. The goal of this setting is to balance the time spent processing new/outstanding requests and prevent either from fully monopolizing the I/O thread’s processing time. It’s recommended that your application decrease this value if computationally expensive or long-running future callbacks are used (via cass_future_set_callback()), otherwise this can be left unchanged.