Skip to content

Latest commit

 

History

History
131 lines (101 loc) · 6.9 KB

README.md

File metadata and controls

131 lines (101 loc) · 6.9 KB
Raw

Socket options

SocketOptions controls various low-level parameters related to TCP connections between the driver and Cassandra.

You can provide these when initializing the cluster:

Cluster cluster = Cluster.builder()
        .addContactPoint("127.0.0.1")
        .withSocketOptions(
                new SocketOptions()
                        .setConnectTimeoutMillis(2000))
        .build();

If you don't, the driver uses a default configuration (see the javadocs of the setter methods of SocketOptions for the default values).

You can retrieve and change the options at runtime:

SocketOptions socketOptions = cluster.getConfiguration().getSocketOptions();
socketOptions.setConnectTimeoutMillis(3000);
  • changes to the read timeout will be taken into account for future request executions;
  • changes to any other option will be taken into account for future connections (connections that were already opened at the time of the change are unaffected, they keep the old values).

TCP options

setConnectTimeoutMillis defines how long the driver waits to establish a new connection to a Cassandra node before giving up.

Other options control the usual low-level TCP parameters (refer to their individual javadoc for details): setKeepAlive, setReceiveBufferSize, setReuseAddress, setSendBufferSize, setSoLinger, setTcpNoDelay. Most of these options are not set explicitly by the driver, so they get the default value from the underlying TCP transport. One exception is setTcpNoDelay, which is forced to true (meaning that Nagle's algorithm is disabled for driver connections).

Driver read timeout

setReadTimeoutMillis controls how long the driver waits for a response from a given Cassandra node before considering it unresponsive.

Cassandra normally provides a guaranteed response time for each type of query, as shown by these parameters in cassandra.yaml (here with their default values):

counter_write_request_timeout_in_ms: 5000
range_request_timeout_in_ms: 10000
read_request_timeout_in_ms: 5000
request_timeout_in_ms: 10000
truncate_request_timeout_in_ms: 60000
write_request_timeout_in_ms: 2000

The goal of setReadTimeoutMillis is to give up on a node if it took longer than these thresholds to reply, on the assumption that there's probably something wrong with it. Therefore it should be set higher than the server-side timeouts.

The default value is 12 seconds. For truncate queries (where the server timeout is 60 seconds) or aggregate queries (where read_request_timeout_in_ms applies per page of processed data, not to the whole query), you can increase the timeout on a specific statement:

session.execute(
        new SimpleStatement("TRUNCATE tmp").setReadTimeoutMillis(65000));

Do not set the read timeout too low, or the driver might give up on requests that had a chance of succeeding.

If the timeout is reached, the driver will receive an OperationTimedOutException, and invoke onRequestError on the retry policy to decide what to do (the default is to retry on the next node in the query plan).

Limiting overall query time

It should be clear by now that setReadTimeoutMillis is per node, not per query. If the driver retries on 4 different nodes, the overall execution time could theoretically be up to 4 times the read timeout. If you want a per query timeout, use the following pattern:

import com.google.common.base.Throwables;

ResultSet execute(Statement statement, long timeout, TimeUnit unit)
        throws InterruptedException, TimeoutException {
    ResultSetFuture future = session.executeAsync(statement);
    try {
        return future.get(timeout, unit);
    } catch (ExecutionException e) {
        throw Throwables.propagate(e.getCause());
    } catch (TimeoutException e) {
        future.cancel(true);
        throw e;
    }
}

A complementary approach is to enable speculative executions, to have the driver query multiple nodes in parallel. This way you won't have to wait for the full timeout if the first node is unresponsive.

Driver read timeout vs. server read timeout

Unfortunately, the term "read timeout" clashes with another concept that is not directly related: a Cassandra node may reply with a READ_TIMEOUT error when it didn't hear back from enough replicas during a read query.

To clarify:

  • driver read timeout: the driver did not receive any response from the current coordinator within SocketOptions.setReadTimeoutMillis. It invokes onRequestError on the retry policy with an OperationTimedOutException to decide what to do.
  • server read timeout: the driver did receive a response, but that response indicates that the coordinator timed out while waiting for other replicas. It invokes onReadTimeout on the retry policy to decide what to do.

We might rename SocketOptions.setReadTimeoutMillis in a future version to clear up any confusion.