Query timestamps

In Cassandra, each mutation has a microsecond-precision timestamp, which is used to order operations relative to each other.

There are various ways to assign it:

CQL USING TIMESTAMP

You can explicitly provide the timestamp in your CQL query:

session.execute("INSERT INTO my_table(c1, c2) values (1, 1) " +
    "USING TIMESTAMP 1432815430948040");

Client-side generation

This is enabled by default if you’re using the driver 3.0+ and a version of Cassandra that supports native protocol v3 or above.

A client timestamp will be sent by each query. It is generated by a TimestampGenerator. The default implementation is AtomicMonotonicTimestampGenerator.

You can specify another generator at initialization:

Cluster.builder().addContactPoint("127.0.0.1")
    .withTimestampGenerator(new MyCustomTimestampGenerator())
    .build();

In addition, you can also override the default timestamp on a per-statement basis:

Statement statement = new SimpleStatement(
    "UPDATE users SET email = 'x@y.com' where id = 1");
statement.setDefaultTimestamp(1234567890);
session.execute(statement);

Server-side generation

This is the “legacy” behavior if you’re connected to a Cassandra version that only supports protocol v2 or below. The server will assign a timestamp based on the time it receives the query.

This can be a problem when the order of the writes matter: with unlucky timing (different coordinators, network latency, etc.), two successive requests from the same client might be processed in a different order server-side, and end up with out-of-order timestamps. If protocol v3 is not an option, the only workaround is to add USING TIMESTAMP in your queries.

The driver ships with two implementations of TimestampGenerator:

  1. AtomicMonotonicTimestampGenerator, which guarantess monotonicity of timestamps for all threads;
  2. ThreadLocalMonotonicTimestampGenerator, which guarantees per-thread monotonicity of timestamps.

There is less contention using ThreadLocalMonotonicTimestampGenerator, but beware that there is a risk of timestamp collision with this generator when accessed by more than one thread; only use it when threads are not in direct competition for timestamp ties (i.e., they are executing independent statements).

Accuracy

Both implementations strive to achieve microsecond resolution on a best-effort basis. But in practice, the real accuracy of generated timestamps is largely dependent on the granularity of the underlying operating system’s clock.

For most systems, this minimum granularity is millisecond, and the sub-millisecond part of generated timestamps is simply a counter that gets incremented until the next clock tick, as provided by System.currentTimeMillis().

On some systems, however, it is possible to have a better granularity by using a JNR call to gettimeofday. This native call will be used when available, unless the system property com.datastax.driver.USE_NATIVE_CLOCK is explicitly set to false.

To check what’s available on your system:

  • make sure your Cluster uses a TimestampGenerator;
  • configure your logging framework to use level INFO for the category com.datastax.driver.core.ClockFactory;

  • look for one of the following messages at startup:

    INFO  com.datastax.driver.core.ClockFactory - Using java.lang.System clock to generate timestamps
INFO  com.datastax.driver.core.ClockFactory - Using native clock to generate timestamps

Monotonicity

The aforementioned implementations also guarantee that returned timestamps will always be monotonically increasing, even if multiple updates happen under the same millisecond.

Note that to guarantee such monotonicity, if more than one timestamp is generated within the same microsecond, or in the event of a system clock skew, both implementations might return timestamps that drift out in the future.

When this happens, the built-in generators log a periodic warning message in the category com.datastax.driver.core.TimestampGenerator. See their non-default constructors for ways to control the warning interval.

Summary

As shown in the previous sections, there are multiple ways to provide a timestamp, some of which overlap. The order of precedence is the following:

  1. if there is a USING TIMESTAMP clause in the CQL string, use that over anything else;
  2. otherwise, if a default timestamp was set on the statement and is different from Long.MIN_VALUE, use it;
  3. otherwise, if a generator is specified, invoke it and use its result if it is different from Long.MIN_VALUE;
  4. otherwise, let the server assign the timestamp.

Steps 2 and 3 only apply if native protocol v3 or above is in use.