Partial Cassandra Writes

Overview

This document contains observation on Batch execution with different configuration for partial writes.

Tests and Observations

For testing different scenarios keyspace and table created

CREATE KEYSPACE test WITH replication = {'class':'SimpleStrategy', 'replication_factor' : 2};

CREATE TABLE user(name text, id text, PRIMARY KEY (name));

In java

PreparedStatement preparedStatement = session.prepare("insert into test.user (id, name) values (?, ?)");
PreparedStatement preparedStatement2 = session.prepare("update test.user set id='id_updated-1' where name=?");
PreparedStatement preparedStatement3 = session.prepare("update test.user set id='id_updated-2' where name=?");


int i = 1;
while(i <= 1000) {
	batchStatement.add(preparedStatement.bind("id_"+i, "user-" + i));
	++i;
}

batchStatement.add(preparedStatement2.bind("user-1"));
batchStatement.add(preparedStatement3.bind("user-2"));

In cassandra.yaml file change timeout to verify partial write

write_request_timeout_in_ms: 10

The above query will throw exception

Observation

The data got inserted into user table and got updated as above batch execution.
The WriteTimeoutException.getWriteType() == BATCH, ensures data is inserted into the table

Approach

To handle above WriteTimeoutException we can use BatchStatement.Type

Batch type of LOGGED, guarantees atomic insertion of data, and WriteTimeoutException.getWriteType() == BATCH, eventually written to the appropriate replicas and the developer doesn't have to do anything.

Thus, the above two checks handles partial write.