question

There are situations where updates to the database must be performed in sequence and cannot be interrupted until it has completed. This is usually not a problem for relational databases since records are locked by the database when a process is writing to it.

A typical example is when a user registers for a new account. Registering a new username is problematic when there are hundreds or thousands of sessions running concurrently.

Usernames are unique and can only be allocated once. To allocate a username, the following events must be performed in sequence (serial):

1. Check that the username has not already been registered (read).
2. If (1) is true, allocate the username (write).

For this sequence of events to succeed, it must be isolated from all other requests such that another user session cannot update the table between the check (step 1) and allocation (step 2) above. It needs to be linearizable (in programming terms) when multiple processes (user sessions) are performing the same events concurrently.

Cassandra achieves this isolation and linearizable (serial) consistency using the Paxos protocol. Paxos achieves consensus in distributed environments using a quorum-based algorithm.

The serial consistency levels are used during the Paxos phase of lightweight transactions, that is the read-before-write when it checks for the condition (for example, if a username does not exists). The consistency level for the reads in the Paxos phase are different from the normal consistency level because they need to prevent updates to the data to guarantee that the LWT condition is true for the lifetime of the operation. If the LWT condition is true then the write is performed using the normal consistency level specified.

See Jonathan Ellis' blog post on Lightweight transactions in Cassandra for a detailed description of the Paxos protocol algorithm used in Cassandra. Cheers!