4.4 Distributed-System Replication Logs

1. What They Are

• Data changes between replicas are generally tracked through replication logs, with several formats.

2. Categories

2.1. Physical Logs

• Which page was modified, what was the original value, and what is the updated value.

2.2. Logical Logs

• Which record was modified. They are further divided into Statement and Row.

2.2.1. Statement

• The original statement requested by the client.
• Disadvantage: some statements, such as time functions, can have side effects.

2.2.2. Row

• For row insertion, the log records the new values of the relevant columns.
• For row deletion, the log indicates that this row was deleted.
• For row updates, the log records the new values of all columns.
• Disadvantage: takes up a lot of space. For example, deleting 10,000 rows requires recording that 10,000 rows were deleted.

3. Examples

• MySQL’s server-layer bin-log is a logical log and supports Statement and Row.
• MySQL’s storage-engine-layer redo-log is a physical log.
• Redis uses AOF, which is a logical log.
• Kafka’s log itself is the data.
• ZooKeeper uses?
• Elasticsearch does not use a log; instead, the primary shard sends synchronization requests in parallel to replica shards.