After that introduction to synchronous replication and dynamically adjustable durability, I want to focus on a feature called the replication slots.
What is the purpose of a replication slot? Let's consider the following example: There is a master and a slave. On the master, a large transaction is executed and the network connection is not fast enough to ship all of the data in time. At some point, the master removes its transaction log (checkpoint). If the slave is too far behind, a resync is needed. As we have already seen, the wal_keep_segments setting can be used to reduce the risk of failing replication. The question is this: what is the best value for the wal_keep_segments setting? Sure, more is better, but how much is best?
Replication slots will solve this problem for us: if we are using a replication slot, a master can only recycle the transaction log once it has been consumed by all replicas. The advantage here is that a slave can never fall behind so much that a resync is needed.
The trouble is, what if we shut down a replica without telling the master about it? The master would keep a transaction log forever and the disk on the primary server would eventually fill up, causing unnecessary downtime.
To reduce this risk for the master, replication slots should only be used in conjunction with proper monitoring and alerting. It is simply necessary to keep an eye on open replication slots that could potentially cause issues or might not be in use anymore.
In PostgreSQL, there are two types of replication slot:
- Physical replication slots
- Logical replication slots
Physical replication slots can be used for standard streaming replication. They will make sure that data is not recycled too early. Logical replication slots do the same thing. However, they are used for logical decoding. The idea behind logical decoding is to give users a chance to attach to the transaction log and decode it with a plugin. A logical transaction slot is, therefore, some sort of tail -f for database instances. It allows the user to extract changes made to the database—and therefore to the transaction log—in any format and for any purpose. In many cases, a logical replication slot is used for logical replication.