As my tenure on the Drupal Association was concluding, I was asked to give the Keynote for …
One of the most common concerns people have is how to ensure that their application is safe, secure and available in the event of an emergency. Often I have found that people are mistakenly believe that they are protected when in fact they often have ignored potential scenarios.
The principles explained apply equally well in RDBMSs, MongoDB and other databases.
Potential scenarios to protect against
Power circuit failure
Data center failure
To prevent drive failure use multiple drives in a single machine for high availability. RAID 10 provides the best performance with high availability. RAID 10 consists of a minimum of 4 disks which are split into mirrored pairs. The raid controller stripes across the pairs.
Replication is the act of mirroring the data from one server onto another. It will protect against any failure of one (or more) of machines. Commonly two or more machines are used. If properly configured it can also protect against most of the failures identified above. I will present a few different configurations and what they protect against.
In addition to standard Master Slave replication MongoDB provides an additional replication configuration called replica sets. Replica sets are similar to master slave, but they are set aware with automatic failover and recovery. Replica sets are a minimum of 3 nodes. 3 nodes can be 3 database nodes, or 2 database nodes + 1 arbiter. The arbiter doesn’t handle queries or store data, but is around to provide a third perspective to cast a vote when determining the status of the set. In master slave you can have the same setup but it requires you to participate as the arbiter.
Configuration 1 :
2 nodes + 1 arbiter (or 3 nodes) on same rack (same switch and power circuit).
- Protects against single drive failure (if not using RAID 0 or 10).
- Protects against single node failure.
Configuration 2 :
2 nodes + 1 arbiter on 3 different racks (different switch and power circuit).
- Protects against the above + failure of switch or power circuit.
Configuration 3 :
4 nodes + 1 arbiter on 3 different racks (different switch and power circuit).
- Protects against the above + two nodes failing or a node failure at the same time as a switch or power failure.
Configuration 4 :
4 nodes + 1 arbiter in 3 different data centers.
- Protects against all the above + data center failure.
Configuration 5 :
Any combination of the above but swap out the arbiter for a delayed secondary with priority 0 (hidden).
Some Protection against data loss in the event of Intrusion, Fat finger (sysadmin / DBA accidentally deleting / changing something), programmer error.
Only partial protection as any data written during the delay window will be lost. If not caught during the window (or windows) provides no protection.
A backup consists of a dump of the data ideally stored in a remote secure location. It’s important to ensure that the backup format used has an easy and consistent import mechanism. Any scripts you use to backup should have a counterpart restore written along with them. The worst possible time to discover issues with your backup is when you are trying to restore them. Make sure restoration is rock solid.
For permanent disaster recovery and for compliance with a variety of industry regulations it’s important to keep (offsite) backups. Depending on your data size, business and type of data you may want to take backups more or less regularly. A solid backup plan is the only way to ensure that the bulk of your data is never lost in the event of a disaster. It is often not enough to have replication be your sole form of backup as it doesn’t protect against intrusion, fat finger (sysadmin / DBA accidentally deleting / changing something) or programmer error outside of the delay window and not at all without a delayed replicant.
It’s important to understand the pitfalls and how to prevent them. Also recognize that there’s a balance that must be struck. Multiple data centers add additional complexity and bring a number of challenges along with their additional protection.
Lastly, Replication != backup. Please backup your data. If someone accidentally or intentionally maliciously performs a drop operation that same operation will be replicated across all of your machines. The only safety net is a good backup strategy.