Replace Oracle RAC with MariaDB Galera Cluster?
If you want to avoid downtimes in your business, High Availabilty (HA) is a strong requirement which, by definition, makes it possible to access your data all the time without losing (any) data. In this blog we compare two alternatives: Oracle RAC and MariaDB Galera Cluster.
There are several options to implement High Availability. Oracle RAC is a popular and proven HA solution. HA can also be enabled for your data and systems with loadbalancers that make it possible to always access your data. MariaDB Galera Cluster provides similar functionality using synchronous multi-master Galera replication. It is also easier to build and proves to be more cost-effective. Being OpenSource, you may have to pay for support, but not for running the system.
Next, the designs of Oracle RAC and MariaDB Galera Cluster are going to be compared, so you can make up your mind on your own.
With RAC, Oracle instances run on separate nodes, while the data is located on shared storage. All instances access the same files.
To prevent conflicts, the instances must agree on which instance is actually working on a block of data. If a node wants to change a row, it must get exclusive access to that block and store it in its cache. It therefore asks the other nodes whether they have the block. If no other node does, it gets the block from storage.
Even in case of read-access data all the nodes need to communicate this way to get the data as it is done for writing. When the block is modified, the requesting nodes get a consistent read version (which they are not allowed to modify) from the block. This adds latency due to internode communication – there will be read and write access every time a node does not have the block.
The need for communication between the nodes for every access on a table adds overhead. On the other hand, having all blocks advance local locks on a node, e.g. for SELECT FOR UPDATE, are cluster wide locks.
The advantage of RAC is that losing an Oracle node does not harm the service at all. The other nodes will keep providing data (HA of the access). Therefore, you can shut down a node to perform maintenance tasks such as upgrading hardware or software, while reducing unexpected downtime. However, the shared storage – responsible for the data – is a potential single point of failure.
On Oracle RAC distributing read or write access is not optimal because latency is added by additional internode round trips. The best results occur when the application only accesses a fixed part of the data per node, so that no blocks have to be moved around, but it makes the setup more complicated.
MariaDB Galera Cluster
In contrast to Oracle RAC, MariaDB Galera Cluster is a high availability setup with shared-nothing architecture. Instead of having one shared storage (SAN or NAS), every cluster member has its own copy of all the data, thus eliminating the single point of failure.
MariaDB Galera Cluster take care about syncing data even for new nodes. This makes managing the cluster easy, as adding an empty node into the cluster is sufficient. MariaDB Galera Cluster will provide all data for the new node.
Unlike Oracle RAC, accessing a node to read data does not result in internode communication. Instead, communication (and so latency) happens at the time transactions are committed. This is faster than the Oracle RAC approach of acquiring all blocks in advance, but this also means conflicts are found at the time a transaction is committed.
And conflict are found by the internode communication because of the commit. Thats why the same data should not be accessed (at least not at the same time) on different nodes, as this increases the chance of having conflicts. This will not happen when the data is accessed on different nodes one after another. In the case of Oracle RAC the blocks would have to be copied.
This means that a SELECT FOR UPDATE statement is able to fail on commit, as it locks the data locally but not cluster wide. So conflicts with transactions on other nodes can only be found at the time of the commit. That is why the same data should not be accessed at the same time on different nodes, as it increases the chance of having conflicts. This is slightly different to Oracle RAC where accessing data on another node any time later does move the blocks.
While Oracle RAC has a lot of latency moving data blocks into the cache of every node, MariaDB Galera Cluster has an increased likelihood of failing commits.
Like Oracle RAC, single nodes in a MariaDB Galera Cluster can be taken down for maintenance without stopping the cluster. When a node rejoins the cluster, it automatically gets missing transactions via Incremental State Transfer (IST), or it may sync all data using State Snapshot Transfer (SST). If the missing transactions are in a local (configurable) cache of a node, IST is used, if not SST is used.
One drawback of the current Galera version is that Data Definition Language (DDL) commands (CREATE, ALTER, DROP) are run synchronously on the cluster. Therefore the entire cluster stalls until a DDL command finishes. Thats why Magento installations running default configuration do not scale at all on MariaDB Galera Cluster. In general using tools like pt-online-schema-change bypass this limitation. Eliminating this limitation is on the development roadmap.
Oracle RAC and MariaDB Galera Cluster provide similar functionality using different designs. Each one is eliminating maintenance downtime for many tasks and thus gives you more freedom to run applications.
In general Oracle RAC has a lot more latency because of internode communication (including moving all requested data blocks) for read and write access. In MariaDB Galera Cluster the changed dataset is sent around by committing. So only changed datasets are sent.
Despite the obvious similarities, the two databases have quite different architectures. Oracle RAC uses shared storage, while MariaDB Galera Cluster uses a shared-nothing architecture, which is less expensive. Oracle RACs shared storage is quite expensive. The author has observed EMC or NetApp for that, as it is the single point of failure something reliable is needed.
Data on MariaDB Galera Cluster is replicated on all the nodes, which makes it easy to run the cluster spread over different regions. Consequently, your data will be safe even if your datacenter burns down. To have this level of redundancy with Oracle RAC you need a shared storage accordingly, i.e. a Netapp MetroCluster. Beside adding more costs, Netapp MetroCluster requires a network with a round trip latency of less than 10ms, while MariaDB Galera Cluster even runs in Cloud environments in different regions.
With Oracle RAC there are two inherent sources of latency: accessing the shared storage and internode communication for read and write access. While in MariaDB Galera Cluster there is latency for every COMMIT needed by the internode communication to check and send the data to be committed.
Of course MariaDB Galera Cluster is no one-to-one replacement for Oracle RAC. But if your application runs with either Oracle or MySQL/MariaDB, MariaDB Galera Cluster is more than an alternative.