This topic describes the tradeoffs involved with backup and recovery solutions.
With the exception of clustering, solutions can be implemented using features at both the storage and hypervisor layer. Choosing the right technology requires understanding both your requirements and what infrastructure is in use in your environment. The following sections outline some basic choices and the tradeoffs involved.
VMware vSphere's high availability provides the ability to have a VM configuration shared between multiple physical ESX servers. Once the storage has been configured on all physical servers, any server can run the Delphix Engine VM. This allows ESX clusters to survive physical server failure. In the event of failure, the VM is started on a different server and appears to clients as an unexpected reboot with non-zero but minimal downtime. Depending on the length of the outage, this may cause a short pause in I/O and database activity, but longer outages can trigger timeouts at the protocol and database layers that result in I/O and query errors. Such long outages are unlikely to occur in a properly configured environment.
Automatic detection of a failure in an HA environment does not work in all circumstances, and there are cases where the host, storage, or network can hang such that clients are deprived of access, but the systems continue to appear functional. In these cases, a manual failover of the systems may be required.
When configuring a cluster, it is important to provide standby infrastructure with equivalent resources and performance characteristics. Asymmetric performance capabilities can lead to poor performance in the event of a failover. In the worst case of an over-provisioned server, this can cause widespread workload failure and the inability to meet performance SLAs.
VMware provides storage-agnostic snapshots that are managed through the VMware Snapshot Manager. The use of VMware snapshots can, however, cause debilitating performance problems for write-heavy workloads due to the need to manage snapshot redo-log metadata. In order to provide an alternative snapshot implementation, while retaining the existing management infrastructure, VMware has created an API to allow storage vendors to supply their own snapshot implementation. This is only supported in ESX 5.1. Furthermore, the array must support the vStorage APIs. Consult the VMware documentation for supported storage solutions and the performance and management implications.
Storage-based snapshots, by virtue of being implemented natively in the storage array, typically do not suffer from such performance problems and are preferred over VMware snapshots when available. When managing storage-based snapshots, it is critical that all LUNs backing a single VM be part of the same consistency group. Consistency groups provide write order consistency across multiple LUNs and allow snapshots to be taken at the same point in time across the LUNs. This must include all VM configuration, system VMDKs, and VMDKs that hold the dSources and VDBs. Each storage vendor presents consistency groups in a different fashion; consult your storage vendor documentation for information on how to configure and manage snapshots across multiple LUNs.
In the event of a snapshot recovery becoming required, ensure that the Delphix Engine VM is powered off for the duration of the snapshot recovery. Failure to do so can lead to filesystem corruption as you're changing blocks underneath a running system.
Site Recovery Manager (SRM) is a VMware product that provides replication and failover of virtual machines within a vSphere environment. It is primarily an orchestration framework, with the actual data replication performed by a native VMware implementation, or by the storage array through a storage replication adapter (SRA). A list of supported SRAs can be found in the VMware documentation. There is some performance overhead in the native solution, but not of the same magnitude as the VMware snapshot impact. SRAs provide better performance but require that the same storage vendor be used as both source and target, and require resynchronization when migrating between storage vendors.
Storage-based replication can also be used in the absence of SRM, though this will require manual coordination when re-configuring and starting up VMs after failover. The VM configuration, as well as the storage configuration within ESX, will have to be recreated using the replicated storage.
The Delphix Engine also provides native replication within Delphix. This has the following benefits:
The target system is online and active
VDBs can be provisioned on the target from replicated objects
A subset of objects can be replicated
On failover, the objects are started in a disabled state. This allows configuration to be adjusted to reflect the target environment prior to triggering policy-driven actions.
Multiple sources can be replicated to a single target
Note that the Delphix Engine currently only replicates data objects (dSources and VDBs) and environments (source and target services). It does not replicate system configuration, such as users and policies. This provides more flexibility when mapping between disparate environments, but requires additional work when instantiating an identical copy of a system after failover.
There is a large ecosystem of storage and VM-based backup tools, each with its own particular advantages and limitations. VMware provides Data Protector, but there are size limitations (linked to a maximum of 2TB of deduped data) that make it impractical for most Delphix Engine deployments. Most third-party backup products, such as Symantec NetBackup, EMC Networker, and IBM Tivoli Storage Manager, have solutions designed specifically for the backup of virtual machines. Because the Delphix Engine is packaged as an appliance, it is not possible to install third-party backup agents. However, any existing solution that can back up virtual machines without the need for an agent on the system should be applicable to Delphix as well. Check with your preferred backup vendor to understand what capabilities exist.
Some storage vendors also provide a native backup of LUNs. Backup at the storage layer reduces overhead by avoiding data movement across the network but loses some flexibility by not operating within the VMware infrastructure. For example, recreating the VM storage configuration from restored LUNs is a manual process when using storage-based recovery.