High availability is a target you agree, then a set of failure assumptions you design around. Two databases can both claim high availability and behave very differently when a disk fails, a network partition splits the cluster, or an operator runs the wrong migration.
This comparison is for the technical lead or operations manager choosing a design for a business-critical application. It covers the criteria that change the decision, six models and where each tends to fail.
Start with two targets and one failure list
Before comparing designs, write the recovery point objective (RPO), recovery time objective (RTO) and failures you actually expect: instance loss, storage failure, availability-zone loss, region loss, accidental deletion, corrupted deployment, and a malicious actor with database credentials.
A design that survives a disk failure may not survive a deleted table. A design that survives a region outage may still lose data if replication is asynchronous. Match the design to the failure list, not to a vendor's headline.
The comparison criteria
- Recovery point. How much committed data can be lost, and whether the answer changes under load.
- Recovery time. How long failover takes, including DNS, connection pools and retries.
- Failover trigger. Who or what decides that the primary is unhealthy, and how split brain is prevented.
- Application impact. Whether the application must handle retries, read-only mode or a changed connection endpoint.
- Operational burden. Patching, upgrades, monitoring, failover drills and on-call skill.
- Cost drivers. Duplicate capacity, cross-zone or cross-region transfer, licences, managed-service premiums and engineering time.
- Blast radius. Whether one mistake can take down both copies at once.
Six designs compared
1. Backup and restore
This is the baseline, not high availability. Backups protect against deletion and corruption, but recovery means rebuilding a server and replaying data. RPO depends on backup frequency; RTO depends on database size and restore testing. It fits internal tools and low-value workloads, and it should exist under every other design.
2. Warm standby
A second database is kept running and updated periodically, often with log shipping or asynchronous replication. Failover is usually manual. The design is simpler and cheaper than a clustered pair, with a larger RPO and a slower, more deliberate recovery. It suits businesses that can tolerate a short outage but cannot tolerate losing the dataset.
3. Active-passive with synchronous replication
One node accepts writes, with the standby acknowledging them before the primary commits. Failover can be automatic with a cluster manager or manual with a runbook. RPO is close to zero for the replicated scope, but write latency increases with distance, and the cluster introduces quorum and split-brain decisions that the team must understand. This is a strong fit for a single data centre or a metro-area pair.
4. Managed multi-AZ failover
The cloud provider runs the standby, manages replication and triggers failover, usually through a DNS endpoint. The team gets a smaller operational surface and a clear support boundary. The trade-off is less control over failover timing, engine-specific limits, and a service premium. It fits teams without a dedicated database administrator, provided they still test failover and understand what the provider does not cover.
5. Active-active multi-primary
More than one node accepts writes. This can improve read scale and local write latency, but it moves the hard problem into conflict resolution. Last-write-wins is rarely acceptable for inventory, payments or customer records. Active-active is appropriate only when the data model and business rules are designed for it, and when the team can operate conflict detection and repair.
6. Multi-region
A second region adds distance and a separate failure domain. With synchronous replication, latency may be unacceptable; with asynchronous replication, the business must accept a non-zero RPO. Multi-region is justified by a regulatory requirement, a large geographic user base or a recovery objective that a single region cannot meet. It is not a default upgrade.
How the models compare in practice
- Lowest complexity: backup and restore, then warm standby.
- Smallest RPO: synchronous active-passive or managed multi-AZ, within their supported distance.
- Fastest recovery: managed multi-AZ or a well-tested automatic failover cluster.
- Highest write scalability: active-active, at the cost of conflict handling.
- Largest blast radius: any design where both copies share credentials, a region, an account or a deployment pipeline.
- Most dependent on people: manual warm standby and multi-region runbooks.
Decision guidance
- Internal reporting or a small back office: backup and restore with tested restores, plus a documented rebuild time.
- Order management or a customer portal with a short outage tolerance: managed multi-AZ or active-passive synchronous replication.
- Payments or inventory where lost writes are unacceptable: synchronous replication with a proven failover process, and a separate mechanism for logical corruption.
- Global read-heavy product: read replicas first, then consider multi-region. Do not jump to multi-primary without a conflict model.
- Regulated data with a residency requirement: a second region in the required jurisdiction, with the RPO stated in the recovery plan.
Traps that erase the design's value
- Untested failover. A cluster that has never failed over is a hypothesis. Schedule drills and record the time.
- Cached connection strings. Applications and poolers may keep the old primary address long after DNS changes.
- Split brain. Two nodes accepting writes is worse than one node being down. Use quorum and fencing.
- Replication lag hidden by averages. Monitor the maximum lag and the age of the standby, not only the mean.
- Backups in the same account. A compromised credential or a billing failure can remove production and backups together.
- Retry storms. Applications that retry aggressively can finish off a struggling primary. Use backoff and idempotency.
Cost lines to model
Budget for duplicate compute and storage, cross-zone or cross-region data transfer, managed-service premiums, licences for commercial engines, monitoring, and the engineering time for drills and upgrades. Add the cost of the application changes needed for retries and failover, often the least-quoted line. A design that is cheap to run but requires a specialist on call every night is not cheap.
Questions for the vendor or platform team
- What exactly triggers failover, and who can trigger it manually?
- What is the measured RPO and RTO in the last drill, not the design target?
- What happens to in-flight transactions during failover?
- How is split brain prevented, and how is the old primary fenced?
- Which failures are outside the service's responsibility?
- What does the application need to change: retries, idempotency, read-only mode or connection handling?
Metrics
- Measured failover time against the agreed RTO.
- Maximum replication lag and the age of the standby.
- Percentage of restore and failover drills completed on schedule.
- Unplanned downtime minutes by cause.
- Data loss incidents and the number of unreconciled records after recovery.
If you need the recovery design, drills and monitoring built together, start with Managed Technology & Cybersecurity; the pricing page explains how assessment and ongoing operations are scoped. For the recovery testing side, read what backup and disaster recovery testing costs.
Sources
- PostgreSQL Documentation, High Availability, Load Balancing, and Replication, accessed 17 November 2026.
- Microsoft Learn, Always On availability groups, accessed 17 November 2026.
- AWS, Disaster Recovery Workloads on AWS, accessed 17 November 2026.
- Google, Site Reliability Engineering, accessed 17 November 2026.
- MySQL Reference Manual, Replication, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.