Zero-downtime AWS migration for a digital media platform

Client

A digital media and video production business running a content platform for a national audience. Named details are withheld at the client's request.

Business challenge

The platform ran as a single application on managed hosting that could no longer absorb traffic spikes around content releases. Deployments required a maintenance window, so releases were batched, infrequent and correspondingly risky. Growth plans depended on removing both constraints.

Existing technical constraints

• A monolithic application with no service boundaries and limited test coverage
• A single database instance handling both transactional and reporting queries
• Media assets served directly from the application host
• No staging environment that resembled production
• A small team with no capacity to run a long parallel migration

Proposed architecture

The migration used the strangler pattern so that value landed continuously rather than at a single cutover event.

1. Media delivery moved first to object storage behind a CDN, removing the largest source of load from the application
2. Read-heavy reporting queries moved to a replica, separating them from transactional traffic
3. The highest-traffic paths extracted into separate services running on containers
4. The remaining monolith containerised and moved as-is, rather than rewritten
5. Traffic shifted progressively at the edge, with the ability to route back at any point

Implementation

Each step was independently reversible. Database migration ran with replication in place and a verified rollback position before any traffic moved. Cutover was performed by shifting a weighted percentage of traffic and observing error rates and latency at each increment, rather than by switching DNS and hoping. Infrastructure was defined in code from the first environment, so staging genuinely matched production.

Security and governance

Identity was consolidated onto short-lived role-based credentials, removing long-lived keys from the deployment process. Network boundaries were introduced between the public edge, the application tier and the data tier. Logging was centralised so that a request could be traced end to end during cutover.

Outcome

• Zero downtime across the full migration, including database cutover
• Roughly 50% improvement in page load performance, driven mostly by the CDN and media changes
• Deployments moved from scheduled maintenance windows to on-demand
• Traffic spikes around content releases absorbed without manual scaling

Lessons and next phase

Moving media delivery first bought the headroom that made everything afterwards calmer — the highest-value step was also the lowest-risk one. The next phase covers autoscaling policy tuning and extracting the remaining scheduled jobs from the monolith.