A multi-CDN setup is only as smart as the thing choosing between networks. Two architectures dominate, and they fail in usefully different ways.
DNS-based steering
The resolver answers with the CDN of the moment, guided by health checks and performance feeds. It needs no client changes and covers every protocol, which is why it is the default. Its weaknesses are inherited from DNS itself: resolver caching delays failover by seconds to minutes, and poorly located resolvers can misplace users. The failure modes matter more than the feature lists, because routing layers earn their keep precisely on the days something is failing.
Client-side switching
Logic in the application, common in video players, measures delivery live and switches mid-session. Failover becomes near-instant and per-user rather than per-resolver. The costs: engineering investment, per-platform maintenance, and it only protects the traffic that runs the logic. Client-side logic also generates the best measurement data you will ever have, every switch decision is a real-user benchmark, which quietly improves the steering everywhere else.
Health checking deserves its own attention within DNS steering, because the steering is only as honest as the signals feeding it. Synthetic health probes from the steering provider’s vantage points can miss regional degradation your users feel, which is why RUM-fed steering, decisions informed by real user measurements, outperforms probe-fed steering in exactly the messy partial-failure scenarios multi-CDN exists for. When evaluating steering platforms, the question is not whether they health-check but what feeds the decision: rented probes see the internet from data centers, and your users do not live there.
Choosing, or combining
Websites and APIs usually get most of the value from well-tuned DNS steering. Streaming at serious scale increasingly runs both, DNS for coarse placement, the player for fine-grained survival. The combination is the current best practice for premium video. The combination pattern also degrades gracefully: if the player logic fails, DNS steering still holds the floor, which is the kind of redundancy a routing layer should model for everything else.
In practice
Map your traffic to the two mechanisms: web pages, APIs and downloads behind DNS steering with health checks and performance feeds; premium video through player-level switching if you ship one. Then rehearse a provider failure quarterly, with a stopwatch, because the honest metric of any steering design is minutes-to-clean-failover measured under calm conditions, not the architecture diagram’s implication of zero.
Steering design is half of any multi-CDN engagement we run. It is the half that decides whether the diagram works.