The key SLA parameters
A standard SLA for a personalization platform covers several measurable parameters:
| Parameter | Typical threshold | What it means |
|---|---|---|
| Uptime | 99.9-99.99% | The share of time the service is available |
| API p99 latency | 100-300 ms or better | 99% of requests are faster than this threshold |
| API p95 latency | 50-150 ms or better | 95% of requests are faster than this threshold |
| Support response time | 1-4 hours (P1) | Critical incidents |
| RPO / RTO | 1-4 hours | Data and service recovery after a failure |
For recommendation widgets embedded in the synchronous HTTP request of a page, a p99 latency of 100 ms or better is a hard requirement. A slow recommendations API slows the whole page down.
Graceful degradation when the SLA is breached
A correct personalization integration is built on the assumption that the SLA can be breached. The protective patterns:
Timeout plus fallback. The call to the personalization API is bounded at 200 to 300 ms. If no answer arrives, a fallback is shown — bestsellers, popular items in the category, cached recommendations.
Asynchronous loading. The recommendation widget loads after the page’s main content, lazily through JavaScript. An SLA breach then has no effect on how the key content loads.
A recommendations cache. Recommendations are cached at the CDN or application layer for 5 to 15 minutes. When the API is unavailable, the cache is served — slightly stale, but not empty.
Tip: when choosing a personalization platform, ask not only for the uptime figure but for the incident history of the last 12 months and the mean time to recovery. An SLA is a promise; MTTR is the reality.
The SLA and sale periods
Peak load during sales — Black Friday, Cyber Monday — is the real test of an SLA. That is also when a breach hurts most: traffic is at its highest and every conversion is worth more. Before a major sale it is worth doing three things in advance:
- confirm with the provider that they are ready for the load;
- configure an aggressive recommendations cache;
- verify the fallback works on a staging environment.