Graceful Degradation

Partial service beats none

Graceful degradation means a system continues to provide its core function even when some components fail, rather than collapsing entirely. A shopping site whose recommendation engine is down should still let people browse and check out; losing a nice-to-have feature is a far better outcome than an error page for everyone.

The alternative — hard failure, where one broken dependency takes down the whole experience — magnifies small problems into large ones. Designing for degradation means deciding in advance what your system can live without, so a failure in a minor feature never blocks a major one.

Rank your features

You cannot protect everything equally, so decide what matters most. Sort features into tiers: the essential path that must keep working, and the enhancements that may be shed under stress. For an integration, submitting and tracking the core record is likely essential, while enrichment, suggestions, or analytics are candidates to drop first.

This ranking should be an explicit, agreed list, not an accident of which code happens to fail first. When you know the priority order ahead of time, shedding load during an incident becomes a deliberate choice rather than a scramble. Pair it with handling partial failures for the call-level mechanics.

Fallbacks and toggles

Several techniques deliver degradation. A fallback returns a sensible default when a dependency is unavailable — a slightly stale cached value, or an empty result the caller can handle. A feature toggle lets you switch off an expensive or fragile feature without a deploy. Caching lets you keep serving recent data when the source is briefly unreachable.

try:
  recs = recommend(user)
except DependencyError:
  recs = cached_or_empty(user)   # degrade, do not fail
render(page, recs)

The pattern is the same throughout: catch the failure, substitute something acceptable, and keep going.

Tell the user honestly

Degradation should be visible, not silent. If a feature is temporarily unavailable, say so plainly rather than showing stale data as if it were live or failing in a confusing way. A short, honest message — "recommendations are temporarily unavailable" — sets expectations and reduces support load.

Where the degradation stems from a platform dependency, point users to authoritative status so they can see it is being handled. Clear communication turns a degraded experience into a tolerable one; silence turns it into a flood of confused reports and lost trust.