How rate limiting behaves
Public endpoints carry rate-limit checks backed by Redis: each request is evaluated against its endpoint's budget, and responses carry the remaining count so well-behaved clients can self-regulate. Client IP resolution normalizes forwarded addresses, so limits apply to the real caller rather than to a proxy's address.
Deny on failure
When Redis is unreachable, rate limiting denies rather than allows. This is deliberate: an unmeasurable public endpoint is more dangerous than an unavailable one. Clients that receive denial responses during an incident should back off and retry with jitter rather than escalate concurrency.
Channel-specific limits
Messaging channels carry their own caps on top of endpoint budgets — SMS dispatch is capped per minute with a bounded retry budget per message. These limits protect both the tenant's sender reputation and the carriers' thresholds. If reminders are throttled, the notification log shows the rate state per message rather than silent drops.
Respect the signal
The remaining-count header is a signal, not a challenge. Systems that treat rate-limit responses as invitations to retry harder degrade everyone's experience — including their own.