[rajtherengan]

Summary

With policy = redis and sync_rate >= period (e.g. sync_rate = 1 with a second limit), counters are not shared across workers and nodes. The first request in a window runs INCR against Redis and caches the result per worker; later requests are counted locally and pushed by a timer that fires sync_rate seconds later — after the window has already closed. Other workers and nodes never see those increments while they matter, so the effective limit becomes roughly limit × workers × nodes.

This PR implements option 4 from the issue, plus a warning:

• Any period whose length is <= sync_rate is treated as realtime for that period only: it is counted in Redis on every request, same as sync_rate = -1.
• Longer periods keep async batching. For example, second + minute with sync_rate = 1 → second is realtime, minute stays async.
• A warning is logged when the fallback applies, since Redis traffic goes up for those periods. It is logged once per worker, per plugin instance and period; making it once per node would need a shared dict — happy to add that if preferred.

I chose this over rejecting the configuration (option 1) so existing declarative configs keep loading on upgrade. Strict validation could still be added in a future major release.

The change is contained to the Redis policy in kong/plugins/rate-limiting/policies/init.lua: a per-period is_realtime() check used by usage() and increment(). increment() also skips scheduling a sync timer when every limited period is realtime.

Note: this makes the late-sync path in #14998 (EXPIREAT receiving milliseconds, so recreated keys never expire) much rarer, but does not fix it for async periods. The two changes are independent.

Checklist

• The Pull Request has tests
• A changelog file has been created under changelog/unreleased/kong
• There is a user-facing docs PR — docs(rate-limiting): document sync_rate realtime fallback for short periods developer.konghq.com#7400

Issue reference

Fixes #15018

[CLAassistant]

All committers have signed the CLA.