[ratelimiting] bug when cassandra counter exceeds the limit

[smanolache]

Hello,

For some reason (non-atomicity of the read-modify-write counter operation in cassandra?) the request counter of the ratelimiting_metrics may exceed the configured limit. In this case

local remaining = conf.limit - current_usage

in kong/plugins/ratelimiting/access.lua becomes negative.

The condition

if remaining == 0 then

is then false and the request is passed through even if the limit is exceeded.

I propose to replace the condition with

if remaining <= 0 then

[thibaultcha]

Yes, the atomicity of the read-write operation is absolutely not guaranteed. I've been wanting to tweak the consistency but even for a single node Kong, single node Cassandra cluster, there is no atomicity, so there is an error rate especially for the second period.

I agree with the proposed change, if you want to submit a PR.

[smanolache]

Submitted the pull request.

[sonicaghi]

thank you!

[thibaultcha]

Yep closing this now