Questions to client side caching broadcast mode.

[proost]

Hi! It seems like rueidis more focuses on client side caching default mode.

Do you have plans to improve client side caching broadcast mode?

[rueian]

It does support broadcast mode: https://github.com/redis/rueidis?tab=readme-ov-file#broadcast-mode-client-side-caching

What would you like to be improved?

[proost]

In my case, some prefix usually has low and steady write requests but sometime has write requests spike. So At that time, it can hurt overall performance.

I wonder it is possible to divide connections. one is for just data, the other one is for invalidation.

Or if you think, it is already possible, could you tell me how to do it? i can't find solution.

[rueian]

Invalidations and responses to a same key should be on the same connection to avoid racing and avoid caching staled data. You can try increasing the PipelineMultiplex for more connections on different keys or probably we need to find why the current implementation doesn’t work well with broadcast spikes.

The on going work #712 could also be helpful.

[proost]

Invalidations and responses to a same key should be on the same connection to avoid racing and avoid caching staled data.

Yes, you are right.

can try increasing the PipelineMultiplex for more connections on different keys

It makes sense.

Hmm... in this case using two connection, one is used only with local cached keys and the other one is used only without local cached keys, looks like a possibility to mitigate above issue, but hacky. Or am i using client side caching wrong way?

[rueian]

Is it possible for you to create a short example that demonstrates the spike and the performance decline? I’m not sure if there’s any room for improvement.

[proost]

TL;DR) rueidis works right. my application logic is wrong.

I tried what i said, but still the issue happen. Root cause is after getting invalidate message and refreshing eagerly makes burst read requests to redis. So it makes worse performance sometimes. I changed refreshing asynchronous and distributed way all servers, and the issue is gone.

Sorry for confusing you and Thank you for answer!