vllm-project / vllm-project/aibrix

Redis synchronized multi-instance prefix cache Question

Open
#2,401 3 comments 0 reactions 0 assignees View on GitHub
Dominant language
Go
Stars
5.1k
Forks
694
Avg merge
1d 19h
Merged PRs (30d)
98

Description

### 🚀 Feature Description and Motivation

@varungup90

Hello, I'm glad that you have implemented a Redis synchronized multi-instance prefix cache instance to solve the cache consistency problem across multiple instances. I have a few questions regarding the implementation details and hope to get your answers.

1. Why not directly put all the KV cache into Redis for direct queries instead of having each instance maintain a copy? Is it because frequent Redis queries result in high IO and performance degradation? However, this approach would require each replica to allocate a large amount of memory to store all the KV cache, which may lead to strong consistency issues for the data?

2. the current code supports a standalone Redis , why not support Redis cluster?

3. In the prefix-cache routing policy, the getRequestCounts data is currently not supported for multi-instance synchronization, which still causes data consistency issues when running multiple instances.

4. Redis key format: `aibrix:{namespace}:e:{entityId}` — `{namespace}` is the hash tag. Namespace must be non-empty.

if only can set one namespace, so that all the data falls into the same slot on Redis (cluster mode)? how to slove it

Thank you

### Use Case

gateway-plugin multi instance deploy

### Proposed Solution

_No response_

Contributor guide

Open the contributing guide

Research direction

Start by locating the Redis synchronized multi-instance prefix-cache implementation and the prefix-cache routing policy, then inspect how getRequestCounts and the key format aibrix:{namespace}:e:{entityId} are handled. The issue currently asks several design questions about Redis Cluster, namespace slotting, and consistency, so a concrete scope and acceptance criteria are needed before implementation can be considered done.

Written by the indexing model from the issue text.

Assessment

Tech stack
go, redis
Domain
distributed-systems
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.