Hi, All! I recently began introducing StarRocks into our system. 🙂 and I have a question 👀 When using shared-data mode, even if the data cache is enabled, is it difficult to achieve sub-second query latency for ad-hoc queries because of the S3 I/O cost? If low latency is important, does that mean it’s better to use shared-nothing mode instead of shared-data? I prefer the shared-data mode because of the operational cost, so this is something I’m thinking about.

it depends on your data access pattern, if the data loaded into the system, has been expired in the cache, or due to scale-in/scale-out, causes cache missing, the query perf will be impacted a lot. One thing is to use

cache select

to proactively warmup the data needed again into cache. If the access pattern can't be predicated, and the query perf is always crucial, shared-nothing might be a better choice then.

I must say, smart folks know to go with shared data.😘

> it depends on your data access pattern, if the data loaded into the system, has been expired in the cache, or due to scale-in/scale-out, causes cache missing, the query perf will be impacted a lot. One thing is to use

cache select

to proactively warmup the data needed again into cache. If the access pattern can't be predicated, and the query perf is always crucial, shared-nothing might be a better choice then. Thanks for the explanation! It makes sense that we should choose the right approach depending on the data access pattern. 👍