Hi team, We’re running StarRocks 3.2 with 3 FEs (1 leader + 2 followers) and a high-throughput analytical workload (~5000 QPS) alongside continuous ingestion (batch). We consistently see higher “Analyze” phase latency on the Leader FE compared to followers — typically 80–200 ms (leader) vs 10–40 ms (followers) at P95. This gap widens under heavy ingestion or DDL activity. Questions: 1. Is it expected that the Leader FE has higher analysis latency? 2. What are the main contributors? 3. Any config or tuning tips (catalog cache, analyze threads, plan cache) to reduce leader load? 4. Recommended way to route read-only queries to followers automatically? Goal: sustain <100 ms P99 latency at ≥5000 QPS with balanced FE load.

Leader FE has more workload compared to the follower, but the latency is not always higher. You can check the cpu and memory (use

top

) usage of the leader fe to see if it's overloaded. Since your QPS is pretty high you can reduce concurrency (

set pipeline_dop = 1;

) For the read-only queries, your application has to connect directly to the follower FE.

@Robert Raharjo "For the read-only queries, your application has to connect directly to the follower FE." Is this happens by default or require a configuration ?

When you start a session, if your app connect to

<leader_ip>:<leader_port>

then the queries will have to go to the leader, but if you use

<follower_ip>:<follower_port>

then you can skip the leader and directly query the follower.

⬆️ @Yakir Gibraltar @Rotem Adhoh Might be interesting for concurrency tests