Hi all — I've posted an RFC proposing intra-query fault tolerance for batch workloads in shared-data clusters: github.com/StarRocks/starrocks/pull/76056 TL;DR: opt-in staged execution mode where exchange output is materialized to shared storage, so a CN failure replays only the lost fragment instances instead of restarting the whole query. Targets long ETL jobs and spot/preemptible deployments; off by default, zero impact when disabled. Aligns with the Roadmap 2026 batch-processing track (#67632). Would appreciate PMC review and feedback on the overall direction, especially the staged-execution scheduling change and the open questions.