# TiFlash Restart Query Impact Knowledge - Optimization ID: `tiflash-restart-query-impact-2026-08-20` - Owner: repository product knowledge selected by the semantic answer owner - Source reference: `APID-29273`, TiDB v8.5.4 release notes, and TiFlash issue `#10266` - Implementation status: merged and deployed to staging - Effective status: `optimized_verified` ## Bounded problem The customer already stated that one TiFlash host needed a restart, supplied v7.5.6, and asked whether running SQL would fail and whether that could be avoided. The original Auto Oncall answer reopened bug and metric discovery, gave only generic replica risk, and asked for information before answering the actual impact question. The missing decision is versioned product behavior and user-intent handling, not OP telemetry. The human answer established that v7.5.6 lacks TiFlash graceful shutdown, so in-flight SQL errors cannot be completely avoided; a low-traffic sequential restart only reduces impact. ## Candidate change - Recall one topic from planned TiFlash restart, exact v7.5.6, or graceful- shutdown language. - Answer impact and avoidance before discussing root cause or requesting more context. - Distinguish replica availability from in-flight MPP task continuity. - Keep production approval and execution human-owned. ## Verification status At `2026-08-20T22:19:47+08:00`, a tool-free offline simulation supplied only the cutoff-visible APID-29273 customer context and the bounded recalled topic; it did not supply the human answer or use OP, Jira, or DB data. The generated answer: - directly said that terminating TiFlash v7.5.6 can fail an in-flight MPP SQL; - said low-traffic, one-host-at-a-time restart reduces but cannot eliminate the error window; - included the v8.5.4 graceful-shutdown boundary without promising zero impact; - explained that replicas do not migrate an already-running MPP task; and - correctly treated `TIFLASH_REPLICA.AVAILABLE=1` as a Boolean availability indicator, not one available instance. An independent tool-free Judge compared this output with the historical human answer and returned `ai_better`, with no missing or incorrect item. PR [#1046](https://github.com/tidbcloud/aisupport/pull/1046) merged the bounded product-knowledge topic and recall contract. Staging image `455679d1` contains that commit and was Ready when the result was reconciled. Because Wetech OP is now intentionally disabled and cannot provide an applicable historical replay, the user accepted the cutoff-bounded, no-provider simulation as the equivalent case-specific evaluation for this semantic problem. The verified simulated customer/internal response was written in place to Jira internal comment `23350886`; readback confirmed `public=false` and the expected version-bounded TiFlash restart guidance. The canonical evaluation was then recorded as `source=analyze_write`, `verdict=ai_better` for the historical owner run and the APID-29273 Base row read back as acceptable. ## Unresolved gaps The unavailable historical OP environment still prevents a provider-backed replay. This verification covers the answer-intent and versioned product- knowledge criterion only; it does not verify historical cluster telemetry or restore Wetech OP evidence acquisition. ## Next validation action Monitor a later TiFlash restart-impact question to confirm the same topic is recalled without manual simulation. If a supported provider becomes available, retain the existing production-operation handoff and validate telemetry only as a separate problem.