[P1B][T09] 实现 Redis 统一租约与 Run/checkpoint 原子防旧 owner 提交 #7

Activity

1. Aias00

   added

   enhancementNew feature or requestNew feature or request

   on Oct 8, 2026
2. Aias00

   added the

   Task

   issue type on Oct 8, 2026
3. Aias00

   added a parent issue on Oct 8, 2026

   • [Roadmap] Redis-first 企业级 Multi-Agent 改造任务总览（无 MQ） argi#117
4. Aias00

   marked this as blocked on Oct 8, 2026

   • [P1B][T08] 实现 Redis 持久化运行记录与原子提交去重 #6
5. Aias00

   marked this as blocking 4 issues on Oct 8, 2026

   • [P1B][T11] 实现 Redis 无 MQ 轮询领取、过期恢复与唤醒索引修复 #8
   • [P1C][T16] 实现 Redis 模型日志与工具回执持久化适配 #11
   • [P2B][T24] 实现 Redis 审批 ticket、原子决议与跨实例恢复 #15
   • [P2B][T25] 提供 Redis 持久化调度与 occurrence 去重执行 #16
6. Aias00

   unmarked this as blocking on Oct 13, 2026

   • [P2B][T25] 提供 Redis 持久化调度与 occurrence 去重执行 #16
7. Aias00

   added

   priority/important-soon高优先级：近期应安排处理的重要问题高优先级：近期应安排处理的重要问题

   on Oct 13, 2026
8. yuluo-yx

   added a commit that references this issue on Oct 18, 2026

   Merge pull request #7 from spring-ai-alibaba/1025-yuluo/dashscope-model

   ...

   289d174
9. PoisonGodv commented on Oct 26, 2026

   

   PoisonGodv

   on Oct 26, 2026

   Contributor

   More actions

   Hi, I'd like to take T09, building on the Redis RunStore from T08 (#27).

   Rough plan:

   • Add a new opt-in adapter so the run lease (claim/renew/release) and checkpoint writes share one owner/fence. This replaces the separate run lock and checkpoint lock.
   • Use one Lua script per operation. Each script checks owner, fence, Redis TIME, and the run/checkpoint revision, then advances the run, checkpoint, journal and outbox together, or writes nothing.
   • Keep all keys of a session in one hash-tag slot and verify the slot at startup. A cross-slot or unsupported deployment fails fast (crossSlotConfigFailsFast). The existing RedisLeasedCheckpointSaver and its key format stay unchanged.
   • Tests against a real Redis via Testcontainers. Cover: two runIds in the same session, stale-owner heartbeat/release/run/checkpoint/journal/outbox writes being rejected, and an injected commit failure leaving no partial checkpoint or run progress with no ownerless fallback. The README will state that async HA failover is not consensus fencing and that RPO0 is not promised.

   Two questions before I start:

   1. refactor(persistence): target durable adapters at loop core #66 moved the Redis module onto argi-loop-core. Should T09 target the loop-core contracts, or is there a pending Core change for a unified lease/commit contract that I should wait for?
   2. Should the "independent JVM" takeover verification be part of this PR, or left to T13?

   Thanks!