redis

Use a fixed namespace template like codeagent:{domain}:{id} with lowercase segments separated by colons. Centralize key construction in a constants class or key factory so business code never hardcodes key strings.

Add random jitter of 0-10% to the base TTL so mass-created keys do not expire simultaneously. For example, a 3600-second base becomes 3600-3960 seconds, spreading expirations and protecting the database from a sudden load spike.

Redis holds high-frequency, short-lifecycle, loss-tolerant runtime state like sessions and agent execution state. MySQL holds long-term facts requiring persistence, audit, or recovery such as conversations, executions, and documents.

Use SET NX with a TTL and a unique token per holder, then release via a Lua script that verifies the token before deleting. Never create locks without TTL or delete a lock unconditionally, since another thread may have acquired it after expiry.

No. Embedding dimensions must match the model that produced them, and index metadata should record the model name and dimension. Switching embedding models requires rebuilding the index and all stored vectors.

Redis is a cache and runtime state store, not a source of truth; data loss must never cause permanent business loss. It also has no cross-store transactions with MySQL, so consistency, retry, and compensation strategies must be designed explicitly.