Jira AI Auto-Fix System

Design an internal tool that pulls new Jira bug tickets, where each bug corresponds to a test failure, and sends each ticket to an AI system that can generate a fix PR in roughly 30 minutes. The service buffers work with a message queue, tracks ticket state in a DB, and needs schema, sharding, and scaling discussion. Requirements Existing Jira continuously produces bug tickets; each ticket corresponds to a test failure. An AI service accepts a Jira ticket and generates a fix PR. AI generation is slow: roughly 30+ minutes per fix PR. Build a service that pulls Jira tickets, buffers them, sends work to the AI service, and tracks each ticket's current state. Store state in a DB; discuss schema design, sharding, and scaling. The workflow must tolerate retries and avoid creating duplicate PRs for the same ticket. Notes Treat the AI service as a slow asynchronous worker, not a synchronous request path. A queue between Jira ingestion and AI execution is the central buffer. Useful state machine: NEW -> QUEUED -> GENERATING -> PR_CREATED -> MERGED | FAILED | NEEDS_REVIEW. Persist state transitions with timestamps and external IDs for Jira issue and PR. Idempotency key should be the Jira ticket ID plus generation attempt. Before creating a PR, check whether a PR already exists for that ticket / attempt. Shard first by ticket or repository ownership if the system spans many repos; shard by Jira project / team if operational ownership matters more than uniform load. Scaling pressure comes from queue backlog and long AI runtime. Track queue age, generation duration, failure rate, duplicate suppression, and PR acceptance rate. Clarify whether humans review every generated PR before merge; the answer changes the downstream notification and approval workflow. Preparation Draw the pipeline: Jira poller/webhook -> dedupe/state DB -> message queue -> AI worker pool -> PR service -> status updater. Prepare the ticket-state schema and the idempotency story before discussing sharding. Walk through failure cases: Jira webhook replay, AI timeout, worker crash after PR creation, and duplicate ticket updates.