HR Master Data System Migration

Project Context

• Alibaba Group's digital management domain included more than 500 systems that Ant Group needed to maintain independently.
• Our team was responsible for a data platform system with a complex call chain. Both the old and new system interfaces had to remain available.
• The central migration goals were interface stability, consistent data, and uninterrupted business workflows.

Core Approach

1. Migration strategy

   • Migrate the protocol, application code, and data interfaces in stages.
   • Forward requests from systems that had not migrated to the old interfaces so calls would continue to work.
   • Validate the rollout at two levels: correctness of application behavior and correctness of data.

2. System characteristics

   • The system handled far more reads than writes. Data synchronization was one-way across heterogeneous networks.
   • There were around 40 write interfaces with relatively low traffic, so the plan was to migrate them together.

3. Stability checks

   • Code behavior: Compare the migrated implementation with the old system.
   • Data: Check synchronized records against the old system's data.
   • Results: Compare read responses and database writes from write interfaces.
   • Business behavior: Check that the resulting workflows remain stable.
   • Recovery: Plan for read/write switching, transactional interfaces, exceptions, and rollback.

4. Observability

   • Across the system: Examine traffic by interface, calling system, and time period.
   • Through the stack: Trace calls end to end with logs from every layer.
   • Build these capabilities early enough to support the staged rollout and stability checks.

5. Performance and timeliness

   • Gateway communication between protocol and code layers, and between code and data layers, could add latency.
   • A cache could help highly time-sensitive workflows; workloads requiring stronger consistency could continue to use the old call path.

Questions and Risks

Q1. What could the migration plan overlook?

• Incomplete system knowledge: With roughly 1,000 interfaces, understanding every behavior is difficult.
• Missing shared patterns: Focusing only on individual interfaces makes it hard to form a coherent migration strategy.
• Response: Build observability for monitoring, logs, and traffic analysis; classify core and high-risk interfaces for closer attention; verify critical dependencies and middleware; and define a clear migration sequence.

Q2. How might mature AI or agent support change the work?

These are potential benefits, not measured outcomes from this migration:

• Automated comparison of interface results could account for differences in parameter order, list order, and time sorting. An agent could read annotations to help identify interface behavior and transactional reads or writes.
• The original estimate was that seven days of manual analysis might fall to hours with agent and LLM assistance, and that some early observability and switching-code work might fall by about 50%. These figures were planning estimates, not verified savings.
• An agent could analyze calling patterns, usage frequency, and business characteristics across upstream systems to reduce blind spots.

Q3. What problems might occur during migration?

1. Code: With around 1,000 interfaces and many upstream systems, time to understand every path is limited. Earlier optimizations or special cases could be missed and discovered during validation.
2. Team coordination: Multiple people may migrate or develop the same area at once. Branch work, pipeline alerts for overlapping changes, and frequent progress updates can reduce collisions.
3. Upstream workflows: Traffic replay may miss infrequent or long-cycle business scenarios. More discussion with upstream teams is needed.
4. Middleware differences: Small behavior differences between old and new middleware may only appear under unusual traffic. Compatibility needs to be checked with the middleware teams.
5. Consistency and freshness: Switching reads and writes while data is synchronized can introduce delay or temporary differences. Handle these according to each workflow's priority and requirements.