ADR-005: Spring Batch Chunk-Oriented Processing for EOD Settlement

1. Problem

At the end of the day, the system must reconcile and settle every transaction of the day. Loading all data into memory in a single Tasklet makes RAM usage proportional to the day's transaction count — unstable, with a risk of OutOfMemoryError as volume grows.

2. Choice

Use Spring Batch with a chunk-oriented step (Reader → Processor → Writer) instead of one Tasklet that processes everything at once.

3. Rationale

• Processing in small chunks keeps memory usage stable regardless of the total number of records.
• Spring Batch checkpoints after every chunk — if the job is interrupted (crash, lost DB connection), it can restart from where it stopped instead of from the beginning.
• faultTolerant() with skip/retry lets locally broken records (e.g. corrupted reconciliation data) be handled without failing the whole job.

4. Trade-offs

• More complex than a simple Tasklet: separate Reader/Processor/Writer and per-chunk transaction boundaries must be designed.
• The chunk size must be chosen carefully — too small increases transaction overhead, too large loses the memory benefit; the exact value will be measured and tuned with real data.