A Coordinated Checkpointing Strategy for Stateful Stream Processing with Minimal Output Staleness in E-Commerce Marketplace Event Pipelines
- Authors
-
-
Keshav Kumar Bista
Department of Computer Science, Mid-Western School of Business Studies, 47 Birendranagar–Kakre Road, Surkhet 21700, NepalAuthor -
Dipesh Ram Kandel
Department of Computer Science, Himalayan Institute of Social Development,12 Gairidhara Ring Path, Kathmandu 44600, NepalAuthor
-
- Abstract
-
Stateful stream processing is a common substrate for e-commerce marketplaces where clicks, cart changes, inventory mutations, payments, and fulfillment events must be joined and aggregated under low-latency constraints. These pipelines frequently rely on periodic checkpointing to ensure exactly-once recovery, yet the coupling between checkpoint coordination and sink commits can make durable outputs lag behind the freshest processed events. The resulting output staleness is operationally visible as delayed inventory decrements, late fraud signals, and inconsistent availability surfaces, even when the engine continues to process events quickly in memory. This paper develops a coordinated checkpointing strategy that targets minimal output staleness while preserving deterministic recovery for stateful operators and transactional sinks. The strategy separates the act of establishing a global recovery point from the act of advancing a durable output frontier, and it coordinates both using watermark-aligned barriers and explicit sink-frontier acknowledgments. The design exploits incremental state snapshots, per-operator changelog compaction, and an output-commit protocol that can advance monotonically with bounded coordination overhead. A staleness-centric cost model is introduced to guide adaptive scheduling of checkpoint and output-frontier advancement under varying load, out-of-order event-time skew, and failure hazard rates. The analysis characterizes trade-offs between barrier alignment delay, snapshot I/O, sink transaction cadence, and expected rollback exposure, and it outlines implementation considerations for marketplace event schemas and multi-tenant resource isolation.
- Downloads
- Published
- 2025-06-04
- Section
- Articles