Multi-Channel Seller Platform —
Zero-Downtime Cloud Migration

The situation

A US e-commerce operations SaaS platform — used by thousands of online merchants to automate their listing management, dynamic pricing, inventory sync, and order fulfillment across multiple sales channels — had outgrown its original on-premise infrastructure. The platform needed to migrate its core backend to AWS to handle growing seller volume, support new marketplace integrations, and sustain performance during peak commercial events.

The migration constraint was absolute. This was a revenue-critical system where any service interruption would cascade directly into seller losses: missed inventory windows, dropped order syncs, stale pricing during high-traffic events. Merchants ran their businesses on this platform 24 hours a day, seven days a week. A conventional big-bang migration cutover was not an option.

A second challenge ran in parallel. The platform needed to expand its marketplace integration layer — adding Walmart Marketplace alongside existing Amazon Seller Central, eBay, and Shopify connectors — while simultaneously rebuilding the underlying infrastructure beneath them. The new services had to be developed, tested against production traffic patterns, and deployed without touching any live seller data until full validation was complete.

What we built

• Full cloud migration of core platform services (on-premise → AWS ECS) using a blue-green deployment strategy — old and new environments ran in parallel throughout, with instant rollback capability at every traffic-shift stage
• New marketplace integration services covering the complete seller workflow: listing creation and sync, competitive repricing triggers, inventory reservation, order ingestion, fulfillment routing, and post-sale status propagation
• AI-powered dynamic repricing engine: a Python ML pipeline that adjusts listing prices in real time based on competitor pricing feeds, demand signals, margin floor rules, and marketplace-specific fee structures — replacing hours of manual repricing with automated sub-minute response cycles
• Idempotent data migration jobs with exactly-once guarantees across all seller records — open orders, active listings, and pending fulfillments migrated live with integrity verification at every step
• API integrations across Amazon Seller Central, eBay, Shopify, Walmart Marketplace, and downstream fulfillment providers — handling auth token rotation, rate-limit backoff, retry logic, and marketplace-specific event lifecycle edge cases
• Real-time SLA observability dashboards (CloudWatch) giving the operations team visibility into per-marketplace sync latency, error rates, and seller-facing uptime throughout the migration window and beyond

Why zero downtime required staged architecture

The primary migration risk wasn't the infrastructure itself — it was the live marketplace session state. Each platform integration maintained active OAuth tokens, webhook subscriptions, and in-flight API cursors. A naive cutover would have invalidated all of them simultaneously, dropping every active listing sync and order notification mid-stream. We rebuilt each integration layer in the new environment first, validated it against production-equivalent traffic in shadow mode, then shifted load incrementally — 5%, 20%, 50%, 100% — with automatic rollback triggers at each threshold.

The data migration presented a separate constraint: open orders and active listings couldn't be frozen during migration — they continued changing in the old system while we moved them. We built event-sourced migration jobs that replayed mutations from a WAL-tailed change stream, ensuring the new system was always within seconds of the old system's state before we switched record ownership. No seller saw a discrepancy.

The AI repricing engine was the new capability delivered on top of the migrated infrastructure. By centralizing pricing data across all marketplaces into a single pipeline, we could apply a unified ML pricing model that previously wasn't possible when each marketplace connector was siloed on separate legacy servers. The repricing engine now processes pricing signals across all connected channels within 90 seconds of a competitor price change.