B2B automotive supply chain engine with a low-memory catalogue client
Enterprise supply chain systems · GPU-accelerated rendering and distributed in-memory stock reservation
A wholesale ordering and stock reservation system for an automotive parts distributor, with a mobile catalogue for its field sales representatives and workshop buyers.
The challenge
Wholesale automotive parts distribution combines complex account-specific pricing tiers, OEM part cross-referencing, and real-time stock allocation across multiple regional warehouses. During peak ordering windows, hundreds of concurrent buyers contend for limited inventory where overselling a single unit disrupts downstream fulfilment.
At the same time, field sales representatives and workshop buyers browse tens of thousands of SKUs and dense mechanical schematics on mid-range mobile devices, requiring consistent high-frame-rate responsiveness within a strict memory ceiling.
- Memory ceiling
- 64 MB maximum client heap budget on constrained devices
- Frame pacing
- Locked 60/120 FPS (< 8.3 ms) across dense schematics
- Concurrency
- Peak wholesale ordering across distributed warehouses
- Inventory invariant
- Zero oversold units under simultaneous contention
Architecture
In-memory domain core
Tiered contract pricing, OEM equivalence graph traversal, and role-based access rules execute directly in application memory. Equivalency lookups traverse compact adjacency graphs and permission checks evaluate 64-bit bitwise masks, avoiding object-relational mapping overhead on high-frequency queries.
Bounded client runtime
The mobile client enforces a strict 64 MB memory budget through viewport virtualisation and pre-sized data buffers, sustaining smooth 60 or 120 FPS scrolling across 50,000-SKU catalogues without unexpected terminations.
Atomic multi-warehouse reservation
Persistent backend worker loops resolve stock reservations through single-roundtrip atomic state scripts inside the in-memory data store. Concurrent orders never lock database rows and cannot claim the same physical unit.
The catalogue client sends orders and lookups to the domain core. Long-running workers reserve stock through atomic scripts in the in-memory store, one round trip per reservation, across warehouses; the database takes orders without row-lock contention.
Hardware and protocols
Direct GPU canvas batching
Exploded mechanical drawings and high-density SKU grids bypass deep UI widget hierarchies, rendering directly on the GPU canvas via batched atlas draw calls and custom fragment shaders packed into contiguous float buffers.
Reference-counted texture pool
Decoded schematic textures are managed in a slab-allocated memory pool that automatically evicts least-recently-used assets when the operating system signals memory pressure, preventing garbage-collection pause spikes.
Zero-copy camera stream scanning
Industrial barcodes and OEM part numbers are decoded directly from raw camera sensor planes in under a single frame interval without allocating intermediate bitmap copies.
Verification
A self-contained, containerised suite runs concurrent virtual ordering workers against the same stock pools and checks that inventory never drifts below zero.
Headless rendering tests profile frame build times and check that heap allocation stays within fixed bounds on synthetic catalogues of 50,000 SKUs, each run in an isolated environment.