AI Artifact Store
C++20 · Distributed Storage
AI Artifact Store
A C++20 distributed artifact store for AI/ML workloads, built around content addressing, resumable transfer, replication, and failure recovery.
Interactive Systems Demo
Node failure · Verified recovery · Replica repair
Story 01
Built for Failure
Replication and verified recovery when a storage node goes offline — repair is explicit, not autonomous.
Step 1
Healthy
Artifact stored with replicated chunks across storage nodes.
Step 2
Node Offline
A real StorageNode process can go offline while metadata may still list replicas.
Step 3
Pull Verified
Pull reads from another available replica and verifies SHA-256 on restore.
Step 4
Repair
Explicit repair restores the desired replication factor — not autonomous self-healing.
Story 02
Object ≠ ObjectLayout
The same raw content can keep the same Object identity while being represented by different chunk layouts such as FixedSize and FastCDC.
Raw Artifact
Object
SHA-256(entire raw object)
FixedSize Layout
physical chunking
Chunks
SHA-256(chunk bytes)
FastCDC Layout
physical chunking
Chunks
SHA-256(chunk bytes)
Chunk
SHA-256(chunk bytes)
Object
SHA-256(entire raw object)
ObjectLayout
the physical chunking representation
Story 03
Measured Trade-off
FastCDC trades chunking throughput for much stronger boundary stability and shifted-content reuse on this deterministic reference workload.
FixedSize
Chunk/hash throughput
2533.5 MiB/s
Shifted-content reuse
0%
FastCDC
Chunk/hash throughput
380.3 MiB/s
Shifted-content reuse
95.8%
Local Release reference workload — not a production SLA.
415
Correctness Tests
6
Process E2Es
3
Storage Nodes
in the interactive demo
95.8%
Shifted-content reuse
FastCDC reference fixture
Architecture
Control Plane · Data Plane
Client-orchestrated transfers with a PostgreSQL-backed metadata service and replicated content-addressed storage nodes.
C++ Client / CLI
Control Plane
Data Plane
Metadata Service
PostgreSQL
Node A
Node B
Node C