Distributed Ecosystem & Six-Engine Architecture
Comprehensive architectural blueprint of the six decoupled engines powering Aarkam Sovereign Data Fabric: Aarkam.IO, Rokka, Kdouja, Abodi, Ahmodi, and AM CLI.
Aarkam System V2 is engineered from first principles as a six-engine decoupled sovereign storage fabric. Unlike monolithic legacy storage appliances or public cloud object stores that bind storage, consensus, encryption, and diagnostics into brittle runtimes, Aarkam decomposes the storage stack into six independently scalable, deployable, and testable subsystems.
The Six Decoupled System Modules
Multi-Tenant UI Portal • Sovereign Identity & IAM • S3 REST SigV4 Gateway Backbone • WORM Object Lock
Stateless Raft MON Consensus Quorum • Consistent Hash Ring • StorageNode Streaming Daemons
Segment-Consolidated PageStore (256MB Segments) • Zero Inode Exhaustion • Direct NVMe Bypass • Arena Skip-List
SIMD AVX2 / AVX-512 Galois Field GF(2⁸) Kernel • 58.3% TCO Hardware Savings • Speculative Hedged Reads
Kernel S.M.A.R.T. Telemetry • NVMe Wear Tracking • ML Failure Runway Forecasting • Zero-Downtime Eviction
Single-Binary Orchestration • 27 Management Domains • POSIX Exit Codes • Structured JSON/YAML & Live Watch
System Modules Cross-Reference Specification
| Module | Primary Responsibility | Deployment Target | Network / Protocol | Default Ports |
|---|---|---|---|---|
| Aarkam.IO | Multi-tenant Web Portal, S3 REST Gateway, IAM, Sovereign Policy | Management Nodes / DMZ Edge | HTTPS, gRPC mTLS | Web UI: 7771 / 57771S3 Gateway: 7772 / 57772 |
| Rokka | Distributed Fabric, Raft MON Quorum, Consistent Hash Ring, Data Streaming | 3+ Coordinator Nodes & all Storage Hosts | gRPC over HTTP/2 with mTLS | Coordinator: 7774 / 57774StorageNode: 7775 / 57775 |
| Kdouja | Direct I/O Chunk Persistence, Segment PageStore, In-memory Skip-List | In-process inside Rokka.StorageNode |
In-process memory spans, Direct NVMe I/O | Embedded Native Core |
| Ahmodi | Cauchy Reed-Solomon Erasure Coding, AVX-512 SIMD Galois Field Kernel | Shared Library in Gateway & Recovery Nodes | SIMD Vector Registers / CPU Bus | In-Process Vector Kernel |
| Abodi | Autonomous S.M.A.R.T. Telemetry, Machine Learning Wear Prediction, Heartbeat Flow | Sidecar on every physical storage node | Local IPC (WMI/sysfs), gRPC telemetry | Sidecar: 7776 / 57776 |
| AM CLI | Single-binary cluster administration, 27 operator domains, rolling upgrades | Operator workstations, Bastion hosts, CI/CD runners | gRPC mTLS to Coordinator & StorageNodes | Client Binary (outbound) |
End-to-End Zero-Allocation Write Path
When an enterprise client or AWS S3 SDK writes an object into the platform, all six engines coordinate deterministically:
- S3 Ingest & Authentication (Aarkam.IO / Gateway): Client transmits an S3
PutObjectrequest over TLS 1.3 to port57772(or57771).Aarkam.IOverifies the AWS SigV4 HMAC-SHA256 signature in constant time usingCryptographicOperations.FixedTimeEquals. - Topology Resolution (Rokka.Coordinator):
Aarkam.IOqueries the active Raft MON leader over gRPC mTLS to resolve the current Consistent Hash Ring topology epoch and target storage nodes. - Client-Side Envelope Encryption (Aarkam.IO): A random 256-bit Data Encryption Key (DEK) is generated per object, wrapped via customer-owned HSM (PKCS#11 / KMIP), and payload chunks are encrypted with AES-256-GCM.
- Cauchy Erasure Coding (Ahmodi): The encrypted bitstream is divided into $K$ data chunks.
Ahmodi's AVX-512 vector kernel computes $M$ Cauchy parity chunks (e.g. 8+3 standard profile) at over 12.5 GB/s per core. - Parallel Network Streaming (Rokka): Chunks stream concurrently across internal gRPC mTLS channels directly to target
Rokka.StorageNodeinstances. - Direct I/O Persistence (Kdouja): Each storage node passes chunks to its embedded
Kdouja.Engine, which executes unbuffered Direct I/O (O_DIRECT) into pre-allocated 256MB segment containers, appending to the WAL and arena skip-list. - Autonomous Health Supervision (Abodi): Concurrently,
Abodievaluates physical drive S.M.A.R.T. counters and I/O latency jitter, feeding health scores intoRokka's consensus ring.
Architectural Isolation Guarantees
- Decoupled Control & Data Planes: Storage coordinator failure never halts active chunk read/write operations. Storage nodes continue serving cached hash ring epochs.
- Zero Inode Exhaustion: Kdouja packs billions of 4KB chunks into 256MB segment containers, reducing filesystem entries by 99.999% and eliminating operating system directory lock contention.
- Zero Memory Allocation on Hot Paths: All chunk transfers utilize
ReadOnlySpan<byte>and pooled unmanaged memory buffers, eliminating .NET Garbage Collection pauses. - Autonomous Pre-Failure Evacuation: Abodi identifies failing hardware up to 72 hours before catastrophic failure, triggering proactive Cauchy RS migrations without emergency reconstruction storms.