Core Technology

FDP-Aware Placement Engine

DataFount goes beyond simple Hot/Cold tiering. We productize FDP with 6-class classification, RUH pool management, and predictive placement powered by online bandit algorithms.

6-Class Classification System

Class A: Interactive Secret ReadPID-0

Read-centric, small random I/O (≤16KB)

p99 Latency

Class B: Secret Update ChurnPID-1

Frequent small overwrites (≤32KB), high overwrite rate

WAF Reduction

Class C: Append-only Audit LogPID-2

Sequential append, high sequential score (≥0.8)

Minimize Mixing

Class D: Metadata IndexPID-3

Mid-size I/O (16KB–256KB) with periodic compaction

Tail Latency

Class E: Policy/Model CachePID-4

Short TTL, frequent eviction, non-critical QoS

GC Isolation

Class F: Attachment BlobPID-5

Large blobs (≥1MB), rare rewrites

Transfer Throughput

Classification Logic Flow

1

Incoming I/O Request

Input: op_type, io_size, object_type, ttl_hint, sequential_score, overwrite_rate

2

Rule Layer 1: Object Type Priority

audit_log → C  |  blob / ≥1MB → F  |  cache / short TTL → E

3

Rule Layer 2: Secret vs Metadata

secret+read+small → A  |  secret+write+churn → B  |  metadata+compaction → D

4

Tie-Break: Score Layer

Per-class scoring → class = argmax(score_A .. score_F)

4-Module Architecture Pipeline

Module 1

FDP Capability Detector

Detect SSD FDP support

Queries PIDs, RUH status via nvme-cli / xNVMe

Module 2

Placement Policy Engine

6-class classification + PID assignment

Rule-based → Score tie-break → Bandit tuner

Module 3

RUH Pool Manager

Per-class pool lifecycle

Pool exhaustion, rotation, throttling, compaction

Module 4

I/O Execution Layer

Userspace I/O path

xNVMe / io_uring_cmd / SPDK

RUH Pool Management

Short-term Pool(Class B, E)

Fast rotation, GC self-contained

Mid-term Pool(Class F)

Chunk/delta/compression pipeline

Long-term Pool(Class C)

Weekly/monthly segments

Compaction Pool(Class D)

Isolated compaction windows

Class A — Minimal write pool, primarily read cache/prefetch integration