Ontology 存储结构与读写方式

一、OSv2 并不是单一的数据库或索引

总体架构图

• Ontology 后端是一组微服务协作程序。怎么说呢，Ontology 对象最终由 Object Storage V2 对外提供对象查询、搜索、关联遍历与访问能力。但 OSv2 不是传统意义上的"单一数据库"，而是围绕对象模型建立的存储、索引与查询运行程序。说到官方原话，"a system backed by multiple specialized object databases in parallel"——目标不是统一所有存储形态。而是按不同的访问模式、性能、索引、 需求挑不同的底层。
• 官方未公开 OSv2 底层用了哪一种引擎也没披露对象数据在物理层面的真实组织结构。其实，从公开文档和架构图能推断的是：
  • "Ontology Object" 是逻辑视图。不一定等同于底层物理结构;
  • OSv2 更像一层对象化的索引 + 查询运行时；
  • 底层事实数据可来自 Dataset、Streaming Datasource、RDBMS、NoSQL；不过，
  • Funnel 负责把这些数据编排、映射、索引为可查询的 Ontology Object；话说回来，
  • OSS则对外提供统一的对象查询接口。
• 读写职责分开: 写入由 Object Data Funnel`统一编排进对象库，查询由 Object Set Service `对外服务。这是 OSv2 相对 OSv1最显眼的架构变化。话说回来，

OSv2 涉及哪些主要服务？

下表列出 Ontology 读写过程中的主要服务；实际运行中涉及的服务可能更多。

OSv2 存储 / 索引架构推测

. OSv2 更像一个"对象语义索引运行时"

. 推测的分层架构

%{init:{'flowchart':{'rankSpacing':。'nodeSpacing':,'padding': }}}%
flowchart TB
%% =========================
%% Theme Styling
%% =========================
classDef source fill:#E3F2FD,stroke:#1E88E5,color:#0D47A1,stroke-width:1px;classDef funnel fill:#E8F5E9,stroke:#43A047。color:#1B5E20,stroke-width:1px;按理说，classDef metadata fill:#FFF3E0。stroke:#FB8C00,color:#E65100,stroke-width:1px;classDef storage fill:#F3E5F5。stroke:#8E24AA,color:#4A148C,stroke-width:1px;classDef index fill:#E0F7FA。stroke:#00ACC1,color:,stroke-width:1px;classDef runtime fill:#FCE4EC。stroke:#D81B60,color:#880E4F,stroke-width:1px;classDef api fill:#EFEBE9,stroke:#6D4C41。color:#3E2723,stroke-width:1px;%% =======================
%% Data Sources Layer
%% =======================
subgraph DS
D1
D2
D3
D4
end
%% =======================
%% Ingestion Layer
%% =======================
subgraph FL
F1
end
%% =======================
%% Metadata Layer
%% =======================
subgraph MD
M1
end
%% =======================
%% OSv2 Core
%% =======================
subgraph OS
subgraph ST
S1
S2
S3
end
subgraph IN
I1
I2
I3="" r1
R22

R33

end
end
%% =======================
%% API Layer
%% =======================
subgraph API
A11

A22

end
Flow diagram details omitted for brevity...
style DS fill:#F0F8FF;stroke:black;其实，stroke-dasharray:"";
style FL fill:white;怎么说呢，stroke:black;stroke-dasharray:"";style MD fill:white;不过，stroke:black;stroke-dasharray:"";style OS fill:white;stroke:black;class D10,D20,D30,D40 source;class F10 funnel;class M10 metadata;其实，class S10,S20,S30 storage;class I10,I20,I30 index;class R10,R20,R30 runtime;class A110,A210 api;

➚User Pain Point #1 – 学习曲线陡峭: The sheer number of services and layers often feels like navigating a multi‑tiered factory—developers struggle to map ir business logic onto this architecture. 

➚User Pain Point #2 – 性能疑问: The mix of KV stores,inverted indexes,graph indexes。and bitmap indexes can lead to confusion about which backend actually serves a given query type. 

. 推测专用对象库维度清单

➚User Pain Point #3 – 权限细粒度控制: The platform applies ACLs at both type and field level across multiple backends—keeping track of who can see what becomes tedious without a clear audit trail. 

二、多种来源的数据如何成为 Ontology 对象?

MDO和普通单源都遵循一样流程，但要额外满足以下约束：

• No MapType/StructType 列；
• Deterministic 主键；
• SRC 所在位置决定权限。Pain point: "When adding a new backing datasource you must re‑validate all existing objects because primary key could change。".

三，Funnel 数据管道四阶段流程示意

 CL
CL --> MG
MG --> IDX
IDX --> HY
HY --> OSV;

SRC -.-> CL;EDIT) -.-> MG;

NOTE right of CL : “Only diffs are persisted” NOTE right of MG : “Merge with edit edits” NOTE right of IDX : “Generate attribute & link indexes” NOTE right of HY : “Load into live cache”

% Flow triggers: SRC ->|New transaction| CL;EDIT ->|User action| MG;note left of HY : “Persisted asynchronously to dataset”  ，  ，

Pain point:
The four stages mean that a single write path involves multiple passes over data – developers worry about latency spikes when large changes occur.

四，三类查询对应何种物理指数?说起来，

#➚Pain point: "Mapping high‑level OSDK expressions onto low‑level indices requires deep knowledge of internal routing." 

五。“Search Around” 与 JOIN 的区别

typescript import { Objects } from '@foundry/sdk';const employee = Objects.search.employee.filter);const reports = employee.searchAroundToReport;

• No explicit JOIN syntax needed – relationship encoded in LinkTypes.
Pain point: "Developers accustomed to SQL joins may mistakenly attempt filters after Search Around instead of before," "causing unnecessary compute overhead." \

---

六。一致性 & 延迟：读后写 vs 列表分页扫描

• ACTION → Funnel → offset applied instantly => read-after-write guarantees.
Pain point: "For long list scans you cannot rely on consistency – paging may skip or duplicate objects if writes happen concurrently." • Solution hint: use single-object reads or pre-filter before pagination.

---

Sidenote:The term OQL is only used inside Agent Studio LLM tooling – not an alternative API for developers.

---

**Write Path Overview** mermaid sequenceDiagram autonumber participant CLI as Client Workshop/O SDK participant ACT as ActionsService participant OMS as OMSMetaDB participant FN as FunctionValidator participant FUN as ObjectDataFunnel participant OSV as OSVLiveIndex & PersistentStore participant AUD as AuditLog opt get contract -> OMS->ACT->OMS ACT->;Validate parameters & permissions opt includes function validation -> ACT->;FN->，FN returns VALID or INVALID;opt Write -> ACT->;FUN Queue，FUN->;FUN ApplyOffsetLive;FUN->，FUN PersistToDataset;opt side effects -> AUD record;AUD ->，AUD flush webhook if configured;END ACT->，CLI ValidateActionResponse \r\\ ApplyActionResponse\r\\ \r\\ Pain point #7: "Because all writes go through Actions/Funnel you have to understand offset sequencing else concurrent edits may conflict." Pain point #8: "Flush interval is 6 hrs by default – you might lose recent edits on crash unless you configure WAL."

---

**Security Model Across Path** * Note that MDO nullification happens inside OSS before returning result.

---

**Real world examples** *Military monitoring scenario* *Industrial manufacturing scenario*

---

**Comparison With Traditional Models**

---

**Commonly Confused Terms**

---

### 参考资料
• …etc ，

Where checked?不过，