Data Virtualization: Architecture, Use Cases & Limits

On this page本页目录

What Is Data Virtualization?什么是数据虚拟化？

This focused article is part of the federated queries and data virtualization guide; use the pillar guide to compare related concepts, methods, and implementation decisions across the full topic.

本文是联邦查询与数据虚拟化指南内容集群中的专题文章；如需比较完整主题下的相关概念、方法与实施决策，请返回基石指南。

Data virtualization is a data-access approach that presents governed logical views across one or more sources without requiring every dataset to be copied into a new store before it can be queried. A virtualization layer resolves source locations and schemas, applies access rules, plans distributed work, pushes supported operations toward sources, combines results, and exposes a stable interface to consumers.

数据虚拟化是一种数据访问方式，通过受治理的逻辑视图连接一个或多个来源，不要求先把每个数据集复制到新存储才能查询。虚拟化层解析源位置与Schema、应用访问规则、规划分布式工作、把受支持的操作下推到源系统、合并结果，并向使用者暴露稳定接口。

“Without moving data” is useful shorthand, not an absolute promise. Products may cache results, materialize hot subsets, spill intermediate data, or export outputs. The defining choice is that the logical model is not dependent on first centralizing every source. This makes data virtualization useful for current cross-source views, constrained movement, rapid integration, and coexistence during migration—but it also makes network behavior, connector capability, source capacity, and failure handling part of every query.

“不移动数据”是便于理解的简称，并非绝对承诺。产品可能缓存结果、物化热点子集、暂存中间数据或导出输出。关键选择在于：逻辑模型不依赖先集中每个来源。这使数据虚拟化适合当前跨源视图、受限数据移动、快速集成和迁移共存，但也意味着网络行为、连接器能力、源容量和失败处理都会进入每次查询。

Data Virtualization Architecture: The Required Layers数据虚拟化架构：必需层次

Connectors and source contracts连接器与源契约

Translate protocols, authentication, schemas, types, capabilities, and errors for databases, files, services, and other approved sources.

为数据库、文件、服务和其他获准来源转换协议、身份验证、Schema、类型、能力与错误。

Metadata and logical models元数据与逻辑模型

Describe virtual objects, relationships, business definitions, ownership, lineage, freshness, and versioned dependencies without hiding uncertainty.

描述虚拟对象、关系、业务定义、负责人、血缘、新鲜度和版本依赖，同时不掩盖不确定性。

Planner and optimizer规划器与优化器

Resolve views, estimate work, choose join placement and order, push down supported operations, control data movement, and allocate execution resources.

解析视图、估算工作、选择连接位置与顺序、下推受支持操作、控制数据移动并分配执行资源。

Policy and identity策略与身份

Authenticate consumers, authorize sources and fields, enforce row and column rules, mask sensitive values, and record audit evidence.

认证使用者、授权来源与字段、执行行列规则、遮蔽敏感值并记录审计证据。

Execution and acceleration执行与加速

Run distributed fragments, merge results, and optionally use caches, result reuse, materialized views, or workload routing under explicit freshness rules.

执行分布式片段、合并结果，并按明确新鲜度规则选择使用缓存、结果复用、物化视图或负载路由。

Consumption and observability使用与可观测性

Expose supported SQL, APIs, files, or tool connections while recording plans, latency, source load, failures, lineage, policy decisions, and cost signals.

暴露受支持的SQL、API、文件或工具连接，同时记录计划、延迟、源负载、失败、血缘、策略决定和成本信号。

How Data Virtualization Works at Query Time数据虚拟化在查询时如何工作

Resolve identity and policy. Authenticate the consumer, evaluate permitted virtual objects, rows, columns, purposes, and source credentials before planning.解析身份与策略。在规划前认证使用者，评估允许的虚拟对象、行、列、用途和源凭据。
Expand the logical request. Resolve view definitions, mappings, calculations, joins, filters, and dependent objects to create a logical plan.展开逻辑请求。解析视图定义、映射、计算、连接、过滤条件和依赖对象，形成逻辑计划。
Inspect connector capabilities. Determine which predicates, projections, aggregations, limits, functions, and joins each source can execute with equivalent semantics.检查连接器能力。确定每个来源能以等价语义执行哪些谓词、投影、聚合、限制、函数和连接。
Build a physical distributed plan. Choose pushdown, join order and location, parallelism, data movement, cache use, timeout, memory, and spill behavior.构建物理分布式计划。选择下推、连接顺序与位置、并行度、数据移动、缓存、超时、内存和暂存行为。
Execute guarded source requests. Apply limits, workload identity, retries only where safe, circuit breakers, and source-specific concurrency controls.执行受保护的源请求。应用限制、负载身份、仅在安全时重试、熔断器和源专属并发控制。
Combine and return results. Merge fragments, enforce remaining rules, attach freshness and provenance where supported, and deliver through the consumer interface.合并并返回结果。合并片段、执行剩余规则、在支持时附加新鲜度与来源，并通过使用接口交付。
Record the execution. Capture query plan, accessed sources, policy outcomes, bytes or rows moved, latency, cache state, failures, and lineage for diagnosis.记录执行。捕获查询计划、访问来源、策略结果、移动字节或行数、延迟、缓存状态、失败和血缘，用于诊断。

The planner can only push down work a connector and source support with compatible semantics. A filter that is safe for one database may not be safe for another because collation, null handling, time zones, or function behavior differ. Unsupported operations execute elsewhere, potentially moving far more data than the logical query suggests.

规划器只能下推连接器和来源支持且语义兼容的工作。由于排序规则、null处理、时区或函数行为不同，对一个数据库安全的过滤条件可能不适用于另一个数据库。不受支持的操作会在其他位置执行，移动的数据可能远超逻辑查询表面显示的数量。

Data Virtualization Use Cases That Fit适合数据虚拟化的场景

Fit depends on freshness, source tolerance, and workload shape适配度取决于新鲜度、源承受能力和负载形态
Use case场景	Why virtualization helps虚拟化的作用	Condition前提
Current cross-source operations当前跨源运营视图	Combines recent states without waiting for a full batch copy无需等待完整批量复制即可组合较新状态	Sources tolerate bounded queries and snapshot limitations are explicit来源能承受受限查询，并明确快照限制
Migration coexistence迁移共存	Presents a stable logical interface while domains move in phases领域分阶段迁移时保持稳定逻辑接口	Versioning, dual-read semantics, and retirement plan are governed版本、双读语义和退役计划受治理
Restricted movement数据移动受限	Executes approved access near data and returns minimized results在数据附近执行获准访问，仅返回最小结果	Policy can be enforced end to end; querying is actually permitted可端到端执行策略，且查询本身确实获准
Discovery and prototyping发现与原型	Tests cross-source value before building durable pipelines在建设持久管道前验证跨源价值	Prototype is not silently promoted without production controls原型不会在缺少生产控制时静默上线
Long-tail governed access长尾受治理访问	Avoids a separate copy pipeline for every low-frequency question避免为每个低频问题建设独立复制管道	Latency expectations allow live or selectively cached execution延迟期望允许实时或选择性缓存执行

A successful use case has a clear consumer, business grain, freshness need, source contract, performance objective, security rule, and failure response. “Create one view of all enterprise data” is not a testable use case until those decisions are narrowed.

成功用例必须明确使用者、业务粒度、新鲜度需求、源契约、性能目标、安全规则和失败响应。“建立所有企业数据的统一视图”在这些决定被收窄前，不是可测试用例。

When Data Virtualization Is the Wrong Default何时不应默认选择数据虚拟化

Fragile operational sources脆弱的运营来源

Avoid unbounded analytical scans when latency-sensitive systems lack replicas, quotas, workload isolation, or safe pushdown.

当延迟敏感系统缺少副本、配额、负载隔离或安全下推时，应避免无界分析扫描。

Stable heavy history稳定的重历史负载

Repeated full-history joins and aggregates often belong in a warehouse or lakehouse optimized for the workload.

重复的全历史连接与聚合通常更适合针对该负载优化的数仓或湖仓。

Strict predictable latency严格可预测延迟

A live plan inherits network and source variance. Materialize when the SLA cannot tolerate those dependencies.

实时计划会继承网络与来源波动；当SLA无法承受这些依赖时应物化。

Offline independence离线独立运行

If consumers must continue while sources are unavailable, durable copied data or an explicitly sufficient cache is required.

如果来源不可用时使用者仍必须继续工作，就需要持久复制数据或明确足够的缓存。

Other warning signs include unsupported proprietary APIs, incompatible transaction or snapshot semantics, complex transformations with poor pushdown, high egress costs, cross-region restrictions, uncontrolled ad hoc concurrency, and consumers who cannot understand freshness or partial-result behavior. A hybrid design is often better than an ideological “never copy” rule.

其他警示包括：不受支持的专有API、不兼容的事务或快照语义、难以下推的复杂转换、高出口成本、跨区域限制、失控的即席并发，以及使用者无法理解新鲜度或部分结果行为。混合设计通常优于意识形态式的“永不复制”。

What to Prepare Before Building a Virtualization Layer建设虚拟化层前需要准备什么

Consumer contract: personas, interfaces, query shapes, concurrency, result grain, freshness, latency, availability, and partial-result policy.使用者契约：角色、接口、查询形态、并发、结果粒度、新鲜度、延迟、可用性和部分结果政策。
Source inventory: owners, locations, schemas, volumes, update patterns, indexes, statistics, replicas, maintenance windows, quotas, credentials, and support boundaries.来源清单：负责人、位置、Schema、规模、更新模式、索引、统计信息、副本、维护窗口、配额、凭据和支持边界。
Meaning and mappings: keys, grain, units, code sets, time zones, join rules, precedence, null behavior, and accepted inconsistency.含义与映射：键、粒度、单位、代码集、时区、连接规则、优先级、null行为和可接受不一致。
Security: data classification, purpose, identity propagation, service accounts, row and column policy, masking, residency, retention, and audit needs.安全：数据分类、用途、身份传递、服务账号、行列策略、遮蔽、驻留、保留和审计需求。
Failure budget: source and network timeouts, retry safety, circuit breakers, stale cache rules, degraded mode, recovery ownership, and rollback.失败预算：来源与网络超时、重试安全、熔断、陈旧缓存规则、降级模式、恢复负责人和回滚。
Evidence set: representative and adversarial queries, expected results, query plans, source baselines, privacy-safe fixtures, and acceptance thresholds.证据集：代表性与刁钻查询、预期结果、查询计划、源基线、隐私安全样例和验收阈值。

How to Implement Data Virtualization Step by Step如何逐步实施数据虚拟化

Choose one bounded data product. Define the user decision, output grain, approved sources, freshness and latency objectives, and conditions that make the result unusable.选择一个边界明确的数据产品。定义用户决定、输出粒度、获准来源、新鲜度与延迟目标，以及让结果不可用的条件。
Baseline each source. Measure representative queries, indexes, statistics, concurrency headroom, maintenance periods, rate limits, and failure behavior before federation.建立每个来源的基线。在联邦前测量代表性查询、索引、统计信息、并发余量、维护时段、速率限制和失败行为。
Prove connector semantics. Test authentication, types, nulls, time zones, collation, precision, supported functions, pushdown, cancellation, and error translation.证明连接器语义。测试认证、类型、null、时区、排序规则、精度、支持函数、下推、取消和错误转换。
Layer virtual models. Separate source-faithful views, integration views, and consumer-facing products so mappings, policy, and breaking changes remain traceable.分层虚拟模型。分开源忠实视图、集成视图和面向使用者的产品，让映射、策略和破坏性变化保持可追溯。
Design the physical path. Decide pushdown, join placement, data transfer, parallelism, memory, spill, caching, materialization, and workload routing for named query patterns.设计物理路径。针对命名查询模式决定下推、连接位置、数据传输、并行度、内存、暂存、缓存、物化和负载路由。
Apply policy end to end. Enforce least privilege at both virtualization and source layers; test identity, row and column rules, masking, logging, and denied access.端到端应用策略。在虚拟化层和源层同时执行最小权限；测试身份、行列规则、遮蔽、日志与拒绝访问。
Test correctness, load, and failure. Compare trusted results, inspect plans, run concurrency and soak tests, interrupt sources, expire credentials, and verify degraded behavior.测试正确性、负载与失败。对比可信结果、检查计划、运行并发与耐久测试、中断来源、让凭据过期，并验证降级行为。
Release narrowly and observe. Start with approved consumers and quotas; monitor latency, source pressure, pushdown, transferred data, errors, policy events, freshness, and cost.小范围发布并观察。从获准使用者与配额开始，监测延迟、源压力、下推、传输数据、错误、策略事件、新鲜度和成本。
Materialize by evidence. Move stable or expensive patterns only when telemetry shows that copied data improves reliability, cost, or service objectives.依据证据物化。只有遥测表明复制数据能改善可靠性、成本或服务目标时，才移动稳定或昂贵模式。

Data Virtualization Example: Available-to-Promise Inventory数据虚拟化示例：可承诺库存

Hypothetical example: this architecture and formula illustrate decisions only. They are not an InfiniSynapse customer case, benchmark, or product guarantee.

假设示例：该架构和公式仅用于说明决定，不代表InfiniSynapse客户案例、基准测试或产品保证。

Suppose a service team needs current available-to-promise inventory by region. On-hand stock is in a PostgreSQL operational database, approved reservations are in a cloud warehouse, and relatively stable safety-stock policy is stored as partitioned files. A virtual product exposes one governed result without requiring a new full copy of all three sources.

假设服务团队需要按区域查看当前可承诺库存。现有库存位于PostgreSQL运营数据库，已批准预留位于云数仓，相对稳定的安全库存政策保存在分区文件中。虚拟数据产品暴露一个受治理结果，无需为三个来源新建完整副本。

Illustrative query-path decisions示意性查询路径决定
Stage阶段	Decision决定	Proof证明
Request请求	Require region and item filter; return result timestamp and source freshness要求区域与商品过滤；返回结果时间和来源新鲜度	Reject unbounded requests and verify policy before planning拒绝无界请求，并在规划前验证策略
Operational source运营来源	Push region, item projection, and supported aggregation to a read replica把区域、商品投影和受支持聚合下推到只读副本	Inspect the remote plan and source CPU, latency, rows, and cancellation检查远程计划、源CPU、延迟、行数与取消
Warehouse数仓	Aggregate approved reservations to the same item-region grain把已批准预留聚合到相同商品-区域粒度	Reconcile status filters and time-window semantics对账状态过滤与时间窗口语义
Policy files政策文件	Cache the approved small reference under a declared expiry and version按声明过期时间和版本缓存获准的小型参考数据	Test stale, missing, duplicate, and future-effective policy records测试陈旧、缺失、重复和未来生效的政策记录
Virtual result虚拟结果	Illustrative formula: on hand minus approved reservations minus safety stock示意公式：现有库存减已批准预留再减安全库存	Compare trusted fixtures and define behavior for negative or partial results比较可信样例，并定义负数或部分结果的行为

The source states may not represent one atomic snapshot. The product must state its consistency model, freshness per source, and whether a partial answer is prohibited. If the operational source cannot meet the guarded workload, the correct change may be a replica, change-data capture, or materialized inventory product—not a more aggressive virtual query.

这些源状态不一定代表同一个原子快照。数据产品必须说明一致性模型、每个来源的新鲜度，以及是否禁止部分答案。如果运营来源无法承受受保护负载，正确调整可能是增加副本、变更数据捕获或物化库存产品，而不是发出更激进的虚拟查询。

Data Virtualization vs Federation, ETL, Warehouses, and Semantic Layers数据虚拟化与联邦、ETL、数仓及语义层的区别

These patterns can be combined in one architecture这些模式可以在同一架构中组合
Pattern模式	Primary role主要作用	Main trade-off主要权衡
Data virtualization数据虚拟化	Governed logical access and reusable models across sources跨源受治理逻辑访问与可复用模型	Inherits source, connector, and network behavior at query time查询时继承来源、连接器和网络行为
Data federation数据联邦	Execute one query across multiple systems跨多个系统执行一个查询	Often focuses on query mechanics; broader modeling and governance may be separate通常聚焦查询机制；更广建模与治理可能另行实现
ETL or ELTETL或ELT	Move and transform data into a persistent target移动并转换数据到持久目标	Adds pipelines and copies but isolates consumption from live sources增加管道与副本，但让使用负载与实时来源隔离
Warehouse or lakehouse数仓或湖仓	Store governed history for optimized analytical workloads为优化分析负载存储受治理历史	Requires ingestion, storage, synchronization, and lifecycle control需要摄取、存储、同步和生命周期控制
Semantic layer语义层	Define governed business entities, dimensions, metrics, and consumption meaning定义受治理业务实体、维度、指标和使用含义	May run over virtual or physical data; does not itself guarantee source execution可运行在虚拟或物理数据之上；本身不保证源执行
Database view数据库视图	Abstract queries inside one database context在单一数据库上下文中抽象查询	Usually lacks heterogeneous cross-system connectors and distributed planning通常缺少异构跨系统连接器与分布式规划

Data fabrics and data meshes are broader organizational or architectural approaches, not substitutes with one fixed runtime behavior. A fabric may use virtualization, catalogs, integration, lineage, and automation; a mesh may publish domain data products through virtual or materialized paths. Evaluate the concrete execution and governance, not the label.

数据织物和数据网格是更广的组织或架构方法，不是具有固定运行行为的替代品。数据织物可使用虚拟化、目录、集成、血缘与自动化；数据网格可通过虚拟或物化路径发布领域数据产品。应评估具体执行与治理，而不是标签。

Design Performance Around Pushdown and Data Movement围绕查询下推与数据移动设计性能

Performance begins with the physical plan, not a generic promise that federation is fast. Inspect whether projections reduce columns, predicates reduce rows, limits and aggregations run remotely, and joins occur where they move the least data without changing semantics. Statistics must be sufficiently current for cost decisions, and the optimizer needs conservative fallbacks when they are missing.

性能始于物理计划，而不是“联邦很快”的笼统承诺。应检查投影是否减少列、谓词是否减少行、限制与聚合是否远程执行，以及连接是否在不改变语义的前提下于数据移动最少的位置发生。统计信息必须足够新以支持成本决定；缺失时优化器需采用保守回退。

Measure per query pattern: plan time, source time, transfer time, merge time, rows or bytes scanned and returned, spills, cache state, and consumer latency percentiles.按查询模式测量：规划时间、源时间、传输时间、合并时间、扫描与返回行数或字节、暂存、缓存状态和使用者延迟分位数。
Protect sources: use replicas where appropriate, bounded filters, quotas, priorities, concurrency pools, cancellation, timeouts, workload windows, and cost guards.保护来源：适当使用副本、有界过滤、配额、优先级、并发池、取消、超时、负载窗口与成本保护。
Cache by contract: define key, scope, maximum age, invalidation, privacy, isolation, warm-up, fallback, and lineage. A cache without a freshness contract creates hidden copies.按契约缓存：定义键、范围、最大年龄、失效、隐私、隔离、预热、回退和血缘。没有新鲜度契约的缓存会形成隐藏副本。
Materialize selectively: persist frequently repeated, expensive, stable, or offline-required outputs; keep volatile or movement-restricted fields virtual when evidence supports it.选择性物化：持久化频繁重复、昂贵、稳定或必须离线使用的输出；在证据支持时让易变或受移动限制字段保持虚拟。

Test cold and warm states separately. A demonstration powered by a warm cache, small sample, and idle sources does not prove production behavior under concurrent ad hoc queries or connector degradation.

应分别测试冷态与热态。由热缓存、小样本和空闲来源支撑的演示，不能证明并发即席查询或连接器退化时的生产行为。

Govern Access Across Every Layer跨每个层次治理访问

A virtual layer can centralize policy presentation, but it does not erase source authorization. Decide whether end-user identity is propagated, exchanged for a scoped token, or represented by a service account. Broad shared credentials can turn a logically restricted view into a high-impact bypass if source permissions, query rewriting, caches, exports, or logs are misconfigured.

虚拟层可以集中呈现策略，但不会消除源授权。必须决定最终用户身份是被传递、交换成限域令牌，还是由服务账号代表。若源权限、查询改写、缓存、导出或日志配置错误，宽泛共享凭据可能让逻辑受限视图变成高影响绕过通道。

Classify virtual objects and derived fields; trace them back to source classifications.分类虚拟对象与派生字段，并追溯到源分类。
Test row filters, column denial, masking, purpose limits, aggregation leakage, and inference through joins.测试行过滤、列拒绝、遮蔽、用途限制、聚合泄露和通过连接进行推断。
Keep credentials in managed secret systems; rotate and revoke them; do not embed them in views or client files.把凭据保存在受管理的密钥系统中，执行轮换与吊销，不要嵌入视图或客户端文件。
Audit requester, purpose, logical object, physical sources, policy decision, exported destination, and administrative changes.审计请求者、用途、逻辑对象、物理来源、策略决定、导出目的地和管理变更。
Apply residency and minimization to data in transit, cache, spills, temporary files, query history, samples, and logs—not only the original source.把驻留与最小化要求应用到传输数据、缓存、暂存、临时文件、查询历史、样本和日志，而不仅是原始来源。

How to Validate a Data Virtualization Deployment如何验证数据虚拟化部署

1Correctness正确性

2Performance性能

3Protection保护

4Recovery恢复

Correctness: compare trusted fixtures and materialized reference results; test keys, grain, types, nulls, time zones, collation, duplicates, late data, source skew, and inconsistent snapshots.正确性：比较可信样例和物化参考结果；测试键、粒度、类型、null、时区、排序规则、重复、延迟数据、源偏差和不一致快照。
Performance: inspect local and remote plans; record cold and warm latency percentiles, scanned and transferred data, planning time, spills, cache behavior, throughput, and concurrency.性能：检查本地与远程计划；记录冷热延迟分位数、扫描与传输数据、规划时间、暂存、缓存行为、吞吐与并发。
Source protection: monitor source CPU, memory, I/O, locks, connections, queueing, replica lag, API quotas, cancellation, and recovery after test load.源保护：监测源CPU、内存、I/O、锁、连接、排队、副本延迟、API配额、取消和测试负载后的恢复。
Security: prove allowed and denied identities, row and column policies, masking, cache isolation, export controls, secrets, audit completeness, and administrative separation.安全：证明允许与拒绝身份、行列策略、遮蔽、缓存隔离、导出控制、密钥、审计完整性和管理分离。
Reliability: interrupt each dependency; verify timeout, circuit breaker, error message, no false completeness, permitted stale behavior, retry safety, and recovery objectives.可靠性：中断每个依赖；验证超时、熔断、错误消息、不会虚假完整、获准陈旧行为、重试安全和恢复目标。
Evolution: change source fields, types, permissions, statistics, and connector versions; confirm impact detection, contract tests, rollback, lineage, and consumer communication.演进：改变源字段、类型、权限、统计信息和连接器版本；确认影响检测、契约测试、回滚、血缘和使用者沟通。

Set acceptance thresholds from the consumer contract before testing. An average latency alone is not enough; tails, failures, source pressure, freshness, correctness, and denied-access behavior decide whether the virtual product is safe.

应在测试前根据使用者契约设定验收阈值。仅有平均延迟不够；尾部延迟、失败、源压力、新鲜度、正确性和拒绝访问行为共同决定虚拟数据产品是否安全。

Explore Approved Multi-Source Data with InfiniSynapse使用InfiniSynapse探索获准的多源数据

InfiniSynapse’s public site describes direct connections to supported databases and multi-source federated analysis without requiring a complex migration first. For an analysis task, prepare approved source access, clear business definitions, join keys and grain, a bounded question, privacy rules, expected freshness, and validation examples. Then use the workspace to analyze relevant connected sources through natural-language questions.

InfiniSynapse官网说明其支持直接连接受支持数据库，并在无需先完成复杂迁移的情况下进行多源联邦分析。针对分析任务，请准备获准来源访问、清晰业务定义、连接键与粒度、有界问题、隐私规则、预期新鲜度和验证样例，再在工作区用自然语言问题分析相关已连接来源。

Do not treat that workflow as proof that InfiniSynapse is a general-purpose data virtualization platform. This page does not claim it publishes virtual SQL schemas or data services, exposes optimizer and pushdown controls, manages enterprise caches, replaces a warehouse, propagates identity to every source, enforces source policies, or guarantees cross-source transaction consistency. Confirm supported connectors and visible behavior for the actual task; keep architecture, credentials, approvals, load protection, and deployment in governed systems.

不要把该工作流视为InfiniSynapse是通用数据虚拟化平台的证明。本页不声称它会发布虚拟SQL Schema或数据服务、暴露优化器与下推控制、管理企业缓存、取代数仓、把身份传递到每个来源、执行源策略，或保证跨源事务一致性。应针对实际任务确认受支持连接器和可见行为，并把架构、凭据、批准、负载保护与部署留在受治理系统中。

Analyze a bounded question across approved sources跨获准来源分析边界明确的问题

Bring approved connections, definitions, keys, grain, privacy constraints, freshness expectations, and test examples. Use InfiniSynapse for the supported multi-source analysis task; validate results and retain infrastructure controls in their systems of record.

请准备获准连接、定义、键、粒度、隐私约束、新鲜度期望和测试样例。使用InfiniSynapse完成受支持的多源分析任务；验证结果，并把基础设施控制保留在其事实系统中。

Analyze approved connected sources分析获准的已连接来源

Data Virtualization FAQ数据虚拟化常见问题

What is data virtualization?

什么是数据虚拟化？

Data virtualization is a data-access approach that presents governed logical views across one or more sources while hiding location, format, and access differences. Queries are planned against the sources at request time, with selective caching or materialization used when required rather than copying every dataset first.

数据虚拟化是一种数据访问方式，通过受治理的逻辑视图连接一个或多个来源，同时隐藏位置、格式和访问差异。查询在请求时针对源系统规划，仅在需要时选择性缓存或物化，而不是先复制每个数据集。

How does data virtualization work?

数据虚拟化如何工作？

A virtualization layer uses connectors and metadata to expose logical views, applies identity and policy, creates a distributed query plan, pushes supported filters or computations to sources, combines returned results, and records execution, lineage, freshness, and failures.

虚拟化层使用连接器和元数据暴露逻辑视图，应用身份与策略，创建分布式查询计划，把受支持的过滤或计算下推到源系统，合并返回结果，并记录执行、血缘、新鲜度和失败。

What is the difference between data virtualization and ETL?

数据虚拟化与ETL有什么区别？

ETL or ELT moves and persists data in a target store before consumption. Data virtualization usually resolves a logical view at query time against existing sources. They are complementary: stable heavy workloads may be materialized while current or restricted data remains virtual.

ETL或ELT先把数据移动并持久化到目标存储，再供使用；数据虚拟化通常在查询时针对现有来源解析逻辑视图。两者可以互补：稳定重负载可物化，而要求当前状态或受移动限制的数据保持虚拟访问。

When should you not use data virtualization?

何时不应使用数据虚拟化？

Avoid relying on live virtualization alone when sources cannot tolerate analytical load, networks are unreliable, connectors cannot push down important work, consumers require predictable subsecond performance, long historical scans dominate, or the workload must continue while sources are offline.

当源系统无法承受分析负载、网络不可靠、连接器不能下推关键工作、使用者需要可预测的亚秒级性能、主要负载是长历史扫描，或源系统离线时业务仍必须运行，就不应只依赖实时数据虚拟化。

Authoritative Sources and Next Steps权威来源与下一步

Implementation details depend on the deployed engine and connector. AWS explains the logical abstraction approach and use cases; IBM discusses virtualization-layer latency, freshness, and source-strain design considerations; Trino documents which operations may be pushed through a connector and why semantics matter; Microsoft documents external-table data virtualization with PolyBase. These are official examples, not evidence that every product supports the same sources, pushdowns, caching, security, or consistency.

实施细节取决于实际部署的引擎与连接器。AWS说明逻辑抽象方式与用例；IBM讨论虚拟化层的延迟、新鲜度和源压力设计考量；Trino记录哪些操作可通过连接器下推以及语义为何重要；Microsoft记录使用PolyBase外部表进行数据虚拟化。这些是官方示例，并不证明每个产品都支持相同来源、下推、缓存、安全或一致性。