What Is Data Federation?什么是数据联邦？
Place this specific workflow in context with the federated queries and data virtualization guide, which connects the definitions, alternatives, validation steps, and related implementation guides.
可通过联邦查询与数据虚拟化指南理解本专题在整体流程中的位置；该指南串联了定义、替代方案、验证步骤与相关实施文章。
Data federation is a query-time integration pattern that lets one logical request access and combine data from multiple autonomous sources while most source data remains in place. A federation layer resolves logical objects to physical sources, decomposes the request, pushes supported filters or calculations toward those sources, transfers selected results, and merges them for the consumer.
数据联邦是一种查询时集成模式，让一个逻辑请求访问并组合多个自治来源的数据，同时让大部分源数据保留在原处。联邦层把逻辑对象解析到物理来源，分解请求，将受支持的筛选或计算下推到来源，传输经过选择的结果，再为使用者汇合结果。
The promise is not “zero movement.” Metadata, subqueries, intermediate rows, and final results still cross boundaries. Nor is federation automatically real time: freshness depends on each source, connector, transaction boundary, cache, and query start time. A trustworthy design makes those boundaries visible instead of hiding them behind a single endpoint.
数据联邦并不承诺“零移动”：元数据、子查询、中间行和最终结果仍会跨越边界。它也不自动等同于实时，因为新鲜度取决于各来源、连接器、事务边界、缓存和查询开始时间。可信设计会明确这些边界，而不是把它们藏在单一端点之后。
Data Federation vs. Virtualization, Replication, and ETL数据联邦与虚拟化、复制及ETL的区别
| Pattern模式 | Data placement数据位置 | Best fit适用场景 | Primary cost主要代价 |
|---|---|---|---|
| Data federation数据联邦 | Read selected data from autonomous sources at query time查询时从自治来源读取经过选择的数据 | Fresh, selective, cross-source questions新鲜、选择性强的跨源问题 | Network, source load, distributed failure网络、源端负载、分布式故障 |
| Data virtualization数据虚拟化 | Often uses federation underneath a broader logical delivery layer常在更广逻辑交付层下使用联邦机制 | Reusable logical models, views, catalogs, and interfaces可复用逻辑模型、视图、目录和接口 | Model governance and platform operations模型治理与平台运维 |
| Replication or CDC复制或CDC | Maintain copies in another system在另一系统维护副本 | Repeatable reads and source isolation with bounded lag有界延迟下的重复读取与源端隔离 | Storage, consistency, pipeline recovery存储、一致性、管道恢复 |
| ETL or ELTETL或ELT | Move and transform data into a target把数据移动并转换到目标 | Large repeated scans, history, curated analytics大范围重复扫描、历史记录、策展分析 | Pipeline latency, duplication, maintenance管道延迟、重复与维护 |
| Data mesh or data fabric数据网格或数据编织 | Organizational or architectural model; may use several access patterns组织或架构模型，可采用多种访问模式 | Ownership and interoperability across domains跨域所有权与互操作性 | Operating model, standards, adoption运作模式、标准与采用 |
A hybrid is normal: federate a small fresh dimension, materialize a heavy fact table, and replicate a source that cannot safely serve analytics. “Federation versus integration” is not absolute; federation is one integration pattern.
混合方式很常见：联邦一个小而新鲜的维度，物化重型事实表，并复制无法安全承载分析的来源。“联邦与集成”不是绝对二选一；联邦本身就是一种集成模式。
Design a Data Federation Architecture设计数据联邦架构
Names, owners, types, keys, units, freshness, classification, and physical locations let a planner resolve a request without pretending schemas are identical.
名称、所有者、类型、键、单位、新鲜度、分类和物理位置让规划器解析请求，而不假装各Schema完全相同。
The planner decomposes work, estimates cardinality and cost, chooses join order and location, and decides what can be pushed down or materialized.
规划器分解工作，估算基数与成本，选择连接顺序和位置，并决定哪些工作能够下推或物化。
Connectors translate capabilities, identity, types, predicates, cancellation, pagination, and errors. Workers combine returned streams under memory and spill limits.
连接器转换能力、身份、类型、谓词、取消、分页和错误；执行节点在内存和溢写限制下组合返回流。
Authorization, lineage, budgets, rate limits, audit events, health, timeouts, circuit breakers, cache rules, and evidence surround every plan.
授权、血缘、预算、限速、审计事件、健康状态、超时、断路器、缓存规则和证据围绕每个计划运行。
Keep the control path separate from the data path. A catalog outage should have a defined effect on cached plans; an execution worker should not silently bypass policy; and source credentials should be scoped to the connector and workload, not shared globally.
应分离控制路径与数据路径。目录故障对缓存计划的影响必须明确；执行节点不能静默绕过策略；来源凭据应限定到连接器和工作负载，而不是全局共享。
Prepare Inputs Before Implementing Data Federation实施数据联邦前准备输入
- Workload envelope: representative queries, consumers, concurrency, latency objective, freshness tolerance, scan range, output size, and failure behavior.工作负载范围：代表性查询、使用者、并发、延迟目标、新鲜度容忍、扫描范围、输出大小和故障行为。
- Source inventory: owners, engines, versions, regions, maintenance windows, read replicas, statistics, quotas, supported operations, and escalation contacts.来源清单：所有者、引擎、版本、区域、维护窗口、只读副本、统计信息、配额、支持操作和升级联系人。
- Join contract: keys, cardinality, null behavior, deduplication, time zones, units, codes, collation, decimal precision, and slowly changing definitions.连接契约：键、基数、空值行为、去重、时区、单位、代码、排序规则、小数精度和缓慢变化的定义。
- Security contract: identity mapping, least privilege, row and column policy, purpose, residency, encryption, logging, retention, and incident response.安全契约：身份映射、最小权限、行列策略、用途、驻留、加密、日志、保留和事件响应。
- Acceptance baseline: authoritative control queries, expected invariants, source-load ceiling, bytes-moved budget, latency percentiles, and rollback owner.验收基线：权威对照查询、预期不变量、源端负载上限、移动字节预算、延迟分位数和回退负责人。
If ownership or semantics are unknown, federation only makes ambiguity faster. Resolve the contract before adding a convenient shared endpoint.
如果所有权或语义未知，联邦只会更快地传播歧义。应先解决契约，再增加方便的共享端点。
How Data Federation Works from Request to Result数据联邦如何从请求走到结果
- Parse and bind.解析与绑定。 Resolve logical names, functions, identities, and policy context.解析逻辑名称、函数、身份和策略上下文。
- Discover capabilities.发现能力。 Read schemas, statistics, connector features, locality, and freshness state.读取Schema、统计信息、连接器能力、位置和新鲜度状态。
- Decompose.分解。 Split the logical plan into source-specific subplans and federation-stage work.把逻辑计划拆分为来源子计划和联邦阶段工作。
- Optimize.优化。 Push projections, predicates, limits, aggregations, or same-source joins only when supported and correct.仅在受到支持且语义正确时下推列、谓词、限制、聚合或同源连接。
- Execute under guardrails.在护栏内执行。 Open scoped sessions, enforce budgets, stream pages, retry safe operations, and propagate cancellation.打开限定会话，执行预算，流式读取分页，仅重试安全操作并传播取消。
- Reconcile and combine.协调并组合。 Convert types, normalize semantics, join or union intermediates, and record provenance.转换类型、规范语义、连接或合并中间结果，并记录来源。
- Return with evidence.带证据返回。 Expose freshness, partial-result status, plan, source timing, bytes moved, policy decisions, and query identifier.公开新鲜度、部分结果状态、计划、来源耗时、移动字节、策略决策和查询标识。
Implement Data Federation in Seven Controlled Steps用七个受控步骤实施数据联邦
- Choose one bounded question. Start with a selective read-only use case whose source owners and truth checks are known.选择一个有边界的问题。从来源所有者和真值检查已知的选择性只读用例开始。
- Prove connector semantics. Test authentication, TLS, metadata discovery, types, nulls, time zones, collation, cancellation, paging, and error translation.证明连接器语义。测试认证、TLS、元数据发现、类型、空值、时区、排序规则、取消、分页和错误转换。
- Publish a minimal logical contract. Expose only needed objects and document keys, units, freshness, owners, and policy.发布最小逻辑契约。只暴露必要对象，并记录键、单位、新鲜度、所有者和策略。
- Capture a baseline. Run source-native control queries and record results, timing, plans, and source metrics.捕获基线。运行来源原生对照查询，并记录结果、耗时、计划和来源指标。
- Enable pushdown deliberately. Compare explain plans and bytes moved for supported and deliberately unsupported expressions.有意识地启用下推。比较受支持及故意不受支持表达式的执行计划和移动字节。
- Exercise failure and concurrency. Add slow responses, schema change, expired credentials, network interruption, cancellation, and peak parallel requests.演练故障与并发。加入慢响应、Schema变化、凭据过期、网络中断、取消和峰值并行请求。
- Approve and monitor an envelope. Set source, network, latency, concurrency, and result-size limits; route outliers to materialization.批准并监控工作负载范围。设置来源、网络、延迟、并发和结果大小限制，把异常负载路由到物化。
Protect Correctness Across Types, Time, and Semantics跨类型、时间与语义保护正确性
A syntactically successful cross-source join can still be wrong. Verify decimal rounding, timestamp zones and precision, string collation and case, character encoding, boolean conventions, null ordering, identifier normalization, and unsupported types. Google BigQuery documentation, for example, notes that external results are converted to GoogleSQL types and unsupported types can fail; this is a product-specific illustration of a general boundary, not a universal implementation rule.
语法成功的跨源连接仍可能在业务上错误。应验证小数舍入、时间戳时区与精度、字符串排序和大小写、字符编码、布尔约定、空值排序、标识符规范化及不受支持类型。例如，Google BigQuery文档说明外部结果会转换为GoogleSQL类型，不受支持类型可能失败；这是一般边界的产品实现示例，并非所有系统的统一规则。
Then validate meaning: “active customer,” revenue date, currency basis, deletion status, and regional code may differ by source. Create explicit mappings, effective dates, owners, and exception paths. Reconcile row counts, key coverage, aggregates, duplicate rates, and unmatched join keys against source-native controls. Preserve the query start time and each source observation time so consumers know whether the combined result is a transactionally consistent snapshot or a best-effort observation.
还要验证含义：“活跃客户”、收入日期、币种基础、删除状态和地区代码在不同来源可能不同。应建立明确映射、生效日期、所有者和异常路径，并使用来源原生对照核对行数、键覆盖、聚合、重复率和未匹配连接键。保留查询开始时间及各来源观察时间，让使用者知道组合结果是事务一致快照，还是尽力而为的观察。
Optimize Pushdown Without Harming Source Systems在不伤害源系统的前提下优化下推
Predicate and projection pushdown reduce rows and columns crossing the network. Depending on connector capability and semantics, aggregations, limits, top-N operations, or same-source joins may also be delegated. Trino documents that support is connector-specific and recommends reading the explain plan; PostgreSQL’s foreign-data wrapper similarly limits remote clauses when safety conditions are not met. Never infer pushdown from a green query alone.
谓词和列下推会减少跨网络的行列数。根据连接器能力和语义，还可委托聚合、限制、Top-N或同源连接。Trino文档说明支持情况因连接器而异，并建议读取执行计划；PostgreSQL外部数据包装器同样会在不满足安全条件时限制远程子句。不能仅根据查询成功就推断已下推。
Measure source CPU, I/O, connections, locks, queue time, scanned rows, returned rows, bytes moved, spill, memory, and latency percentiles together. Use read replicas where appropriate, workload groups, per-source concurrency, query timeouts, scan and result caps, admission control, and cancellation propagation. Statistics must be fresh enough for join ordering; when estimates are unreliable, prefer guarded plans or pre-materialize the risky side.
应同时测量源端CPU、I/O、连接、锁、排队时间、扫描行、返回行、移动字节、溢写、内存和延迟分位数。适当使用只读副本、工作负载组、每来源并发、查询超时、扫描与结果上限、准入控制和取消传播。统计信息必须足够新鲜以支持连接排序；估算不可靠时，应选择有护栏的计划或预先物化风险侧。
Carry Identity, Policy, and Lineage Through Federation在联邦中传递身份、策略与血缘
Decide whether the source sees the end-user identity, a scoped service identity, or a delegated token. Document where row and column policy is enforced and which layer wins if rules conflict. A federation endpoint must not widen access merely because two individually permitted sources can now be joined; combined data can create new sensitive inferences.
应决定来源看到终端用户身份、限定服务身份还是委托令牌，并记录行列策略在哪一层执行，以及规则冲突时哪一层优先。联邦端点不能因为两个来源分别允许访问，就自动扩大权限；组合数据可能产生新的敏感推断。
Encrypt control and data paths, isolate connector secrets, rotate credentials, restrict egress, log source objects and policy decisions, and correlate central and source audit identifiers. Record logical-to-physical lineage, transformations, data location, purpose, freshness, cache or temporary-result retention, and deletion. Test denied columns, denied rows, revoked users, expired tokens, and direct-source versus federated equivalence.
加密控制与数据路径，隔离连接器密钥，轮换凭据，限制出站，记录来源对象和策略决策，并关联中央与来源审计标识。记录逻辑到物理血缘、转换、数据位置、用途、新鲜度、缓存或临时结果保留及删除。测试拒绝列、拒绝行、已撤销用户、过期令牌以及直连与联邦权限等价性。
Design for Slow Sources, Schema Change, and Partial Results为慢源、Schema变化与部分结果设计
One slow or unavailable source can hold the whole result open. Define per-source connect and read timeouts, overall deadlines, cancellation, bounded retries with jitter, circuit breakers, concurrency isolation, backpressure, and spill limits. Retry only idempotent read work, and never return partial data as a complete answer. If partial results are allowed, label missing sources, coverage, observation time, and degraded semantics in the response.
一个缓慢或不可用的来源就可能阻塞整个结果。应定义每来源连接与读取超时、整体截止时间、取消、带抖动的有界重试、断路器、并发隔离、背压和溢写限制。只重试幂等读取，绝不能把部分数据当作完整答案返回。若允许部分结果，响应中必须标注缺失来源、覆盖范围、观察时间和降级语义。
Schema evolution is also a failure mode. Version connector contracts, detect additions, removals, type changes, renamed objects, and capability regressions, then quarantine incompatible changes. Cache metadata and plans only with explicit invalidation. During a catalog or policy outage, choose fail-closed behavior for sensitive access and document whether already running queries are cancelled.
Schema演进也是一种故障。应为连接器契约设定版本，检测新增、删除、类型变化、对象重命名和能力退化，再隔离不兼容变化。元数据和计划缓存必须有明确失效机制。目录或策略服务故障时，敏感访问应选择关闭失败，并说明正在运行的查询是否取消。
Use a Decision Framework for Federation or Materialization用决策框架选择联邦或物化
| Signal信号 | Favor federation偏向联邦 | Favor movement or materialization偏向移动或物化 |
|---|---|---|
| Freshness新鲜度 | Latest source state is essential最新来源状态至关重要 | A governed snapshot or bounded lag is acceptable可接受受治理快照或有界延迟 |
| Selectivity选择性 | Strong filters and small results强筛选与小结果 | Repeated wide scans or large shuffles重复大范围扫描或大型重分布 |
| Source impact来源影响 | Read capacity is isolated and controlled读取容量已隔离并受控 | Operational SLOs or locks are at risk业务SLO或锁存在风险 |
| Consistency一致性 | Best-effort current observation is valid当前尽力观察有效 | One repeatable cross-source point is required需要一个可重复的跨源时间点 |
| Availability可用性 | All required sources meet the query SLO所有必需来源满足查询SLO | Consumers need independence from source outages使用者需要独立于来源故障 |
Materialize only the expensive or unstable fragment when possible. This preserves freshness for selective dimensions while isolating large facts or unreliable sources.
可行时只物化昂贵或不稳定的片段，从而为选择性维度保留新鲜度，同时隔离大型事实或不可靠来源。
Hypothetical Example: Federated Order Availability假设示例：联邦订单可用量分析
Illustrative scenario, not a customer case: an analyst needs open-order exposure by customer tier and current stock. Orders are in PostgreSQL, customer tiers are in a cloud warehouse, and available inventory is exposed by a read-only operational API. The team defines customer_id, SKU, currency conversion time, order status, and inventory observation time before writing the federated query.
这是说明性场景，不是客户案例：某分析师需要按客户等级和当前库存查看未结订单风险。订单位于PostgreSQL，客户等级位于云数据仓库，可用库存由只读业务API提供。团队在编写联邦查询前，先定义customer_id、SKU、币种转换时间、订单状态和库存观察时间。
The planner pushes the open-status and date filters plus order aggregation to PostgreSQL, projects only tier and customer key from the warehouse, and requests only relevant SKUs from the API. The returned aggregates and dimensions are joined centrally. Illustrative acceptance thresholds: no more than 20 concurrent queries, a 60-second deadline, a documented API call budget, zero unexplained reconciliation difference, and explicit failure if inventory is unavailable. These numbers are examples, not product guarantees.
规划器把未结状态、日期筛选和订单聚合下推到PostgreSQL，只从仓库投影等级与客户键，并仅向API请求相关SKU，再在中央连接返回的聚合和维度。示例验收阈值：并发不超过20、截止时间60秒、明确的API调用预算、无法解释的核对差异为零，并在库存不可用时显式失败。这些数字只是示例，不是产品保证。
Load tests reveal that month-end scans exceed the source envelope. The team materializes the historical order aggregate nightly while keeping customer tier and current inventory federated. The result is a workload-specific hybrid, not a claim that one pattern wins everywhere.
负载测试发现月末扫描超出源端范围。团队于是每晚物化历史订单聚合，同时继续联邦客户等级和当前库存。最终得到的是针对工作负载的混合方案，而不是某一种模式永远优胜的结论。
Validate Data Federation with Reproducible Evidence用可重复证据验证数据联邦
| Area领域 | Test测试 | Blocking evidence阻断证据 |
|---|---|---|
| Correctness正确性 | Reconcile keys, rows, aggregates, nulls, types, time zones, and observation times核对键、行、聚合、空值、类型、时区和观察时间 | Unexplained difference or hidden partial result无法解释的差异或隐藏的部分结果 |
| Plan计划 | Compare source-native and federation explain evidence and bytes moved比较来源原生与联邦执行证据及移动字节 | Unexpected full scan or unsafe pushdown意外全表扫描或不安全下推 |
| Performance性能 | Run representative mixes at peak concurrency while recording source and federation metrics在峰值并发运行代表性组合并记录来源及联邦指标 | Source ceiling, latency objective, or budget exceeded超出来源上限、延迟目标或预算 |
| Security安全 | Compare direct and federated permissions; revoke identities and inspect audit lineage比较直连和联邦权限；撤销身份并检查审计血缘 | Widened access, missing audit, or retained secret权限扩大、审计缺失或密钥残留 |
| Resilience韧性 | Inject latency, timeout, schema drift, expired credentials, cancellation, and unavailable sources注入延迟、超时、Schema漂移、凭据过期、取消和来源不可用 | Silent incompleteness, retry storm, or leaked resources静默不完整、重试风暴或资源泄漏 |
Version evidence by query, connector, engine, source topology, schema version, statistics snapshot, policy, and workload envelope. Re-run after connector upgrades, source patches, contract changes, and materialization changes.
证据应按查询、连接器、引擎、来源拓扑、Schema版本、统计快照、策略和工作负载范围进行版本化。连接器升级、来源补丁、契约变化和物化调整后都应重新运行。
When Data Federation Is the Wrong Path何时不应采用数据联邦
- Repeated full scans: move or materialize the data rather than repeatedly consuming network and source capacity.重复全量扫描：应移动或物化数据，而不是反复消耗网络和来源容量。
- Strict historical consistency: use governed snapshots or pipelines when every source must represent the same repeatable business cutoff.严格历史一致性：当每个来源必须代表同一可重复业务截止点时，应使用受治理快照或管道。
- Unreliable or rate-limited sources: isolate consumers through a copy, cache, or event-driven integration if source availability cannot meet the service objective.不可靠或受限速来源：若来源可用性无法满足服务目标，应通过副本、缓存或事件驱动集成隔离使用者。
- Unresolved semantics or ownership: a unified endpoint cannot repair conflicting definitions, unknown keys, or missing accountability.语义或所有权未解决：统一端点无法修复冲突定义、未知键或责任缺失。
- Write coordination: distributed writes and transactions require a separate architecture; do not infer them from read federation.写入协调：分布式写入和事务需要独立架构，不能从只读联邦推断得到。
Use InfiniSynapse for an Approved Multi-Source Analysis使用InfiniSynapse执行获批的多来源分析
InfiniSynapse’s public site presents direct connections to supported platforms and multi-source joint analysis without requiring a complex migration. This makes it a related analysis entry for a bounded multi-source question. This guide does not claim that InfiniSynapse is a general-purpose federation engine, connector SDK, catalog, semantic layer, policy engine, optimizer, or replacement for source controls.
InfiniSynapse官网说明其可直接连接受支持的平台，并在不要求复杂迁移的情况下进行多来源联合分析，因此它可作为有边界多来源问题的相关分析入口。本指南不声称InfiniSynapse是通用联邦引擎、连接器SDK、目录、语义层、策略引擎、优化器或源端控制替代品。
Before starting, prepare the approved sources, supported connection details, least-privilege credentials, certificate and network requirements, owner approval, join keys, definitions, freshness expectations, query scope, result limits, and source-load budget. Keep architecture, policy enforcement, monitoring, and recovery with the responsible source and platform teams.
开始前，请准备获批来源、受支持连接详情、最小权限凭据、证书与网络要求、所有者批准、连接键、定义、新鲜度预期、查询范围、结果限制和源端负载预算。架构、策略执行、监控和恢复仍由负责的来源与平台团队管理。
When the sources, access, semantics, and workload guardrails are ready, open InfiniSynapse to perform the authorized analysis within the confirmed product workflow.
当来源、访问、语义和工作负载护栏准备就绪后，可打开InfiniSynapse，在已确认的产品工作流内执行获批分析。
Analyze approved data sources分析获批数据来源Data Federation FAQ数据联邦常见问题
What is data federation?
什么是数据联邦？
Data federation is a query-time integration pattern that gives users one logical request path across autonomous data sources. A federation layer discovers source metadata, plans source-specific subqueries, pushes supported work toward each source, transfers selected results, and combines them without first requiring every source record to be copied into one central store.
数据联邦是一种查询时集成模式，让用户通过一个逻辑请求路径访问多个自治数据源。联邦层发现来源元数据，规划针对不同来源的子查询，把受支持的工作下推到各来源，传输经过筛选的结果并进行合并，而不要求先把每条源记录复制到一个中央存储。
How does data federation work?
数据联邦如何工作？
A request is parsed against a logical contract, resolved to physical sources, checked against policy, decomposed into subqueries, optimized for pushdown and data movement, executed through connectors, and merged into a result. The system must also reconcile types and semantics, protect source capacity, handle partial failure, and expose execution evidence.
请求先依据逻辑契约解析并映射到物理来源，经过策略检查后分解为子查询，再围绕下推和数据移动进行优化，通过连接器执行并汇合结果。系统还必须协调类型和语义、保护来源容量、处理部分失败，并提供执行证据。
What is the difference between data federation and data virtualization?
数据联邦与数据虚拟化有什么区别？
Data federation usually names the query mechanics for accessing and combining distributed sources. Data virtualization is often the broader delivery layer around those mechanics, adding stable logical models, reusable views, catalogs, governance, and consumer interfaces. Product terminology varies, so evaluate observable behavior rather than labels.
数据联邦通常指访问和组合分布式来源的查询机制；数据虚拟化通常是在这些机制之上提供的更广交付层，包括稳定逻辑模型、可复用视图、目录、治理和消费接口。产品术语并不统一，因此应评估可观察行为，而不是只看名称。
How is data federation different from replication or ETL?
数据联邦与复制或ETL有什么不同？
Federation reads selected data at query time and commonly leaves systems of record in place. Replication maintains copied data elsewhere, while ETL or ELT moves and transforms data into a target for repeated use. Federation favors freshness and source autonomy; movement patterns favor predictable performance, workload isolation, and historical consistency.
联邦在查询时读取经过筛选的数据，通常让记录系统保留在原处；复制会在别处维护数据副本；ETL或ELT会把数据移动并转换到目标中供重复使用。联邦偏向新鲜度和来源自治，移动模式偏向可预测性能、负载隔离和历史一致性。
When should data be federated instead of moved?
何时应联邦数据而不是移动数据？
Federate when the workload is selective, source data must remain authoritative, freshness matters, copying is restricted or disproportionate, and source capacity can be protected. Move or materialize data when queries repeatedly scan large ranges, require stable historical snapshots, need independent availability, or would overload operational systems.
当负载具有选择性、源数据必须保持权威、新鲜度重要、复制受限或代价不成比例，并且能够保护来源容量时，可以采用联邦。当查询反复扫描大范围、需要稳定历史快照、要求独立可用性，或会压垮业务系统时，应移动或物化数据。
How do you test data federation for production?
如何对数据联邦进行生产验收？
Use representative queries and concurrency to verify result reconciliation, plan and pushdown evidence, bytes moved, source load, latency percentiles, permission equivalence, type and time-zone handling, stale metadata behavior, cancellation, timeout, retry, slow-source isolation, and recovery. Approve a workload envelope, not an unlimited platform claim.
使用代表性查询和并发验证结果核对、执行计划和下推证据、移动字节数、源端负载、延迟分位数、权限等价性、类型与时区处理、陈旧元数据行为、取消、超时、重试、慢源隔离和恢复。批准的是明确的工作负载范围，而不是无限能力声明。
Official Sources and Verification Notes官方来源与验证说明
- Amazon Athena documentation for federated queries, connectors, pushdown, and limitationsAmazon Athena联邦查询、连接器、下推与限制文档
- Google BigQuery introduction to federated queries, type conversion, SQL pushdown, and limitsGoogle BigQuery联邦查询、类型转换、SQL下推与限制介绍
- Trino documentation for predicate, projection, aggregation, join, limit, and top-N pushdownTrino谓词、列、聚合、连接、限制和Top-N下推文档
- PostgreSQL foreign-data wrapper documentation for remote query optimization and execution semanticsPostgreSQL外部数据包装器远程查询优化与执行语义文档
These first-party documents demonstrate concrete implementations, not a universal feature set. Connector behavior, supported operations, locations, limits, security, and pricing can change; verify the exact release and deployment you plan to operate.
这些第一方文档展示具体实现，并不代表统一功能集。连接器行为、支持操作、位置、限制、安全和价格都可能变化；请针对计划运行的确切版本与部署重新验证。