DP-700 Fabric 数据工程师 中英双语通关课 (Bilingual)

Every new workspace gets a starter pool. Switch to a custom pool only when a job needs specific node sizes, and expect slower session start.

The Jobs tab controls admission. Optimistic job admission is the default; turning on Reserve maximum cores switches it off. Session timeout sets when idle notebook sessions expire.

每个新工作区都自带入门池。只有作业需要特定节点规格时才换自定义池，代价是会话启动变慢。作业页控制准入：默认是乐观准入，打开“预留最大核心”就关掉它；会话超时决定空闲笔记本会话多久过期。

When a workspace joins a domain, every item in it gets the domain as metadata. Users can then filter the OneLake catalog by domain. It changes discovery and delegated settings, never who can read the data. Workspace roles and item permissions still decide access.

工作区挂到域以后，里面所有项目都带上这个域标签，用户可以在 OneLake 目录里按域筛选。它只影响查找和下放的设置，不影响谁能读数据。子域没有自己的管理员，沿用父域的。

Reading through an external shortcut copies each file into a workspace cache. Later reads come from the cache instead of the remote cloud, so you pay egress once. Turn it on in the OneLake tab of workspace settings, pick a retention period, and use Reset cache when you need a clean start.

通过外部快捷方式读文件时，文件会存进工作区缓存，之后再读就直接从缓存拿，不再从外部云下载，流出费只付一次。保留期 1 到 28 天，每次读取都重新计时；源端有更新时会重新拉取并刷新缓存。

An Apache Airflow job runs DAGs (directed acyclic graphs, 有向无环图) on a pool set at workspace level. If customizing compute for items is off, every environment uses the starter pool. Pick the pool by how long the runtime must stay available and how many DAGs must run together.

Airflow 作业在工作区级别设定的运行池上跑 DAG。如果关闭了“为项目自定义计算”，所有环境都用入门池。入门池适合开发，自定义池适合生产，需要更多并发就加额外节点。

A workspace connects to exactly one branch and folder. Admins connect and disconnect it. Contributors can commit and update when they have write access to every item. The Source control panel shows which items are new, modified, deleted or in conflict.

For isolated work, branch out to another workspace. You get a new branch and a linked workspace while the shared one stays untouched.

一个工作区只连一个分支和文件夹。连接和断开只有管理员能做；对所有项目有写权限的参与者可以提交和更新。要隔离开发就用“分支到另一个工作区”，会新建分支和关联工作区，原工作区不受影响。冲突时可以接受传入更改、保留当前内容，或在 Git 里合并解决。

Database projects let you treat schema like application code. You can start empty or extract a project from an existing warehouse, edit the files, build, and then publish. Fabric Git integration also stores a SQL database as a SQL project, so the same files work in Git, pipelines and SqlPackage.

Two publish settings matter most. BlockOnPossibleDataLoss stops a deployment that would drop data. DropObjectsNotInSource removes target objects missing from the project.

数据库项目让架构像代码一样管理：可以新建，也可以从现有仓库提取，改文件、编译、发布。Fabric 的 Git 集成也把 SQL 数据库存成 SQL 项目。BlockOnPossibleDataLoss 会阻止可能丢数据的部署，生产环境建议开；DropObjectsNotInSource 会删掉项目里没有的目标对象，要慎用。

Create the pipeline, name its stages, then assign a workspace to each stage. Paired items in the target are overwritten when you deploy, while their ID, URL and permissions stay. Compare first to see what will change.

Deployment rules swap stage specific values. A semantic model can point to a production database, and a notebook can bind to a production lakehouse through a default lakehouse rule. Rules take effect on the next deployment.

先建管道、命名阶段，再给每个阶段分配工作区。部署时目标阶段里配对的项目被覆盖，但 ID、URL 和权限不变；部署前先对比。部署规则用来替换各阶段的值，比如语义模型指向生产数据库、笔记本绑定生产湖仓，下次部署才生效。

Workspace roles are the broadest access control in Fabric. A role applies to all items, current and future, so give the lowest role that does the job and use item permissions for single items. Assign roles to groups to keep membership manageable.

工作区角色是最粗粒度的权限，作用于工作区里现在和以后的所有项目，所以按最低够用原则分配，单个项目用项目权限。查看者能连 SQL 分析端点用 T-SQL 读数据，但不能用 Spark 或湖仓浏览器读。

Item permissions are for people who need one item, not the whole workspace. Only workspace Admins and Members can share, and each recipient always gets Read. Pick the extra permissions by how the person will read the data.

项目权限给只需要单个项目的人。只有工作区管理员和成员能共享，收件人总会得到读取权限。用 T-SQL 查全部表加 ReadData，用 Spark 读文件加 ReadAll，只查少数对象就只给读取再用 GRANT 授权。

Start from the minimum. Share the warehouse with Read only, which is just CONNECT, then grant the objects a user needs. Fabric creates the database user on the first GRANT, so there is no CREATE USER. Add column and row rules on top.

先给最小权限：只共享读取，再按需 GRANT 对象。第一次 GRANT 时 Fabric 自动建数据库用户，不能 CREATE USER。列级安全用列清单授权，查 SELECT * 会报错；行级安全用安全策略和内联表值函数，对 dbo 也生效。

You add a mask with ALTER TABLE and ALTER COLUMN, or define it when the table is created. Users with SELECT see masked values. Workspace Admins, Members and Contributors have CONTROL, so they always see real data.

Adding or changing a mask needs ALTER ANY MASK and ALTER on the table.

用 ALTER TABLE ... ALTER COLUMN ... ADD MASKED WITH 加掩码，也可以建表时定义。有 SELECT 的用户看到的是遮住的值；工作区管理员、成员、参与者有 CONTROL，看到的是真实值。添加或修改掩码需要 ALTER ANY MASK 和表的 ALTER 权限。

Labels come from Microsoft Purview Information Protection, so admins publish them there first. A user with edit rights picks a label on the item. Fabric then helps keep it consistent: downstream inheritance, inheritance when you create an item from another, and default or mandatory labeling.

标签来自 Microsoft Purview 信息保护，管理员先在 Purview 发布。有编辑权限的用户给项目打标签后，Fabric 会通过下游继承、由项目新建时继承，以及默认或强制标签保持一致。下游继承不支持从 Power BI 项目到 Fabric 项目。

Endorsement helps people find reliable content in the OneLake catalog and in item lists. You set it in the item settings under Endorsement. If you are not authorized, Certified and Master data appear greyed out and a link explains how to request it.

背书帮助用户在 OneLake 目录和列表里找到可靠内容，在项目设置的“背书”里设置。没有授权时，“认证”和“主数据”是灰的，旁边有申请说明的链接。认证和主数据要 Fabric 管理员先在租户里启用。

Use the Purview audit log when compliance staff need to search activity without being Fabric admins. Use the Activity Events API or Get-PowerBIActivityEvent when a Fabric admin wants to export events into their own store on a schedule.

合规人员不是 Fabric 管理员时用 Purview 审核日志搜索；Fabric 管理员要定期导出就用活动事件 API 或 Get-PowerBIActivityEvent，后者会自动处理延续令牌。仓库 SQL 审核日志默认关闭，写在仓库的 OneLake 里，删仓库日志也会没。

Admins and Members create the roles. They restrict users who only have Read, such as Viewers, while Admins, Members and Contributors keep full access. Folder access flows down to subfolders, and users can traverse parents without seeing siblings.

A ReadWrite role cannot carry row or column rules.

管理员和成员建角色。角色限制只有读取权限的用户，比如查看者；管理员、成员、参与者仍有完全访问。文件夹权限向下继承，用户能穿过父文件夹但看不到兄弟文件夹。多个角色取并集，行规则用 OR 合并。读写角色不能带行列规则。

Pick the tool by who builds it and what the job needs. A pipeline orchestrates: it copies data at scale with no code and wraps other work in control flow such as ForEach, If condition and Until. Dataflow Gen2 is low code Power Query with more than 300 transformations. A notebook is code first Spark for complex logic and open source libraries.

The three usually work together: a pipeline runs a copy, then a dataflow or a notebook, in one schedule.

按“谁来做、要做什么”选工具。管道负责编排：不写代码就能大规模复制数据，还能用 ForEach、If 条件、Until 把其他工作串起来。Dataflow Gen2 是低代码的 Power Query，有 300 多种转换。笔记本是写代码的 Spark，适合复杂逻辑和开源库。三者常常组合使用：一个管道先复制，再调数据流或笔记本。

A pipeline runs in three ways: on demand, on a schedule or from an event. A schedule needs a start and an end date, a pipeline can hold several schedules, and each can pass parameter values. Event triggers are built on eventstreams and Activator, so OneLake, Azure Blob, job and workspace events can start a run. Copy jobs, dataflows and notebooks have their own schedules too.

Inside the triggered pipeline, @pipeline()?.TriggerEvent?.FileName returns the file that fired the run.

管道有三种启动方式：手动、按计划、由事件触发。计划必须设开始和结束日期，一个管道可以有多个计划，每个计划都能传参数值。事件触发建立在事件流和 Activator 之上，OneLake、Blob、作业和工作区项事件都能启动运行。复制作业、数据流和笔记本也有自己的计划。被触发的管道里可以用 @pipeline()?.TriggerEvent?.FileName 拿到触发它的文件名。

Parameters carry values into a run and stay fixed until it ends. Variables can change during the run through the Set variable activity. Reference values with expressions such as @pipeline().parameters.Region, or embed them in text with @{...}, which always returns a string. A notebook receives values through a cell marked as the parameter cell, and notebookutils lets one notebook run others.

In a pipeline, the Notebook activity sends base parameters that replace the defaults in the parameter cell.

参数把值带进一次运行，直到结束都不变；变量可以在运行中用“设置变量”活动修改。用 @pipeline().parameters.Region 这样的表达式引用参数，或用 @{...} 嵌进文本，结果总是字符串。笔记本通过标记为参数单元格的那一格接收值，notebookutils 让一个笔记本去运行其他笔记本。在管道里，笔记本活动的基本参数会替换参数单元格里的默认值。

A full load copies everything each run, which is simple but slow and costly on big tables. An incremental load copies only new or changed rows after a first full run. A watermark tracks the highest value of a column such as a modified date. Change data capture (CDC) reads inserts, updates and deletes recorded by the source. Copy job supports both and remembers its own state.

In Copy job, pick an update method too: append is the default, merge updates rows by key, and SCD Type 2 keeps history.

全量加载每次复制全部数据，简单但大表又慢又贵。增量加载在第一次全量之后只复制新增或变更的行。水位记录某列（如修改时间）的最大值；CDC 读取源端记录的增删改。复制作业两种都支持，并自动记住状态。复制作业还要选写入方式：默认追加，合并按键更新，SCD 类型 2 保留历史。

Load staging tables first, then dimensions, then facts. Dimensions go first so every member exists before facts look up keys. Give each dimension an integer surrogate key and keep the business key from the source for matching. SCD type 1 overwrites an attribute. SCD type 2 expires the current row and inserts a new version, so history survives. A fact whose lookup fails still loads, pointing to an Unknown member.

先加载暂存表，再处理维度，最后处理事实。维度要先处理，保证事实查键时成员都已存在。每个维度用整数代理键，同时保留源系统的业务键用于匹配。SCD 1 直接覆盖属性；SCD 2 让当前行过期并插入新版本，历史得以保留。查不到维度的事实行照样加载，指向“未知”成员。

Streaming data arrives all the time, so the load never ends. An eventstream is the no code path: add a source such as Azure Event Hubs, apply transformations like Filter or Group by, and route to an eventhouse, a lakehouse, Activator or several destinations at once. Spark Structured Streaming is the code path that writes to a Delta table with a checkpoint.

Set a trigger interval on the stream to write fewer, larger files.

流数据源源不断，加载永不结束。事件流是无代码路线：加一个源（如 Azure 事件中心），做筛选、分组等转换，再路由到事件库、湖仓、Activator，或同时写多个目标。Spark 结构化流是代码路线，带检查点写入 Delta 表。给流设置触发间隔，可以写出更少、更大的文件。

All Fabric data stores keep their data in OneLake, so the choice is about workload and skills. A lakehouse suits Spark users and data of any shape. A warehouse suits T-SQL teams that need multi-table transactions. An eventhouse suits streaming, time series and log data queried with KQL. A SQL database in Fabric suits operational apps that need OLTP.

The eventhouse is the one exception for open format: its OneLake availability is opt in.

Fabric 的所有数据存储都把数据放在 OneLake，所以选择看工作负载和技能。湖仓适合 Spark 用户和各种格式的数据；仓库适合需要多表事务的 T-SQL 团队；事件库适合用 KQL 查询的流、时序和日志数据；Fabric SQL 数据库适合需要 OLTP 的业务应用。事件库是例外：它在 OneLake 中的开放格式副本需要手动开启。

Choose the transformation language that matches where the data lives and who writes the logic. Dataflow Gen2 gives analysts low code Power Query. Notebooks give engineers Spark for large volumes and complex code. T-SQL transforms tables inside a warehouse with set based statements. KQL shapes event data in an eventhouse, for example with an update policy as rows arrive.

T-SQL in a warehouse can read lakehouse and mirrored tables in the same workspace with three-part names.

按数据在哪、谁写逻辑来选转换语言。Dataflow Gen2 给分析师低代码的 Power Query；笔记本给工程师 Spark，处理大数据量和复杂代码；T-SQL 在仓库里用集合式语句转换表；KQL 在事件库里整形事件数据，比如用更新策略在数据到达时转换。仓库里的 T-SQL 可以用三段式名称读取同一工作区的湖仓和镜像表。

A shortcut is a pointer, like a symbolic link, that makes data stored elsewhere appear as a folder in OneLake. Internal shortcuts point to Fabric items. External shortcuts point to storage such as ADLS Gen2, Amazon S3 or Google Cloud Storage. Same tenant shortcuts pass the user's identity by default, while external ones use a connection.

In a KQL database a shortcut behaves like an external table and is queried with external_table().

快捷方式是一个指针，类似符号链接，让存在别处的数据在 OneLake 里显示成文件夹。内部快捷方式指向 Fabric 项，外部快捷方式指向 ADLS Gen2、Amazon S3、Google Cloud Storage 等存储。同租户快捷方式默认传递用户身份，外部快捷方式使用连接。在 KQL 数据库里，快捷方式相当于外部表，用 external_table() 查询。

Mirroring continuously replicates data from operational sources into OneLake as Delta tables, with no pipelines to build. Each mirrored database comes with a SQL analytics endpoint, so you can query it with T-SQL, join it to warehouses with three-part names, or use it in Spark and Direct Lake reports. Background replication compute is free and does not use capacity.

镜像把业务系统的数据持续复制到 OneLake，成为 Delta 表，不用自己建管道。每个镜像数据库自带 SQL 分析端点，可以用 T-SQL 查询，用三段式名称和仓库联接，也能给 Spark 和 Direct Lake 报表用。后台复制计算免费，不占容量。

Choose the tool by what happens around the copy. A Copy activity lives inside a pipeline, so it can run after a Lookup, before a notebook, or inside a loop. A Copy job stands alone, keeps its own schedules and tracks incremental progress, resuming from the end of the last successful run after a failure.

For an on-premises source, install the gateway inside your network, create an on-premises connection that points to the gateway cluster, and pick that connection in the Copy assistant or the Copy job wizard.

复制前后还要做别的事，就把复制活动放进管道，与查找、笔记本、循环串起来；只需要搬数据，就用复制作业，它自带计划，能自动记录增量进度，失败后从上次成功处继续。访问本地数据要先在内网装网关，再在“管理连接和网关”里建本地连接，然后在复制助手或复制作业里选它。

Every Fabric store has a native transformation language. In a lakehouse, a notebook reads files or tables into a DataFrame, reshapes them with PySpark or Spark SQL, and writes Delta tables. In a warehouse, T-SQL statements such as CTAS and INSERT...SELECT build new tables, and they can read lakehouse tables with three-part names. In an eventhouse, KQL reshapes events as they arrive.

每种 Fabric 存储都有自己的转换语言。湖仓里用笔记本把文件或表读成 DataFrame，用 PySpark 或 Spark SQL 处理后写成 Delta 表；仓库里用 CTAS、INSERT...SELECT 等 T-SQL 建新表，还能用三段式名称读取湖仓表；事件屋里用 KQL 在数据到达时重塑事件。湖仓的 SQL 分析终结点只能读，不能写。

Denormalizing trades some storage for simpler and faster queries. Join a chain of normalized tables, such as product, subcategory and category, into one dimension table, and keep keys and measures in the fact table. Reports then filter and group on dimension columns and summarize fact columns. Keep a snowflake only when a dimension is huge or facts relate to it at a higher grain.

反规范化用一点存储换来更简单、更快的查询：把产品、子类别、类别这类规范化链条连接成一张维度表，事实表只留键和度量。报表在维度列上筛选分组，在事实列上汇总。只有维度特别大，或事实表在更粗粒度上关联时才考虑雪花结构。

All three engines follow one idea: choose grouping keys, then apply aggregate functions to each group. Filter rows before grouping to cut work, and filter groups afterwards with HAVING in SQL or with a where after summarize in KQL. In T-SQL all null values in a grouping column form a single group, and most aggregate functions ignore nulls.

三种引擎思路一样：先选分组键，再对每组套聚合函数。分组前先过滤行能少算很多；分组后要过滤，SQL 用 HAVING，KQL 在 summarize 后面接 where。T-SQL 中分组列的所有空值归为一组，大多数聚合函数会忽略空值。

Clean incoming data in a fixed order. First remove duplicates inside the batch, because an insert-only merge still inserts duplicates that exist within the new data. Next handle missing values: drop the row, fill a default, or point a failed dimension lookup to an Unknown member. Then upsert with MERGE. When a fact arrives before its dimension row, insert an inferred member.

按固定顺序清洗：先在批次内去重，因为只插入的 MERGE 遇到新数据内部的重复仍会插进去；再处理缺失值，删行、填默认值，或让查找失败的维度键指向“未知”成员；然后用 MERGE 更新插入。事实先到、维度后到时，先插一个推断成员，等详情到了再更新。

Start from the team's skills and where the data must land. Eventstream suits low-code teams that filter, aggregate and route events to several destinations at once. Spark structured streaming suits engineers who write PySpark and want durable Delta tables in a lakehouse. KQL fits when events already land in an eventhouse and must be reshaped or summarized there.

先看团队技能和数据最终落在哪。事件流适合低代码团队，一边筛选聚合一边把事件同时发往多个目标；Spark 结构化流适合会写 PySpark、要在湖仓里落持久 Delta 表的工程师；事件已经进了事件屋、要在里面重塑或汇总时，用 KQL。

A native table is the default home for streaming data. Eventstream or another connector ingests events, the eventhouse organizes them by arrival time, and every KQL feature applies. A OneLake shortcut leaves the data in a lakehouse, warehouse or external storage and queries it as an external table. Choose a shortcut when another item owns the data and copying it is not worth it.

原生表是流数据的默认归宿：事件流等连接器把事件摄取进来，事件屋按到达时间组织数据，所有 KQL 功能都能用。OneLake 快捷方式把数据留在湖仓、仓库或外部存储，以外部表方式查询。数据归别的项目管、又不值得复制时，用快捷方式。反过来，打开 OneLake 可用性能让其他引擎读原生表。

Query acceleration closes most of the speed gap between a shortcut and an ingested table. The eventhouse caches and indexes data as it lands in OneLake, so new files are picked up without an ingestion pipeline or a duplicate copy to manage. It fits history joined with live streams, or small dimension data mirrored into OneLake. Keep a shortcut standard when it is queried rarely.

查询加速把快捷方式和摄取表之间的速度差距补上大半：事件屋在数据落到 OneLake 时就缓存并建索引，新文件自动跟上，不用维护摄取管道或重复副本。适合历史数据与实时流联查，或镜像进 OneLake 的小维度数据。很少查询的快捷方式保持标准即可。

Build an eventstream from left to right. Add one or more sources, which flow into the default stream. Insert operators to clean or reshape events, then attach destinations. Several destinations can read the same stream without interfering with each other. When other items should reuse a transformed stream, publish it as a derived stream so it appears in Real-Time hub.

事件流从左往右搭：先加一个或多个源，汇入默认流；中间插入运算符清洗或重塑事件；最后接目标。多个目标可以读同一个流，互不干扰。转换后的流要给别的项目复用时，发布成派生流，它会出现在实时中心。

A structured streaming query treats its input as a table that keeps growing. You describe the source, the transformations and the sink, and Spark reruns the plan as new data arrives. The checkpoint makes the query restartable, so Spark resumes from the last committed progress. With a replayable source and a Delta sink, results are written exactly once.

结构化流把输入看成一张不断增长的表：你描述源、转换和输出，Spark 在新数据到来时反复执行。检查点让查询可以重启，从上次提交的进度继续。源可重放、输出是 Delta 时，结果恰好写入一次。按计划跑增量用 availableNow，常驻运行用 processingTime。

Use KQL processing when events already land in an eventhouse. A raw table receives everything, and an update policy parses, filters or enriches each new batch into a clean table. On top of that, a materialized view keeps aggregates such as the latest record per device or daily counts. Query the view by name for full freshness, or with materialized_view() for speed.

事件已经进了事件屋时就用 KQL 处理：原始表接收所有数据，更新策略把每批新数据解析、过滤或丰富后写进清洗表；再在上面建物化视图，维护每台设备最新记录、每日计数等聚合。按名称查视图结果最新，用 materialized_view() 查更快。

Pick the window by how results should be emitted. Eventstream uses the Stream Analytics query language, so its SQL operator supports all five types, and the Group by operator sets a window in its settings. In KQL, summarize by bin(Timestamp, 5m) builds tumbling style buckets. In Spark, groupBy(window(col, "5 minutes")) does the same, a slide duration makes windows overlap, and session_window() groups bursts.

按结果怎么输出来选窗口。事件流使用 Stream Analytics 查询语言，SQL 运算符支持全部五种窗口，分组运算符在设置里选窗口。KQL 里 summarize by bin(Timestamp, 5m) 形成翻转式的时间桶；Spark 里 groupBy(window(col, "5 minutes")) 效果相同，加上滑动间隔就会重叠，session_window() 则按活动段分组。

Start broad, then drill down. The Monitor hub lists job runs across workspaces, with filters for status, item type and time. Opening a pipeline run shows its activities, inputs, outputs and errors. For a copy activity, the performance view splits the duration into phases, which tells you whether the source, the sink or queuing is slow.

Export activity runs to CSV when you need to share or analyze them outside Fabric.

先在监控中心按状态、类型、时间筛选，再打开某次管道运行查看每个活动的输入、输出和错误。复制活动的性能视图会把耗时拆开，帮你判断慢在源端、目标端还是排队。甘特图适合比较多次运行的时长和重叠，活动运行可以导出 CSV。

Spark work shows up as applications. Open one from the Monitor hub or from the item's Recent runs pane to reach the application detail page. Dataflow Gen2 has its own refresh history: each refresh lists the tables it loaded and the activities it ran, with rows and bytes read and written. Pick the view that answers your question.

Spark 作业以应用的形式出现，可从监控中心或项目的“最近运行”进入应用详情页。数据流 Gen2 有自己的刷新历史，每次刷新都列出加载的表和执行的活动，以及读写的行数和字节数。按你要回答的问题选对视图。

An Import refresh copies data, so it can be long and expensive. A Direct Lake refresh, called framing, reads Delta metadata and points the model at the newest Parquet files, so it usually takes seconds. Both report into refresh history, and both consume capacity as background operations you can review in the Capacity Metrics app.

导入刷新会复制全部数据，耗时长、成本高；Direct Lake 刷新叫框架化，只读 Delta 元数据并指向最新文件，通常几秒完成。两者都写入刷新历史，也都作为后台操作消耗容量，可以在容量指标应用里查看。

Fabric offers several alert paths. Schedule failure emails are free and simple, but cover only scheduled runs. Activator rules react to job events such as Started, Succeeded or Failed and can post to Teams or run another item. Inside a pipeline, an Outlook or Teams activity on a failure path sends a custom message with run details.

计划失败邮件免费、简单，但只覆盖计划运行；Activator 规则能对启动、成功、失败等作业事件告警，还能发 Teams 消息或运行其他项目。在管道内部，把 Outlook 或 Teams 活动接在失败分支上，可以发带运行详情的自定义消息。

Start with the failed activity in the run detail view. Its error message and input usually tell you whether the fault is transient, such as a network blip, or an authoring problem, such as a wrong parameter type. Transient faults call for retries; authoring problems call for a fix and a rerun from the failed activity.

先在运行详情里看失败活动的错误和输入，判断是网络抖动这类瞬时故障，还是参数类型写错这类设计问题。瞬时故障靠重试，设计问题要修正后从失败的活动重跑。

A Dataflow Gen2 problem shows up in one of two places. A publish problem appears as an indicator next to the dataflow name after you publish your draft. A refresh problem appears in refresh history, where each failed table or activity carries its own error details. Some failures have well known causes and fixes.

数据流 Gen2 的问题出现在两个地方：发布草稿后，发布错误会以名称旁的提示出现；刷新错误出现在刷新历史中，每个失败的表或活动都有自己的错误详情。上面列出几类常见失败及其处理方法。

Notebook failures fall into layers. Session and access errors stop you before code runs. Capacity errors such as HTTP 430 mean the capacity has no room for another Spark job. Code errors surface in the cell output and the Spark UI, where the Executors and Stages tabs reveal memory pressure and data skew.

笔记本故障分层：会话和权限错误让代码根本跑不起来；HTTP 430 这类容量错误说明容量已经放不下新的 Spark 作业；代码错误出现在单元格输出和 Spark UI 里，执行器页和阶段页能看出内存压力和数据倾斜。

Most Eventhouse ingestion problems are found with management commands. The .show ingestion failures command lists failures from ingestion commands with the table, details and whether the failure is permanent or transient. Update policies add a second place to fail, because their query runs on every new batch and must return the target table's exact schema.

For trends over time, workspace monitoring stores ingestion results in the EventhouseIngestionResultsLogs table.

Eventhouse 摄取问题大多用管理命令排查。.show ingestion failures 列出摄取命令的失败，含表名、详情和永久或瞬时类型。更新策略多了一个失败点：它对每批新数据运行查询，输出必须与目标表架构完全一致。要看长期趋势，可用工作区监控的摄取结果日志表。

Select a node on the editor canvas, then read the lower pane. Data insights charts show whether events arrive, back up or fail at that node. Runtime logs list the engine's messages by severity. Schema problems show up as conversion errors at a destination, so compare the inferred schema with the target table.

在编辑器画布上选中节点，再看下方窗格。数据洞察图表显示事件在该节点是否到达、积压或失败；运行时日志按级别列出引擎消息。架构问题通常在目标处表现为数据转换错误，要把推断的架构和目标表对比。

Fabric Data Warehouse speaks T-SQL, but not every SQL Server statement. When a script fails, first check the surface area list, then the transaction rules. Locks are taken per table, so concurrent updates and long DDL can block or abort each other. For history, the queryinsights views keep completed queries for 30 days.

Fabric 数据仓库使用 T-SQL，但不支持 SQL Server 的全部语句。脚本失败时先查表面区域列表，再看事务规则。锁按表获取，并发更新和长时间 DDL 会互相阻塞或中止。查历史用 queryinsights 视图，已完成查询保留 30 天。

A shortcut behaves like a symbolic link, so it depends on three things outside itself: the target path must still exist, the user needs permission on both the shortcut path and the target, and external shortcuts need a working cloud connection. Check them in that order.

快捷方式类似符号链接，依赖三件事：目标路径仍然存在，用户对快捷方式路径和目标都有权限，外部快捷方式的云连接可用。按这个顺序检查。

Streaming and frequent small writes leave a Delta table with many small files, and every reader pays for them. OPTIMIZE rewrites them into right sized files. Add ZORDER BY for columns you often filter on together, and VORDER when Power BI or other read heavy consumers matter. New workspaces write without V-Order by default.

When you combine them, Spark compacts first, then applies Z-Order, then V-Order. Run VACUUM after compaction to drop the replaced files. It is a Spark command, so it fails in the SQL analytics endpoint.

流式写入和频繁的小批写入会让 Delta 表堆积大量小文件，拖慢所有读取。OPTIMIZE 把它们合并成大小合适的文件；经常一起筛选的列加 ZORDER BY，给 Power BI 等读多场景用时加 VORDER。新工作区默认写入时不开 V-Order。两者一起用时顺序是：先装箱压缩，再 Z-Order，最后 V-Order。压缩后运行 VACUUM 删掉被替换的旧文件；它是 Spark 命令，在 SQL 分析终结点里跑不了。

Copy speed is capped by the slowest of the source, the destination and the network between them. Scale at two levels. Inside one Copy activity, raise intelligent throughput optimization and let the degree of copy parallelism add threads. Across many tables, run copies at the same time from a ForEach with Sequential cleared.

With a partition option set, the copy runs as many partition queries in parallel as the degree of copy parallelism allows.

复制速度取决于源、目标和两者之间网络中最慢的一环。可以在两个层面扩展：单个复制活动内，提高智能吞吐优化，让复制并行度增加线程；多张表时，用 ForEach 并取消勾选 Sequential，让多个复制同时跑。设置了分区选项后，并行运行的分区查询数由复制并行度决定。

The warehouse compacts files and creates statistics for you, but your design still matters. Use the smallest data types that fit, keep join columns on the same type, and load in batches with COPY INTO instead of many single row inserts. When a load changes data a lot, update statistics before users arrive so their first queries do not pay for the refresh.

仓库会自动压缩文件、创建统计信息，但设计仍然重要：用够用的最小数据类型，连接列类型保持一致，用 COPY INTO 批量加载而不是逐行插入。大批数据变化后，在用户来之前先更新统计信息，免得早上的第一批查询替你等统计刷新。

Queued ingestion waits to build bigger batches, so data shows up later but costs less. If data arrives too late and volume is low, shorten the batching time on that table. Keep the data people query most in the hot cache, and let the retention policy drop old data automatically. Data that is outside the hot cache but inside retention stays queryable, only slower.

排队引入会先攒成大批次再写入，数据出现得晚一点但更省钱。数据量小又嫌延迟高时，缩短该表的批处理时间。常查的数据放进热缓存，旧数据交给保留策略自动删除；在保留期内但不在热缓存的数据照样能查，只是慢一些。

Start with the engine. The native execution engine runs supported Spark operators in vectorized C++ and falls back to the JVM for anything it cannot handle, such as JSON or XML reads. Turn it on in the environment Acceleration tab, or per session with spark.native.enabled. Then match the resource profile to the workload.

For a slow stage, look for skew in the Spark UI, keep adaptive query execution on, and tune shuffle partitions.

先从引擎入手：原生执行引擎用向量化 C++ 执行支持的算子，遇到不支持的（比如读 JSON、XML）自动回退到 JVM。在环境的 Acceleration 页打开，或在会话里设 spark.native.enabled。然后按负载选资源配置文件。某个阶段慢时，在 Spark UI 里查倾斜，保持自适应查询执行开启，再调整洗牌分区数。

Every engine runs faster when it touches less data. In T-SQL against a warehouse or SQL analytics endpoint, name the columns you need, filter before joins, and keep join columns on the same data type to avoid conversions. Lakehouse tables read through the endpoint also need healthy files, so compact them and avoid high cardinality partitions.

In KQL, put datetime filters first, then term filters on strings, then numeric filters, and leave column scans such as contains for last.

任何引擎都是处理的数据越少越快。对仓库或 SQL 分析终结点写 T-SQL 时，只列需要的列，先筛选再连接，连接列类型保持一致以免转换。通过终结点读取的湖仓表也要文件健康：定期压缩，避免高基数分区。KQL 里先放时间筛选，再放字符串整词筛选，然后数值筛选，contains 这类扫描放最后。