sql - 突触顶级父层次结构

Question

这是表结构；

code    name                      under 

1       National Sales Manager    1
2       regional sales manager    1 
3       area sales manager        2 
4       sales manager             3

如何获得如下所示的顶级父层次结构；

code    name                      under     ultimateparent

1       National Sales Manager    1         1
2       regional sales manager    1         1
3       area sales manager        2         1
4       sales manager             3         1

在常规 SQL 服务器上，我将使用递归 CTE，如 SQL Server function to get top level parent in hierarchy 中所示。 。但是，突触数据库不支持它。

Answer 1

您有几个选择。我将在这里描述三个:

1. 如果您附近有 Azure SQL DB 并且卷不太大，请使用 CREATE EXTERNAL TABLE 呈现该 Azure SQL DB 中的表，或者仅使用 Azure 数据工厂 (ADF) 进行复制数据结束后，执行递归 CTE，然后使用 ADF 将其移植回来。或者在这些数据进入 SQL 池之前使用某种预处理。
2. 递归 CTE 只是一天结束时的一种循环，因此 Synapse 支持 WHILE。现在显然这种类型的循环不能很好地转换为 Synapse，因为它很啰嗦，但对于层次结构深度较低的小体积来说可能是一个选择。您需要权衡以这种方式无效地使用 MPP 架构与编写替代方案之间的权衡。

我编写了选项 2 的示例，仅运行几行就花费了 20 多秒。通常我会认为这是 Not Acceptable ，但正如前面提到的，您可以权衡替代方案:

IF OBJECT_ID('dbo.someHierarchy') IS NOT NULL
    DROP TABLE dbo.someHierarchy;

CREATE TABLE dbo.someHierarchy (
    code        INT NOT NULL,
    [name]      VARCHAR(50) NOT NULL,
    under       INT NOT NULL
)
WITH
    (
    DISTRIBUTION = ROUND_ROBIN,
    HEAP
    );


INSERT INTO dbo.someHierarchy ( code, [name], under )
SELECT 1, 'National Sales Manager', 1
UNION ALL
SELECT 2, 'Regional Sales Manager', 1
UNION ALL
SELECT 3, 'Area Sales Manager', 2
UNION ALL
SELECT 4, 'Sales Manager', 3

INSERT INTO dbo.someHierarchy ( code, [name], under )
SELECT 5, 'Lead Bob', 5
UNION ALL
SELECT 6, 'Main Bob', 5
UNION ALL
SELECT 7, 'Junior Bob 1', 6
UNION ALL
SELECT 8, 'Junior Bob 2', 6

INSERT INTO dbo.someHierarchy ( code, [name], under )
SELECT 9, 'Jim - CEO', 9
UNION ALL
SELECT 10, 'Tim - CFO', 9
UNION ALL
SELECT 11, 'Rob - CIO', 9
UNION ALL
SELECT 12, 'Bob - VP', 10
UNION ALL
SELECT 13, 'Shon - Director', 12
UNION ALL
SELECT 14, 'Shane - VP', 11
UNION ALL
SELECT 15, 'Sheryl - VP', 11
UNION ALL
SELECT 16, 'Dan - Director', 15
UNION ALL
SELECT 17, 'Kim - Director', 15
UNION ALL
SELECT 18, 'Carlo - PM', 16
UNION ALL
SELECT 19, 'Monty - Sr Dev', 18
UNION ALL
SELECT 20, 'Chris - Sr Dev', 18




IF OBJECT_ID('tempdb..#tmp') IS NOT NULL DROP TABLE #tmp;

CREATE TABLE #tmp (
    xlevel          INT NOT NULL,
    code            INT NOT NULL,
    [name]          VARCHAR(50) NOT NULL,
    under           INT NOT NULL,
    ultimateParent  INT NOT NULL
    );
    

-- Insert first level; similar to anchor section of CTE
INSERT INTO #tmp ( xlevel, code, [name], under, ultimateParent )
SELECT 1 AS xlevel, code, [name], under, under AS ultimateParent
FROM dbo.someHierarchy
WHERE under = code;


-- Loop section
DECLARE @i INT = 1

WHILE EXISTS (
    SELECT * FROM dbo.someHierarchy h
    WHERE NOT EXISTS ( SELECT * FROM #tmp t WHERE h.code = t.code )
    )
BEGIN

    -- Insert subsequent levels; similar to recursive section of CTE
    INSERT INTO #tmp ( xlevel, code, [name], under, ultimateParent )
    SELECT t.xlevel + 1, h.code, h.[name], h.under, t.ultimateParent
    FROM #tmp t
        INNER JOIN dbo.someHierarchy h ON t.code = h.under
    WHERE h.under != h.code
      AND t.xlevel = @i;

    -- Increment counter
    SET @i += 1

    -- Loop guard
    IF @i > 99
    BEGIN
        RAISERROR( 'Too many loops!', 16, 1 )
        BREAK
    END
END

SELECT 'loop' s, *
FROM #tmp
ORDER BY code, xlevel;

结果:

条件是 WHILE EXISTS 循环是一种特别昂贵的方法，因此也许有一种更简单的方法来处理您的数据。

第三种替代方法是使用 Azure Synapse Notebook 和库，例如 GraphFrames走层次结构。有更简单的方法可以做到这一点，但我发现连接组件方法能够确定最终的管理器。使用 GraphFrames 的优点之一是它允许更复杂的图形查询，例如如果需要的话可以使用主题。 此笔记本使用 Spark (Scala) 版本:

上传 graphFrames library 的正确版本到 Spark:

%%configure -f
{
    "conf": {
        "spark.jars": "abfss://{yourContainer}@{yourDataLake}.dfs.core.windows.net/synapse/workspaces/{yourWorkspace}/sparkpools/{yourSparkpool}/libraries/graphframes-0.8.1-spark2.4-s_2.11.jar",
    }
}

为您的环境配置带有大括号的元素。

导入相关库:

import org.apache.spark.sql._
import org.apache.spark.sql.functions._
import org.graphframes._

从专用 SQL 池获取数据并将其分配给数据帧:

// Get a table from Synapse dedicated SQL pool, select / rename certain columns from it to vertices and edge dataframes
val df = spark.read.synapsesql("yourDB.dbo.someHierarchy")
val v = df.selectExpr("code AS id", "name AS empName", "under")

v.show

// Reformat the code/under relationship from the original table 
// NB Exclude because in graph terms these don't have an edge
val e = df.selectExpr("code AS src", "under AS dst", "'under' AS relationship").where("code != under")

e.show

从顶点和边数据帧创建图形框架:

// Create the graph frame
val g = GraphFrame(v, e)
print(g)

为connectedComponents设置检查点:

// The connected components adds a component id to each 'group'
// Set a checkpoint to start
sc.setCheckpointDir("/tmp/graphframes-azure-synapse-notebook")

针对数据运行连接组件算法:

// Run connected components algorithm against the data
val cc = g.connectedComponents.run() // doesn't work on Spark 1.4
display(cc)

加入原始顶点数据帧和连接组件算法的结果，并将其写回 Azure Synapse 专用 SQL 池:

val writeDf = spark.sqlContext.sql ("select v.id, v.empName, v.under, cc.component AS ultimateManager from v inner join cc on v.id = cc.id")

//display(writeDf)
writeDf.write.synapsesql("someDb.dbo.someHierarchy2", Constants.INTERNAL)

结果:

我有一种感觉，会有一种更简单的方法可以用笔记本来完成此任务，但期待看到一些替代方案。在此处对 Synapse 上的递归 CTE 的反馈项目进行投票:

https://feedback.azure.com/forums/307516-azure-synapse-analytics/suggestions/14876727-support-for-recursive-cte