如何在 PostgreSQL 中展开 JSON 树的所有节点？答案

【问题标题】：How to unroll all nodes of a JSON tree in PostgreSQL?如何在 PostgreSQL 中展开 JSON 树的所有节点？
【发布时间】：2019-09-04 07:34:08
【问题描述】：

让我们有一个带有 jsonb 列的表，该列表示节点的树结构。要生成所有树节点的平面列表，每行一个，在使用令人惊叹的递归 CTE 的简单树结构的情况下很容易做到：

CREATE TEMPORARY TABLE api_schema (id INT, content JSONB);
INSERT INTO api_schema VALUES (1, '[{"name": "A", "category": "tuple", "children": [{"name": "B", "category": "datapoint"}, {"name": "C", "category": "datapoint"}]}]');
INSERT INTO api_schema VALUES (2, '[{"name": "D", "category": "tuple", "children": [{"name": "E", "category": "tuple", "children": [{"name": "F", "category": "datapoint"}]}]}]');

WITH RECURSIVE schema_objects (id, object) AS (
  SELECT id, jsonb_array_elements(content) FROM api_schema
  UNION
  SELECT id, jsonb_array_elements(object->'children') FROM schema_objects
  WHERE object->>'category' != 'datapoint'
) SELECT * FROM schema_objects;

棘手的部分是递归公式中需要更多逻辑时。在我的例子中，除了datapoint（没有孩子）和tuple（孩子是一个列表）类别之外，还有一个multivalue类别（孩子是一个单一的节点）。如何让CTE处理这种情况？

CTE 的天真重写是这样的：

INSERT INTO api_schema VALUES (3, '[{"name": "D", "category": "multivalue", "children": {"name": "E", "category": "tuple", "children": [{"name": "F", "category": "datapoint"}]}}]');

WITH RECURSIVE schema_objects (id, object) AS (
  SELECT id, jsonb_array_elements(content) FROM api_schema
  UNION
  SELECT id, CASE WHEN jsonb_typeof(object->'children') = 'array'
               THEN jsonb_array_elements(object->'children')
               ELSE object->'children'
             END AS object
  FROM schema_objects
  WHERE object->>'category' != 'datapoint'
) SELECT * FROM schema_objects;

但是，这里的问题是这在 Postgres 10 中不起作用：

ERROR:  set-returning functions are not allowed in CASE

我们可以创建两个 SELECT，每个 SELECT 都涵盖一个单独的类别吗？这是不允许的：

WITH RECURSIVE schema_objects (id, object) AS (
  SELECT id, jsonb_array_elements(content) FROM api_schema
  UNION
  (SELECT id, jsonb_array_elements(object->'children') FROM schema_objects WHERE object->>'category' = 'tuple'
   UNION
   SELECT id, object->'children' FROM schema_objects WHERE object->>'category' = 'multivalue')
) SELECT * FROM schema_objects WHERE id=1;

ERROR:  recursive reference to query "schema_objects" must not appear more than once

互联网上流传的一个想法是使用 CTE 来分解 CASE，但我们已经在 CTE 中，所以这甚至无法编译：

WITH RECURSIVE schema_objects (id, object) AS (
  SELECT id, jsonb_array_elements(content) FROM api_schema
  UNION
  WITH schema_children (id, children) AS (
    SELECT CASE jsonb_typeof(object->'children') WHEN 'array' THEN object->'children' ELSE jsonb_build_array(object->'children') END AS children
    FROM schema_objects
    WHERE object->>'category' != 'datapoint'
  )
  SELECT id, jsonb_array_elements(children)
  FROM schema_children
) SELECT * FROM schema_objects WHERE id=1;

Postgres 也建议使用横向 FROM，但不清楚在单个“表”的情况下如何组合。

【问题讨论】：

标签： json postgresql recursion tree

【解决方案1】：

在撰写问题时，我在早期的尝试中发现了一个错误:)。在 FROM 中使用子选择可以解决问题，将非数组重新捆绑为数组：

WITH RECURSIVE schema_objects (id, object) AS (
  SELECT id, jsonb_array_elements(content) FROM api_schema

  UNION

  SELECT id, jsonb_array_elements(children)
  FROM (SELECT id, CASE jsonb_typeof(object->'children') WHEN 'array' THEN object->'children' ELSE jsonb_build_array(object->'children') END AS children
    FROM schema_objects
    WHERE object->>'category' != 'datapoint'
  ) s
) SELECT * FROM schema_objects

但是，这种方法的一个非常大的普遍缺点是它非常慢 - 对于 10000 ~ 30 个节点的树，我正在查看需要分钟的查询。

【讨论】：