Strange duplicate behavior from GROUP_CONCAT of two LEFT JOINs of GROUP_BYs

Question

Here is all my tables' structure and the query (please focus on the last query, appended below). As you see in the fiddle, here is the current output:

+---------+-----------+-------+------------+--------------+
| user_id | user_name | score | reputation | top_two_tags |
+---------+-----------+-------+------------+--------------+
| 1       | Jack      | 0     | 18         | css,mysql    |
| 4       | James     | 1     | 5          | html         |
| 2       | Peter     | 0     | 0          | null         |
| 3       | Ali       | 0     | 0          | null         |
+---------+-----------+-------+------------+--------------+

It's correct and all fine.

---

Now I have one more existence named "category". Each post can has only one category. And I also want to get top two categories for each user. And here is my new query. As you see in the result, some duplicates happened:

+---------+-----------+-------+------------+--------------+------------------------+
| user_id | user_name | score | reputation | top_two_tags |   top_two_categories   |
+---------+-----------+-------+------------+--------------+------------------------+
| 1       | Jack      | 0     | 18         | css,css      | technology,technology  |
| 4       | James     | 1     | 5          | html         | political              |
| 2       | Peter     | 0     | 0          | null         | null                   |
| 3       | Ali       | 0     | 0          | null         | null                   |
+---------+-----------+-------+------------+--------------+------------------------+

See? css,css, technology, technology. Why these are duplicate? I've just added one more LEFT JOIN for categories, exactly like tags. But it doesn't work as expected and even affects on the tags either.

---

Anyway, this is the expected result:

+---------+-----------+-------+------------+--------------+------------------------+
| user_id | user_name | score | reputation | top_two_tags |        category        |
+---------+-----------+-------+------------+--------------+------------------------+
| 1       | Jack      | 0     | 18         | css,mysql    | technology,social      |
| 4       | James     | 1     | 5          | html         | political              |
| 2       | Peter     | 0     | 0          | null         | null                   |
| 3       | Ali       | 0     | 0          | null         | null                   |
+---------+-----------+-------+------------+--------------+------------------------+

Does anybody know how can I achieve that?

---

CREATE TABLE users(id integer PRIMARY KEY, user_name varchar(5));
CREATE TABLE tags(id integer NOT NULL PRIMARY KEY, tag varchar(5));
CREATE TABLE reputations(
    id  integer PRIMARY KEY, 
    post_id  integer /* REFERENCES posts(id) */, 
    user_id integer REFERENCES users(id), 
    score integer, 
    reputation integer, 
    date_time integer);
CREATE TABLE post_tag(
    post_id integer /* REFERENCES posts(id) */, 
    tag_id integer REFERENCES tags(id),
    PRIMARY KEY (post_id, tag_id));
CREATE TABLE categories(id INTEGER NOT NULL PRIMARY KEY, category varchar(10) NOT NULL);
CREATE TABLE post_category(
    post_id INTEGER NOT NULL /* REFERENCES posts(id) */, 
    category_id INTEGER NOT NULL REFERENCES categories(id),
    PRIMARY KEY(post_id, category_id)) ;

SELECT
    q1.user_id, q1.user_name, q1.score, q1.reputation, 
    substring_index(group_concat(q2.tag  ORDER BY q2.tag_reputation DESC SEPARATOR ','), ',', 2) AS top_two_tags,
    substring_index(group_concat(q3.category  ORDER BY q3.category_reputation DESC SEPARATOR ','), ',', 2) AS category
FROM
    (SELECT 
        u.id AS user_Id, 
        u.user_name,
        coalesce(sum(r.score), 0) as score,
        coalesce(sum(r.reputation), 0) as reputation
    FROM 
        users u
        LEFT JOIN reputations r 
            ON    r.user_id = u.id 
              AND r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
    GROUP BY 
        u.id, u.user_name
    ) AS q1
    LEFT JOIN
    (
    SELECT
        r.user_id AS user_id, t.tag, sum(r.reputation) AS tag_reputation
    FROM
        reputations r 
        JOIN post_tag pt ON pt.post_id = r.post_id
        JOIN tags t ON t.id = pt.tag_id
    WHERE
        r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
    GROUP BY
        user_id, t.tag
    ) AS q2
    ON q2.user_id = q1.user_id 
    LEFT JOIN
    (
    SELECT
        r.user_id AS user_id, c.category, sum(r.reputation) AS category_reputation
    FROM
        reputations r 
        JOIN post_category ct ON ct.post_id = r.post_id
        JOIN categories c ON c.id = ct.category_id
    WHERE
        r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
    GROUP BY
        user_id, c.category
    ) AS q3
    ON q3.user_id = q1.user_id 
GROUP BY
    q1.user_id, q1.user_name, q1.score, q1.reputation
ORDER BY
    q1.reputation DESC, q1.score DESC ;

Answer 1

Your second query is of the form:

q1 -- PK user_id
LEFT JOIN (...
    GROUP BY user_id, t.tag
) AS q2
ON q2.user_id = q1.user_id 
LEFT JOIN (...
    GROUP BY user_id, c.category
) AS q3
ON q3.user_id = q1.user_id
GROUP BY -- group_concats

The inner GROUP BYs result in (user_id, t.tag) & (user_id, c.category) being keys/UNIQUEs. Other than that I won't address those GROUP BYs.

TL;DR When you join (q1 JOIN q2) to q3 it is not on a key/UNIQUE of one of them so for each user_id you get a row for every possible combination of tag & category. So the final GROUP BY inputs duplicates per (user_id, tag) & per (user_id, category) and inappropriately GROUP_CONCATs duplicate tags & categories per user_id. Correct would be (q1 JOIN q2 GROUP BY) JOIN (q1 JOIN q3 GROUP BY) in which all joins are on common key/UNIQUE (user_id) & there is no spurious aggregation. Although sometimes you can undo such spurious aggregation.

A correct symmetrical INNER JOIN approach: LEFT JOIN q1 & q2--1:many--then GROUP BY & GROUP_CONCAT (which is what your first query did); then separately similarly LEFT JOIN q1 & q3--1:many--then GROUP BY & GROUP_CONCAT; then INNER JOIN the two results ON user_id--1:1.

A correct symmetrical scalar subquery approach: SELECT the GROUP_CONCATs from q1 as scalar subqueries each with a GROUP BY.

A correct cumulative LEFT JOIN approach: LEFT JOIN q1 & q2--1:many--then GROUP BY & GROUP_CONCAT; then LEFT JOIN that & q3--1:many--then GROUP BY & GROUP_CONCAT.

A correct approach like your 2nd query: You first LEFT JOIN q1 & q2--1:many. Then you LEFT JOIN that & q3--many:1:many. It gives a row for every possible combination of a tag & a category that appear with a user_id. Then after you GROUP BY you GROUP_CONCAT--over duplicate (user_id, tag) pairs and duplicate (user_id, category) pairs. That is why you have duplicate list elements. But adding DISTINCT to GROUP_CONCAT gives a correct result. (Per wchiquito's comment.)

Which you prefer is as usual an engineering tradeoff to be informed by query plans & timings, per actual data/usage/statistics. input & stats for expected amount of duplication), timing of actual queries, etc. One issue is whether the extra rows of the many:1:many JOIN approach offset its saving of a GROUP BY.

-- cumulative LEFT JOIN approach
SELECT
   q1.user_id, q1.user_name, q1.score, q1.reputation,
    top_two_tags,
    substring_index(group_concat(q3.category  ORDER BY q3.category_reputation DESC SEPARATOR ','), ',', 2) AS category
FROM
    -- your 1st query (less ORDER BY) AS q1
    (SELECT
        q1.user_id, q1.user_name, q1.score, q1.reputation, 
        substring_index(group_concat(q2.tag  ORDER BY q2.tag_reputation DESC SEPARATOR ','), ',', 2) AS top_two_tags
    FROM
        (SELECT 
            u.id AS user_Id, 
            u.user_name,
            coalesce(sum(r.score), 0) as score,
            coalesce(sum(r.reputation), 0) as reputation
        FROM 
            users u
            LEFT JOIN reputations r 
                ON    r.user_id = u.id 
                  AND r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
        GROUP BY 
            u.id, u.user_name
        ) AS q1
        LEFT JOIN
        (
        SELECT
            r.user_id AS user_id, t.tag, sum(r.reputation) AS tag_reputation
        FROM
            reputations r 
            JOIN post_tag pt ON pt.post_id = r.post_id
            JOIN tags t ON t.id = pt.tag_id
        WHERE
            r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
        GROUP BY
            user_id, t.tag
        ) AS q2
        ON q2.user_id = q1.user_id 
        GROUP BY
            q1.user_id, q1.user_name, q1.score, q1.reputation
    ) AS q1
    -- finish like your 2nd query
    LEFT JOIN
    (
    SELECT
        r.user_id AS user_id, c.category, sum(r.reputation) AS category_reputation
    FROM
        reputations r 
        JOIN post_category ct ON ct.post_id = r.post_id
        JOIN categories c ON c.id = ct.category_id
    WHERE
        r.date_time > 1500584821 /* unix_timestamp(DATE_SUB(now(), INTERVAL 1 WEEK)) */
    GROUP BY
        user_id, c.category
    ) AS q3
    ON q3.user_id = q1.user_id 
GROUP BY
    q1.user_id, q1.user_name, q1.score, q1.reputation
ORDER BY
    q1.reputation DESC, q1.score DESC ;