如何使用多个 CASE 语句优化 SELECT 查询？答案

【问题标题】：How to optimize SELECT query with multiple CASE statements?如何使用多个 CASE 语句优化 SELECT 查询？
【发布时间】：2017-03-13 13:13:00
【问题描述】：

在两人游戏中，PostgreSQL based game 最常调用的语句是返回用户正在玩的游戏列表的 SELECT 查询：

（请原谅截图中的非拉丁字母）

CREATE OR REPLACE FUNCTION words_get_games(in_uid integer)
        RETURNS TABLE (
                out_gid integer, 
                out_created integer,
                out_finished integer,
                out_letters varchar[15][15],
                out_values integer[15][15],
                out_bid integer,
                out_last_tiles jsonb,
                out_last_score integer,
                out_player1 integer,
                out_player2 integer,
                out_played1 integer,
                out_played2 integer,
                out_hand1 text,
                out_hand2 text,
                out_score1 integer,
                out_score2 integer,
                out_female1 integer,
                out_female2 integer,
                out_given1 varchar,
                out_given2 varchar,
                out_photo1 varchar,
                out_photo2 varchar,
                out_place1 varchar,
                out_place2 varchar
        ) AS
$func$
    SELECT 
        g.gid,
        EXTRACT(EPOCH FROM g.created)::int,
        EXTRACT(EPOCH FROM g.finished)::int,
        g.letters,
        g.values,
        g.bid,
        m.tiles,
        m.score,
        /* HOW TO OPTIMIZE THE FOLLOWING CASE STATEMENTS? */
        CASE WHEN g.player1 = in_uid THEN g.player1 ELSE g.player2 END,
        CASE WHEN g.player1 = in_uid THEN g.player2 ELSE g.player1 END,
        CASE WHEN g.player1 = in_uid THEN g.score1 ELSE g.score2 END,
        CASE WHEN g.player1 = in_uid THEN g.score2 ELSE g.score1 END,
        CASE WHEN g.player1 = in_uid THEN s1.female ELSE s2.female END,
        CASE WHEN g.player1 = in_uid THEN s2.female ELSE s1.female END,
        CASE WHEN g.player1 = in_uid THEN s1.given ELSE s2.given END,
        CASE WHEN g.player1 = in_uid THEN s2.given ELSE s1.given END,
        CASE WHEN g.player1 = in_uid THEN s1.photo ELSE s2.photo END,
        CASE WHEN g.player1 = in_uid THEN s2.photo ELSE s1.photo END,
        CASE WHEN g.player1 = in_uid THEN s1.place ELSE s2.place END,
        CASE WHEN g.player1 = in_uid THEN s2.place ELSE s1.place END,
        EXTRACT(EPOCH FROM CASE WHEN g.player1 = in_uid THEN g.played1 ELSE g.played2 END)::int,
        EXTRACT(EPOCH FROM CASE WHEN g.player1 = in_uid THEN g.played2 ELSE g.played1 END)::int,
        ARRAY_TO_STRING(CASE WHEN g.player1 = in_uid THEN g.hand1 ELSE g.hand2 END, ''),
        REGEXP_REPLACE(ARRAY_TO_STRING(CASE WHEN g.player1 = in_uid THEN g.hand2 ELSE g.hand1 END, ''), '.', '?', 'g'),
    FROM words_games g 
        LEFT JOIN words_moves m ON m.gid = g.gid
        -- find move record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
            FROM words_moves m2 
            WHERE m2.gid = m.gid
            AND m2.played > m.played)
    LEFT JOIN words_social s1 ON s1.uid = g.player1
        -- find social record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
            FROM words_social s 
            WHERE s1.uid = s.uid
            AND s.stamp > s1.stamp)
    LEFT JOIN words_social s2 ON s2.uid = g.player2
        -- find social record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
            FROM words_social s 
            WHERE s2.uid = s.uid
            AND s.stamp > s2.stamp)
    WHERE in_uid IN (g.player1, g.player2)
    AND (g.finished IS NULL OR g.finished > CURRENT_TIMESTAMP - INTERVAL '1 day');

$func$ LANGUAGE sql;

正如您在上面的自定义 SQL 函数中所见，为了始终将用户数据返回为 player1、given1、score1，我使用了许多 CASE 语句（以便在需要时可以交换获取的列):

    CASE WHEN g.player1 = in_uid THEN g.score1 ELSE g.score2 END,

我的问题是：是否可以优化上述 SELECT 查询（无需切换到较慢的 PL/pgSQL）？

更新：

the mailing list 的 Geoff 提供了一个很好的建议，即在 JOINing 时已经使用 CASE：

SELECT 
        g.gid,
        EXTRACT(EPOCH FROM g.created)::int,
        EXTRACT(EPOCH FROM g.finished)::int,
        g.letters,
        g.values,
        g.bid,
        m.tiles,
        m.score,
        CASE WHEN g.player1 = in_uid THEN g.player1 ELSE g.player2 END,
        CASE WHEN g.player1 = in_uid THEN g.player2 ELSE g.player1 END,
        CASE WHEN g.player1 = in_uid THEN g.score1 ELSE g.score2 END,
        CASE WHEN g.player1 = in_uid THEN g.score2 ELSE g.score1 END,
        s1.female,
        s2.female,
        s1.given,
        s2.given,
        s1.photo,
        s2.photo,
        s1.place,
        s2.place,
        EXTRACT(EPOCH FROM CASE WHEN g.player1 = in_uid THEN g.played1 ELSE g.played2 END)::int,
        EXTRACT(EPOCH FROM CASE WHEN g.player1 = in_uid THEN g.played2 ELSE g.played1 END)::int,
        ARRAY_TO_STRING(CASE WHEN g.player1 = in_uid THEN g.hand1 ELSE g.hand2 END, ''),
        REGEXP_REPLACE(ARRAY_TO_STRING(CASE WHEN g.player1 = in_uid THEN g.hand2 ELSE g.hand1 END, ''), '.', '?', 'g')
FROM words_games g 
LEFT JOIN words_moves m ON m.gid = g.gid
        -- find move record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
                FROM words_moves m2 
                WHERE m2.gid = m.gid
                AND m2.played > m.played)
LEFT JOIN words_social s1 ON s1.uid = in_uid
        -- find social record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
                FROM words_social s 
                WHERE s1.uid = s.uid
                AND s.stamp > s1.stamp)
LEFT JOIN words_social s2 ON s2.uid = (CASE WHEN g.player1 = in_uid THEN g.player2 ELSE g.player1 END)
        -- find social record with the most recent timestamp
        AND NOT EXISTS (SELECT 1
                FROM words_social s 
                WHERE s2.uid = s.uid
                AND s.stamp > s2.stamp)
WHERE in_uid IN (g.player1, g.player2)
AND (g.finished IS NULL OR g.finished > CURRENT_TIMESTAMP - INTERVAL '1 day');

【问题讨论】：

你怎么知道它是次优的？只有两组值可能（或可能不）需要在外部查询中交换，在查询已经检索到所需的数据（一个用于游戏，另外两个用于玩家)
是的，我不确定。实际上我想优化两点：性能和可读性。
我不会担心 case 表达式，因为它们在选择列表中。
我同意这很丑 ;-] 另一种方法是保留 {player1,player2} 符号，或者在棋盘游戏 {player_black, player_white} 或球类游戏 {team_home, team_visitors} 和/或者将交换留给表示层。
为避免出现多种情况，您可以使用Composite Types：CASE WHEN g.player1 = in_uid THEN (g.player1, g.score1,...)::tp_game_data ELSE (g.player2, g.score2,...)::tp_game_data END

标签： sql postgresql select case postgresql-9.5

【解决方案1】：

lateral 和 distinct on (IMO) 有助于提高可读性。 distinct on 也会对性能产生影响，尽管我无法猜测是正面还是负面。

select
    g.gid,
    extract(epoch from g.created)::int created,
    extract(epoch from g.finished)::int finished,
    g.letters,
    g.values,
    g.bid,
    m.tiles,
    m.score,
    r.*
from
    words_games g
    left join (
        select distinct on (gid, played) *
        from words_moves
        order by gid, played desc
    ) words_moves m on m.gid = g.gid
    left join (
        select distinct on (uid, stamp) *
        from words_social
        order by uid, stamp desc
    ) words_social s1 on s1.uid = g.player1
    left join (
        select distinct on (uid, stamp) *
        from words_social
        order by uid, stamp desc
    ) words_social s2 on s2.uid = g.player2
    cross join lateral (
        select
            g.player1, g.player2,
            extract(epoch from g.player1)::int, extract(epoch from g.player2)::int,
            array_to_string(g.hand1, ''),
            regexp_replace(array_to_string(g.hand2, ''), '.', '?', 'g'),
            g.score1, g.score2,
            s1.female, s2.female,
            s1.given, s2.given,
            s1.photo, s2.photo,
            s1.place, s2.place
        where g.player1 = in_uid
        union all
        select
            g.player2, g.player1,
            extract(epoch from g.player2)::int, extract(epoch from g.player1)::int,
            array_to_string(g.hand2, ''),
            regexp_replace(array_to_string(g.hand1, ''), '.', '?', 'g'),
            g.score2, g.score1,
            s2.female, s1.female,
            s2.given, s1.given,
            s2.photo, s1.photo,
            s2.place, s1.place
        where g.player1 != in_uid
    ) r
where
    in_uid in (g.player1, g.player2)
    and (g.finished is null or g.finished > current_timestamp - interval '1 day')

【讨论】：

【解决方案2】：

由于您担心许多 case 语句并且总是相同的条件，您可以将这个条件拉出并有两个选择，例如

select ...
       g.player1, g.player2,
       extract(epoch from g.played1)::int, extract(epoch from g.played2)::int,
       ...
       g.score1, g.score2,
       ...

和另一个（相同的）选择与交换的列

select ...
       g.player2, g.player1,
       extract(epoch from g.played2)::int, extract(epoch from g.played1)::int,
       ...
       g.score2, g.score1,
       ...

虽然，正如@joop 和@jarlh 已经质疑的那样，首先测试一下，这是否真的是一个性能问题。

【讨论】：

注意：在之前的帖子中，OP 有一个类似（相当庞大）的解决方案，使用 UNION stackoverflow.com/q/40304011/2235885
是的，我之前有 UNION SELECT 与交换字段:-) 但我正在尝试优化此语句，因为它是最常调用的语句。
那么，对于select union 和select case 这两个查询，也许可以查看explain (analyze) 以获得更多洞察力。