优化大型子表的日期查询：GiST 还是 GIN？

Question

问题

72 个子表，每个子表都有一个年份索引和一个站点索引，定义如下：

CREATE TABLE climate.measurement_12_013
(
-- Inherited from table climate.measurement_12_013:  id bigint NOT NULL DEFAULT nextval('climate.measurement_id_seq'::regclass),
-- Inherited from table climate.measurement_12_013:  station_id integer NOT NULL,
-- Inherited from table climate.measurement_12_013:  taken date NOT NULL,
-- Inherited from table climate.measurement_12_013:  amount numeric(8,2) NOT NULL,
-- Inherited from table climate.measurement_12_013:  category_id smallint NOT NULL,
-- Inherited from table climate.measurement_12_013:  flag character varying(1) NOT NULL DEFAULT ' '::character varying,
  CONSTRAINT measurement_12_013_category_id_check CHECK (category_id = 7),
  CONSTRAINT measurement_12_013_taken_check CHECK (date_part('month'::text, taken)::integer = 12)
)
INHERITS (climate.measurement)

CREATE INDEX measurement_12_013_s_idx
  ON climate.measurement_12_013
  USING btree
  (station_id);
CREATE INDEX measurement_12_013_y_idx
  ON climate.measurement_12_013
  USING btree
  (date_part('year'::text, taken));

（外键约束稍后添加。）

由于全表扫描，以下查询运行非常缓慢：

SELECT
  count(1) AS measurements,
  avg(m.amount) AS amount
FROM
  climate.measurement m
WHERE
  m.station_id IN (
    SELECT
      s.id
    FROM
      climate.station s,
      climate.city c
    WHERE
        /* For one city... */
        c.id = 5182 AND

        /* Where stations are within an elevation range... */
        s.elevation BETWEEN 0 AND 3000 AND

        /* and within a specific radius... */
        6371.009 * SQRT( 
          POW(RADIANS(c.latitude_decimal - s.latitude_decimal), 2) +
            (COS(RADIANS(c.latitude_decimal + s.latitude_decimal) / 2) *
              POW(RADIANS(c.longitude_decimal - s.longitude_decimal), 2))
        ) <= 50
    ) AND

  /* Data before 1900 is shaky; insufficient after 2009. */
  extract( YEAR FROM m.taken ) BETWEEN 1900 AND 2009 AND

  /* Whittled down by category... */
  m.category_id = 1 AND

  /* Between the selected days and years... */
  m.taken BETWEEN
   /* Start date. */
   (extract( YEAR FROM m.taken )||'-01-01')::date AND
    /* End date. Calculated by checking to see if the end date wraps
       into the next year. If it does, then add 1 to the current year.
    */
    (cast(extract( YEAR FROM m.taken ) + greatest( -1 *
      sign(
        (extract( YEAR FROM m.taken )||'-12-31')::date -
        (extract( YEAR FROM m.taken )||'-01-01')::date ), 0
    ) AS text)||'-12-31')::date
GROUP BY
  extract( YEAR FROM m.taken )

迟钝来自这部分查询：

  m.taken BETWEEN
    /* Start date. */
  (extract( YEAR FROM m.taken )||'-01-01')::date AND
    /* End date. Calculated by checking to see if the end date wraps
      into the next year. If it does, then add 1 to the current year.
    */
    (cast(extract( YEAR FROM m.taken ) + greatest( -1 *
      sign(
        (extract( YEAR FROM m.taken )||'-12-31')::date -
        (extract( YEAR FROM m.taken )||'-01-01')::date ), 0
    ) AS text)||'-12-31')::date

这部分查询匹配选定的日期。例如，如果用户想要查看有数据的所有年份的 6 月 1 日至 7 月 1 日之间的数据，则上述子句仅与那些日子匹配。如果用户想查看 12 月 22 日和 3 月 22 日之间的数据，同样对于所有有数据的年份，上述子句计算 3 月 22 日是在下一年的 12 月 22 日，因此相应地匹配日期：

目前日期固定为 1 月 1 日至 12 月 31 日，但将参数化，如上所示。

计划中的 HashAggregate 显示成本为 10006220141.11，我怀疑这是天文数字。

对正在执行的测量表（本身既没有数据也没有索引）执行全表扫描。该表从其子表中聚合了 2.73 亿行。

问题

索引日期以避免全表扫描的正确方法是什么？

我考虑过的选项：

• 杜松子酒
• GiST
• 重写 WHERE 子句
• 将 year_taken、month_taken 和 day_taken 列与表格分开

你有什么想法？

谢谢！

Answer 1

您的问题是您有一个取决于日期计算的 where 子句。如果数据库需要获取每一行并在知道日期是否匹配之前对其进行计算，那么数据库就无法使用索引。

除非您将其重写为数据库具有固定范围以检查哪些不依赖于要检索的数据的形式，否则您将始终必须扫描表。

Answer 2

试试这样的：

create temporary table test (d date);

insert into test select '1970-01-01'::date+generate_series(1,50*365);

analyze test

create function month_day(d date) returns int as $$
  select extract(month from $1)::int*100+extract(day from $1)::int $$
language sql immutable strict;

create index test_d_month_day_idx on test (month_day(d));

explain analyze select * from test
  where month_day(d)>=month_day('2000-04-01')
  and month_day(d)<=month_day('2000-04-05');

Answer 3

我认为要在这些分区上高效运行，我会让您的应用在日期范围方面更加智能。让它生成一个实际的日期列表来检查每个分区，然后让它在分区之间生成一个带有 UNION 的查询。听起来您的数据集非常静态，因此日期索引上的 CLUSTER 也可以大大提高性能。