优化两个数百万行表之间的内部联接

Question

我是Postgres的新手，甚至更了解解释如何运作。我在下面有一个典型的查询，我只是替换日期：

explain
select account_id, 
   security_id, 
   market_value_date, 
   sum(market_value) market_value
from market_value_history mvh
inner join holding_cust hc on hc.id = mvh.owning_object_id
where
hc.account_id = 24766
and market_value_date = '2015-07-02'
and mvh.created_by = 'HoldingLoad'
group by account_id, security_id, market_value_date
order by security_id, market_value_date;

附件是解释的截图 hold_cust表的计数是200万行，market_value_history表有1.63亿行

以下是market_value_history和holding_cust的表定义和索引：

我很感激您在调整此查询时可以给我的任何建议。

CREATE TABLE public.market_value_history
(
  id integer NOT NULL DEFAULT nextval('market_value_id_seq'::regclass),
  market_value numeric(18,6) NOT NULL,
  market_value_date date,
  holding_type character varying(25) NOT NULL,
  owning_object_type character varying(25) NOT NULL,
  owning_object_id integer NOT NULL,
  created_by character varying(50) NOT NULL,
  created_dt timestamp without time zone NOT NULL,
  last_modified_dt timestamp without time zone NOT NULL,
  CONSTRAINT market_value_history_pkey PRIMARY KEY (id)
)
WITH (
  OIDS=FALSE
);
ALTER TABLE public.market_value_history
  OWNER TO postgres;

-- Index: public.ix_market_value_history_id

-- DROP INDEX public.ix_market_value_history_id;

CREATE INDEX ix_market_value_history_id
  ON public.market_value_history
  USING btree
  (owning_object_type COLLATE pg_catalog."default", owning_object_id);

-- Index: public.ix_market_value_history_object_type_date

-- DROP INDEX public.ix_market_value_history_object_type_date;

CREATE UNIQUE INDEX ix_market_value_history_object_type_date
  ON public.market_value_history
  USING btree
  (owning_object_type COLLATE pg_catalog."default", owning_object_id, holding_type COLLATE pg_catalog."default", market_value_date);




CREATE TABLE public.holding_cust
(
  id integer NOT NULL DEFAULT nextval('holding_cust_id_seq'::regclass),
  account_id integer NOT NULL,
  security_id integer NOT NULL,
  subaccount_type integer,
  trade_date date,
  purchase_date date,
  quantity numeric(18,6),
  net_cost numeric(18,2),
  adjusted_net_cost numeric(18,2),
  open_date date,
  close_date date,
  created_by character varying(50) NOT NULL,
  created_dt timestamp without time zone NOT NULL,
  last_modified_dt timestamp without time zone NOT NULL,
  CONSTRAINT holding_cust_pkey PRIMARY KEY (id)
)
WITH (
  OIDS=FALSE
);
ALTER TABLE public.holding_cust
  OWNER TO postgres;

-- Index: public.ix_holding_cust_account_id

-- DROP INDEX public.ix_holding_cust_account_id;

CREATE INDEX ix_holding_cust_account_id
  ON public.holding_cust
  USING btree
  (account_id);

-- Index: public.ix_holding_cust_acctid_secid_asofdt

-- DROP INDEX public.ix_holding_cust_acctid_secid_asofdt;

CREATE INDEX ix_holding_cust_acctid_secid_asofdt
  ON public.holding_cust
  USING btree
  (account_id, security_id, trade_date DESC);

-- Index: public.ix_holding_cust_security_id

-- DROP INDEX public.ix_holding_cust_security_id;

CREATE INDEX ix_holding_cust_security_id
  ON public.holding_cust
  USING btree
  (security_id);

-- Index: public.ix_holding_cust_trade_date

-- DROP INDEX public.ix_holding_cust_trade_date;

CREATE INDEX ix_holding_cust_trade_date
  ON public.holding_cust
  USING btree
  (trade_date);

Answer 1

两件事：

1. 正如德米特里指出的那样，你应该考虑在market_value_date字段上创建一个索引。它可能会发布一个完全不同的查询计划，这可能会或可能不会带来其他瓶颈，但它肯定会删除此seq-Scan。
2. 轻微（因为我怀疑它是否影响性能），但其次，如果您没有按设计强制执行字段长度，您可能希望将createdby字段更改为TEXT。从查询中可以看出，它试图将所有createdby字段强制转换为TEXT以进行此查询。