Question

这是我的函数的简化版本，其中包含查询（因此任何变量现在都没有用），并且该函数不会完成，但是如果我单独运行相同的查询，它会在一秒钟内完成。

无法完成的功能

select * from test_function_difference(1);


CREATE OR REPLACE FUNCTION test_function_difference ( 
  p_does_nothing int

)  
RETURNS TABLE(
  t_datum date,
  t_capacity numeric,
  t_used numeric,
  t_category int,
  t_category_name text,  
  t_used_p numeric,
  t_unused_p numeric
)
  VOLATILE
AS $dbvis$

declare
p_sql text := '';
p_execute text := '';
rec record;
begin

p_sql := 
'
 with 
vytizeni as (
  select 
    date_trunc(''day'',mcz.datum)::date as datum ,  
    sum(zd.v_vytizeni)/3600.0 used
  from v_ui_cdc_s5_misto_cas_zdroj_aggregace mcz
  left join (select * , pul_den as den_noc from v_ui_cdc_s5_misto_cas_zdroj_aggregace_zdrobneni) zd on mcz.id = zd.id
  where
    datum between  ''2018-12-31'' and ''2018-12-31''

    and ( zahranicni = 0 or zahranicni is null )
     and den_noc = -1 
  group by 
    date_trunc(''day'',mcz.datum)::date
)
,kapacita as (
  select
    date_trunc(''day'',datum)::date as datum , 
    sum(obsazeni_g)/3600.0 capacity
  from v_ui_cdc_s5_misto_cas_zdroj_aggregace
  where
    datum between  ''2018-12-31'' and ''2018-12-31''

  group by
    date_trunc(''day'',datum)::date 
)
,zdroj as (
  select 
    k.datum,  
    k.capacity,
    v.used,
    -1 category
  from kapacita k 
  join vytizeni v on k.datum = v.datum
)

select
  c.* , 
  kc.nazev::text categeroy_name,
  case when sum(capacity)over(partition by datum) = 0 then 1 else used/sum(capacity)over(partition by datum) end as used_p,
  greatest(1 - case when sum(capacity)over(partition by datum) = 0 then 1 else sum(used)over(partition by datum)/sum(capacity)over(partition by datum) end,0) as unused_p
from  zdroj  c
left join v_ui_cdc_s5_kategorie_cinnosti kc on kc.id = c.category
order by c.datum
';

raise notice '% ' , p_sql;

RETURN QUERY 
execute p_sql;

END;
$dbvis$ LANGUAGE plpgsql

和我分别运行的查询（在533毫秒内完成）

with 
vytizeni as (
  select 
    date_trunc('day',mcz.datum)::date as datum ,  
    sum(zd.v_vytizeni)/3600.0 used
  from v_ui_cdc_s5_misto_cas_zdroj_aggregace mcz
  left join (select * , pul_den as den_noc from v_ui_cdc_s5_misto_cas_zdroj_aggregace_zdrobneni) zd on mcz.id = zd.id
  where
    datum between  '2018-12-31' and '2018-12-31'

    and ( zahranicni = 0 or zahranicni is null )
     and den_noc = -1 
  group by 
    date_trunc('day',mcz.datum)::date
)
,kapacita as (
  select
    date_trunc('day',datum)::date as datum , 
    sum(obsazeni_g)/3600.0 capacity
  from v_ui_cdc_s5_misto_cas_zdroj_aggregace
  where
    datum between  '2018-12-31' and '2018-12-31'

  group by
    date_trunc('day',datum)::date 
)
,zdroj as (
  select 
    k.datum,  
    k.capacity,
    v.used,
    -1 category
  from kapacita k 
  join vytizeni v on k.datum = v.datum
)

select
  c.* , 
  kc.nazev::text categeroy_name,
  case when sum(capacity)over(partition by datum) = 0 then 1 else used/sum(capacity)over(partition by datum) end as used_p,
  greatest(1 - case when sum(capacity)over(partition by datum) = 0 then 1 else sum(used)over(partition by datum)/sum(capacity)over(partition by datum) end,0) as unused_p
from  zdroj  c
left join v_ui_cdc_s5_kategorie_cinnosti kc on kc.id = c.category
order by c.datum

编辑：在将近28分钟后，我就能从函数中获取结果（我也在周日晚上尝试过，这意味着我拥有整个服务器的资源，因为在正常的加载过程中，即使一小时后仍未完成）然后我独立运行查询并在2.1秒后获得结果，这是解释分析

功能：28分钟 https://explain.depesz.com/s/v9xJ

独立查询：2.1秒 https://explain.depesz.com/s/aBri

第二次独立运行430ms https://explain.depesz.com/s/ENva

有趣的是：如果我将时间间隔的开始日期编辑为'2018-12-30'或其他任何日期，该功能也将完成

这意味着

start date = '2018-12-31'
query => finishes under 1 second
function => won't finish
start date = '2018-12-30'
query => finishes under 1 second
function => finishes under 1 second

版本详细信息：x86_64-pc-linux-gnu上的PostgreSQL 10.7，由gcc（GCC）4.8.5 20150623（Red Hat 4.8.5-36）编译，64位

Answer 1

性能差异的原因是该函数内部的执行不使用并行查询，而偶然执行并行执行会选择更好的执行计划。

该功能是否未标记为并行安全？那可能会有所不同。

但是，核心问题是ui_cdc_s5_misto_cas_zdroj_aggregace上扫描的结果行总数的总mies估计，估计为1而不是2243行。

您应该ANALYZE对该表进行更好的估算。如果仅靠这一点不能改善估算值，请尝试在default_statistics_target之前提高ANALYZE。

如果您需要提高default_statistics_target以获得更好的估算，请坚持使用

ALTER TABLE ui_cdc_s5_misto_cas_zdroj_aggregace
   ALTER datum SET STATISTICS <whatever proved useful>

Postgres查询无法完成功能，但如果单独运行，则可以运行

1 个答案: