前言或說切片的想法是如何產生的
故事從這裡開始:
抒情題外話:
正是“那一刻”,因為
那麼,怎樣才能讓客戶滿意,同時提升自己的技能呢?
盡可能地簡化一切,那麼只有兩種方法可以從根本上提高資料庫的效能:
1)粗放路徑-我們增加資源,改變配置;
2)強化路徑-查詢最佳化
我再說一遍,因為當時不再清楚在加速請求中還需要改變什麼,所以選擇了路徑 - 表格設計變更。
初始條件
首先,有這個 ERD(以有條件簡化的方式顯示):
主要特點:
- 多對多關係
- 該表已經有一個潛在的分區鍵
原始請求:
SELECT
p."PARAMETER_ID" as parameter_id,
pc."PC_NAME" AS pc_name,
pc."CUSTOMER_PARTNUMBER" AS customer_partnumber,
w."LASERMARK" AS lasermark,
w."LOTID" AS lotid,
w."REPORTED_VALUE" AS reported_value,
w."LOWER_SPEC_LIMIT" AS lower_spec_limit,
w."UPPER_SPEC_LIMIT" AS upper_spec_limit,
p."TYPE_CALCUL" AS type_calcul,
s."SHIPMENT_NAME" AS shipment_name,
s."SHIPMENT_DATE" AS shipment_date,
extract(year from s."SHIPMENT_DATE") AS year,
extract(month from s."SHIPMENT_DATE") as month,
s."REPORT_NAME" AS report_name,
p."SPARAM_NAME" AS SPARAM_name,
p."CUSTOMERPARAM_NAME" AS customerparam_name
FROM data w INNER JOIN shipment s ON s."SHIPMENT_ID" = w."SHIPMENT_ID"
INNER JOIN parameters p ON p."PARAMETER_ID" = w."PARAMETER_ID"
INNER JOIN shipment_pc sp ON s."SHIPMENT_ID" = sp."SHIPMENT_ID"
INNER JOIN pc pc ON pc."PC_ID" = sp."PC_ID"
INNER JOIN ( SELECT w2."LASERMARK" , MAX(s2."SHIPMENT_DATE") AS "SHIPMENT_DATE"
FROM shipment s2 INNER JOIN data w2 ON s2."SHIPMENT_ID" = w2."SHIPMENT_ID"
GROUP BY w2."LASERMARK"
) md ON md."SHIPMENT_DATE" = s."SHIPMENT_DATE" AND md."LASERMARK" = w."LASERMARK"
WHERE
s."SHIPMENT_DATE" >= '2018-07-01' AND s."SHIPMENT_DATE" <= '2018-09-30' ;
在測試資料庫上執行結果:
價格 :502 997.55
執行時間處理時間:505秒
我們看到了什麼? 基於時間片的常規請求。
讓我們做一個最簡單的邏輯假設:如果有一個時間片的樣本,它會對我們有幫助嗎? 沒錯——分區。
劃分什麼?
乍一看,選擇是顯而易見的 - 使用“SHIPMENT_DATE”鍵對“shipment”表進行聲明性分區(跳得太遠了——最終結果在生產中出了點問題).
如何分區?
這個問題也不是太難。 幸運的是,在 PostgreSQL 10 中,現在有了人工分區機制。
所以:
- 保存來源表的轉儲 - pg_dump 來源表
- 刪除原表—— 刪除表格source_table
- 建立具有範圍分割區的父表 - 建立表格source_table
- 創建部分 - 建立表格source_table,建立索引
- 導入步驟 1 中建立的轉儲 - pg_恢復
分區腳本
為了簡單和方便,步驟 2,3,4、XNUMX、XNUMX 已合併為一個腳本。
所以:
保存來源表的轉儲
pg_dump postgres --file=/dump/shipment.dmp --format=c --table=shipment --verbose > /dump/shipment.log 2>&1
刪除來源表+建立範圍分割區的父表+建立分割區
--create_partition_shipment.sql
do language plpgsql $$
declare
rec_shipment_date RECORD ;
partition_name varchar;
index_name varchar;
current_year varchar ;
current_month varchar ;
begin_year varchar ;
begin_month varchar ;
next_year varchar ;
next_month varchar ;
first_flag boolean ;
i integer ;
begin
RAISE NOTICE 'CREATE TEMPORARY TABLE FOR SHIPMENT_DATE';
CREATE TEMP TABLE tmp_shipment_date as select distinct "SHIPMENT_DATE" from shipment order by "SHIPMENT_DATE" ;
RAISE NOTICE 'DROP TABLE shipment';
drop table shipment cascade ;
CREATE TABLE public.shipment
(
"SHIPMENT_ID" integer NOT NULL DEFAULT nextval('shipment_shipment_id_seq'::regclass),
"SHIPMENT_NAME" character varying(30) COLLATE pg_catalog."default",
"SHIPMENT_DATE" timestamp without time zone,
"REPORT_NAME" character varying(40) COLLATE pg_catalog."default"
)
PARTITION BY RANGE ("SHIPMENT_DATE")
WITH (
OIDS = FALSE
)
TABLESPACE pg_default;
RAISE NOTICE 'CREATE PARTITIONS FOR TABLE shipment';
current_year:='0';
current_month:='0';
begin_year := '0' ;
begin_month := '0' ;
next_year := '0' ;
next_month := '0' ;
FOR rec_shipment_date IN SELECT * FROM tmp_shipment_date LOOP
RAISE NOTICE 'SHIPMENT_DATE=%',rec_shipment_date."SHIPMENT_DATE";
current_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE");
current_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE") ;
IF to_number(current_month,'99') < 10 THEN
current_month := '0'||current_month ;
END IF ;
--Init borders
IF begin_year = '0' THEN
first_flag := true ; --first time flag
begin_year := current_year ;
begin_month := current_month ;
IF current_month = '12' THEN
next_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 year') ;
ELSE
next_year := current_year ;
END IF;
next_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 month') ;
END IF;
-- Check current date into borders NOT for First time
IF to_date( current_year||'.'||current_month, 'YYYY.MM') >= to_date( begin_year||'.'||begin_month, 'YYYY.MM') AND
to_date( current_year||'.'||current_month, 'YYYY.MM') < to_date( next_year||'.'||next_month, 'YYYY.MM') AND
NOT first_flag
THEN
CONTINUE ;
ELSE
--NEW borders only for second and after time
begin_year := current_year ;
begin_month := current_month ;
IF current_month = '12' THEN
next_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 year') ;
ELSE
next_year := current_year ;
END IF;
next_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 month') ;
END IF;
partition_name := 'shipment_shipment_date_'||begin_year||'-'||begin_month||'-01-'|| next_year||'-'||next_month||'-01' ;
EXECUTE format('CREATE TABLE ' || quote_ident(partition_name) || ' PARTITION OF shipment FOR VALUES FROM ( %L ) TO ( %L ) ' , current_year||'-'||current_month||'-01' , next_year||'-'||next_month||'-01' ) ;
index_name := partition_name||'_shipment_id_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_ID") TABLESPACE pg_default ' ) ;
--Drop first time flag
first_flag := false ;
END LOOP;
end
$$;
導入轉儲
pg_restore -d postgres --data-only --format=c --table=shipment --verbose shipment.dmp > /tmp/data_dump/shipment_restore.log 2>&1
檢查分區結果
結果是什麼? 執行計劃的全文又大又乏味,因此很可能僅限於最終的數字。
這是
成本: 502 997.55
執行時間處理時間: 505秒。
已經成為
成本: 77 872.36
執行時間處理時間: 79秒。
相當不錯的結果。 降低了成本和執行時間。 因此,分區的使用給出了預期的效果,並且通常不會令人驚訝。
讓顧客高興
測試結果提交給客戶審核。 經過審查後,他們得到了一個有點出乎意料的結論:“太好了,對‘數據’表進行分區。”
是的,但我們檢查了一個完全不同的「shipment」表;「data」表格沒有「SHIPMENT_DATE」欄位。
沒問題,新增、更改。 最主要的是客戶對結果滿意,實施細節並不是特別重要。
對主表「資料」進行分區
總的來說,沒有出現什麼特別的困難。 當然,分區演算法已經發生了一些變化。
將“SHIPMENT_DATA”欄位加入到“data”表中
psql -h хост -U база -d юзер
=> ALTER TABLE data ADD COLUMN "SHIPMENT_DATE" timestamp without time zone ;
將「data」表中「SHIPMENT_DATA」欄位的值填入為「shipment」表中同名欄位的值
-----------------------------
--update_data.sql
--updating for altered table "data" to values of "shipment_data" from the table "shipment"
--version 1.0
do language plpgsql $$
declare
rec_shipment_data RECORD ;
shipment_date timestamp without time zone ;
row_count integer ;
total_rows integer ;
begin
select count(*) into total_rows from shipment ;
RAISE NOTICE 'Total %',total_rows;
row_count:= 0 ;
FOR rec_shipment_data IN SELECT * FROM shipment LOOP
update data set "SHIPMENT_DATE" = rec_shipment_data."SHIPMENT_DATE" where "SHIPMENT_ID" = rec_shipment_data."SHIPMENT_ID";
row_count:= row_count +1 ;
RAISE NOTICE 'row count = % , from %',row_count,total_rows;
END LOOP;
end
$$;
儲存「資料」表的轉儲
pg_dump postgres --file=/dump/data.dmp --format=c --table=data --verbose > /dump/data.log 2>&1</source
重新建立分區表“data”
--create_partition_data.sql
--create partitions for the table "wafer data" by range column "shipment_data" with one month duration
--version 1.0
do language plpgsql $$
declare
rec_shipment_date RECORD ;
partition_name varchar;
index_name varchar;
current_year varchar ;
current_month varchar ;
begin_year varchar ;
begin_month varchar ;
next_year varchar ;
next_month varchar ;
first_flag boolean ;
i integer ;
begin
RAISE NOTICE 'CREATE TEMPORARY TABLE FOR SHIPMENT_DATE';
CREATE TEMP TABLE tmp_shipment_date as select distinct "SHIPMENT_DATE" from shipment order by "SHIPMENT_DATE" ;
RAISE NOTICE 'DROP TABLE data';
drop table data cascade ;
RAISE NOTICE 'CREATE PARTITIONED TABLE data';
CREATE TABLE public.data
(
"RUN_ID" integer,
"LASERMARK" character varying(20) COLLATE pg_catalog."default" NOT NULL,
"LOTID" character varying(80) COLLATE pg_catalog."default",
"SHIPMENT_ID" integer NOT NULL,
"PARAMETER_ID" integer NOT NULL,
"INTERNAL_VALUE" character varying(75) COLLATE pg_catalog."default",
"REPORTED_VALUE" character varying(75) COLLATE pg_catalog."default",
"LOWER_SPEC_LIMIT" numeric,
"UPPER_SPEC_LIMIT" numeric ,
"SHIPMENT_DATE" timestamp without time zone
)
PARTITION BY RANGE ("SHIPMENT_DATE")
WITH (
OIDS = FALSE
)
TABLESPACE pg_default ;
RAISE NOTICE 'CREATE PARTITIONS FOR TABLE data';
current_year:='0';
current_month:='0';
begin_year := '0' ;
begin_month := '0' ;
next_year := '0' ;
next_month := '0' ;
i := 1;
FOR rec_shipment_date IN SELECT * FROM tmp_shipment_date LOOP
RAISE NOTICE 'SHIPMENT_DATE=%',rec_shipment_date."SHIPMENT_DATE";
current_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE");
current_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE") ;
--Init borders
IF begin_year = '0' THEN
RAISE NOTICE '***Init borders';
first_flag := true ; --first time flag
begin_year := current_year ;
begin_month := current_month ;
IF current_month = '12' THEN
next_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 year') ;
ELSE
next_year := current_year ;
END IF;
next_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 month') ;
END IF;
-- RAISE NOTICE 'current_year=% , current_month=% ',current_year,current_month;
-- RAISE NOTICE 'begin_year=% , begin_month=% ',begin_year,begin_month;
-- RAISE NOTICE 'next_year=% , next_month=% ',next_year,next_month;
-- Check current date into borders NOT for First time
RAISE NOTICE 'Current data = %',to_char( to_date( current_year||'.'||current_month, 'YYYY.MM'), 'YYYY.MM');
RAISE NOTICE 'Begin data = %',to_char( to_date( begin_year||'.'||begin_month, 'YYYY.MM'), 'YYYY.MM');
RAISE NOTICE 'Next data = %',to_char( to_date( next_year||'.'||next_month, 'YYYY.MM'), 'YYYY.MM');
IF to_date( current_year||'.'||current_month, 'YYYY.MM') >= to_date( begin_year||'.'||begin_month, 'YYYY.MM') AND
to_date( current_year||'.'||current_month, 'YYYY.MM') < to_date( next_year||'.'||next_month, 'YYYY.MM') AND
NOT first_flag
THEN
RAISE NOTICE '***CONTINUE';
CONTINUE ;
ELSE
--NEW borders only for second and after time
RAISE NOTICE '***NEW BORDERS';
begin_year := current_year ;
begin_month := current_month ;
IF current_month = '12' THEN
next_year := date_part('year' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 year') ;
ELSE
next_year := current_year ;
END IF;
next_month := date_part('month' ,rec_shipment_date."SHIPMENT_DATE" + interval '1 month') ;
END IF;
IF to_number(current_month,'99') < 10 THEN
current_month := '0'||current_month ;
END IF ;
IF to_number(begin_month,'99') < 10 THEN
begin_month := '0'||begin_month ;
END IF ;
IF to_number(next_month,'99') < 10 THEN
next_month := '0'||next_month ;
END IF ;
RAISE NOTICE 'current_year=% , current_month=% ',current_year,current_month;
RAISE NOTICE 'begin_year=% , begin_month=% ',begin_year,begin_month;
RAISE NOTICE 'next_year=% , next_month=% ',next_year,next_month;
partition_name := 'data_'||begin_year||begin_month||'01_'||next_year||next_month||'01' ;
RAISE NOTICE 'PARTITION NUMBER % , TABLE NAME =%',i , partition_name;
EXECUTE format('CREATE TABLE ' || quote_ident(partition_name) || ' PARTITION OF data FOR VALUES FROM ( %L ) TO ( %L ) ' , begin_year||'-'||begin_month||'-01' , next_year||'-'||next_month||'-01' ) ;
index_name := partition_name||'_shipment_id_parameter_id_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_ID", "PARAMETER_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_lasermark_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("LASERMARK" COLLATE pg_catalog."default") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_shipment_id_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_parameter_id_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("PARAMETER_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_shipment_date_idx';
RAISE NOTICE 'INDEX NAME =%',index_name;
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_DATE") TABLESPACE pg_default ' ) ;
--Drop first time flag
first_flag := false ;
END LOOP;
end
$$;
載入在步驟 3 中建立的轉儲。
pg_restore -h хост -юзер -d база --data-only --format=c --table=data --verbose data.dmp > data_restore.log 2>&1
為舊資料建立一個單獨的部分
---------------------------------------------------
--create_partition_for_old_dates.sql
--create partitions for keeping old dates
--version 1.0
do language plpgsql $$
declare
rec_shipment_date RECORD ;
partition_name varchar;
index_name varchar;
begin
SELECT min("SHIPMENT_DATE") AS min_date INTO rec_shipment_date from data ;
RAISE NOTICE 'Old date is %',rec_shipment_date.min_date ;
partition_name := 'data_old_dates' ;
RAISE NOTICE 'PARTITION NAME IS %',partition_name;
EXECUTE format('CREATE TABLE ' || quote_ident(partition_name) || ' PARTITION OF data FOR VALUES FROM ( %L ) TO ( %L ) ' , '1900-01-01' ,
to_char( rec_shipment_date.min_date,'YYYY')||'-'||to_char(rec_shipment_date.min_date,'MM')||'-01' ) ;
index_name := partition_name||'_shipment_id_parameter_id_idx';
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_ID", "PARAMETER_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_lasermark_idx';
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("LASERMARK" COLLATE pg_catalog."default") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_shipment_id_idx';
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_parameter_id_idx';
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("PARAMETER_ID") TABLESPACE pg_default ' ) ;
index_name := partition_name||'_shipment_date_idx';
EXECUTE format('CREATE INDEX ' || quote_ident(index_name) || ' ON '|| quote_ident(partition_name) ||' USING btree ("SHIPMENT_DATE") TABLESPACE pg_default ' ) ;
end
$$;
最終結果:
這是
成本: 502 997.55
執行時間處理時間:505秒
已經成為
成本: 68 533.70
執行時間處理時間: 69秒
值得,相當值得。 考慮到一路走來,我們或多或少掌握了 PostgreSQL 10 中的分區機制——這是一個非常好的結果。
抒情題外話
是否有可能做得更好 - 是的,你可以!為此,您需要使用物化視圖。
建立物化視圖 LASERMARK_VIEW
CREATE MATERIALIZED VIEW LASERMARK_VIEW
AS
SELECT w."LASERMARK" , MAX(s."SHIPMENT_DATE") AS "SHIPMENT_DATE"
FROM shipment s INNER JOIN data w ON s."SHIPMENT_ID" = w."SHIPMENT_ID"
GROUP BY w."LASERMARK" ;
CREATE INDEX lasermark_vw_shipment_date_ind on lasermark_view USING btree ("SHIPMENT_DATE") TABLESPACE pg_default;
analyze lasermark_view ;
我們再次重寫請求:
使用物化視圖查詢
SELECT
p."PARAMETER_ID" as parameter_id,
pc."PC_NAME" AS pc_name,
pc."CUSTOMER_PARTNUMBER" AS customer_partnumber,
w."LASERMARK" AS lasermark,
w."LOTID" AS lotid,
w."REPORTED_VALUE" AS reported_value,
w."LOWER_SPEC_LIMIT" AS lower_spec_limit,
w."UPPER_SPEC_LIMIT" AS upper_spec_limit,
p."TYPE_CALCUL" AS type_calcul,
s."SHIPMENT_NAME" AS shipment_name,
s."SHIPMENT_DATE" AS shipment_date,
extract(year from s."SHIPMENT_DATE") AS year,
extract(month from s."SHIPMENT_DATE") as month,
s."REPORT_NAME" AS report_name,
p."STC_NAME" AS STC_name,
p."CUSTOMERPARAM_NAME" AS customerparam_name
FROM data w INNER JOIN shipment s ON s."SHIPMENT_ID" = w."SHIPMENT_ID"
INNER JOIN parameters p ON p."PARAMETER_ID" = w."PARAMETER_ID"
INNER JOIN shipment_pc sp ON s."SHIPMENT_ID" = sp."SHIPMENT_ID"
INNER JOIN pc pc ON pc."PC_ID" = sp."PC_ID"
INNER JOIN LASERMARK_VIEW md ON md."SHIPMENT_DATE" = s."SHIPMENT_DATE" AND md."LASERMARK" = w."LASERMARK"
WHERE
s."SHIPMENT_DATE" >= '2018-07-01' AND s."SHIPMENT_DATE" <= '2018-09-30';
我們得到另一個結果:
這是
成本: 502 997.55
執行時間處理時間:505秒
已經成為
成本: 42 481.16
執行時間處理時間: 43秒。
當然,儘管這樣一個有希望的結果具有欺騙性,但想法需要更新。 所以接收資料的總時間不會有太大幫助。 但作為一個實驗,它是相當有趣的。
事實上,事實證明,再次感謝
後記
所以,客戶很滿意。 和 到 利用這種情況。
新任務:您能想出什麼來深化和擴展?
然後我想起來了——夥計們,我們沒有對 PostgreSQL 資料庫進行監控。
實際上,AWS 上仍然有一些以 Cloud Watch 形式進行的監控。 但這種監控對 DBA 來說有什麼好處呢? 一般來說,幾乎沒有。
如果您有機會為自己做一些有用且有趣的事情,您就不能不利用這個機會...
為了
這就是我們如何來到最有趣的部分:
3 年 2018 月 XNUMX 日。
決定開始研究監控 PostgreSQL 查詢效能的可用功能。
但這是一個完全不同的故事。
待續…
來源: www.habr.com