表现 – Postgres哈希加盟速度

发布时间：2020-12-13 16:01:27 所属栏目：百科来源：网络整理

导读：我有3个表,我希望使用Postgres 9.1中的内部联接,读取,设备和device_patients加入.下面是每个表的缩写模式. 读 – 约250,000行 CREATE TABLE reads ( id serial NOT NULL,device_id integer NOT NULL,value bigint NOT NULL,read_datetime timestamp without

我有3个表,我希望使用Postgres 9.1中的内部联接,读取,设备和device_patients加入.下面是每个表的缩写模式.

读 – 约250,000行

CREATE TABLE reads
  (
    id serial NOT NULL,device_id integer NOT NULL,value bigint NOT NULL,read_datetime timestamp without time zone NOT NULL,created_at timestamp without time zone NOT NULL,updated_at timestamp without time zone NOT NULL,CONSTRAINT reads_pkey PRIMARY KEY (id )
  )
WITH (
  OIDS=FALSE
);
ALTER TABLE reads
  OWNER TO postgres;

CREATE INDEX index_reads_on_device_id
  ON reads
  USING btree
  (device_id );

CREATE INDEX index_reads_on_read_datetime
  ON reads
  USING btree
  (read_datetime );

设备 – 约500行

CREATE TABLE devices
(
  id serial NOT NULL,serial_number character varying(20) NOT NULL,CONSTRAINT devices_pkey PRIMARY KEY (id )
)
WITH (
  OIDS=FALSE
);
ALTER TABLE devices
  OWNER TO postgres;

CREATE UNIQUE INDEX index_devices_on_serial_number
  ON devices
  USING btree
  (serial_number COLLATE pg_catalog."default" );

patient_devices – 约25,000行

CREATE TABLE patient_devices
(
  id serial NOT NULL,patient_id integer NOT NULL,issuance_datetime timestamp without time zone NOT NULL,unassignment_datetime timestamp without time zone,CONSTRAINT patient_devices_pkey PRIMARY KEY (id )
)
WITH (
  OIDS=FALSE
);
ALTER TABLE patient_devices
  OWNER TO postgres;

CREATE INDEX index_patient_devices_on_device_id
  ON patient_devices
  USING btree
  (device_id );

CREATE INDEX index_patient_devices_on_issuance_datetime
  ON patient_devices
  USING btree
  (issuance_datetime );

CREATE INDEX index_patient_devices_on_patient_id
  ON patient_devices
  USING btree
  (patient_id );

CREATE INDEX index_patient_devices_on_unassignment_datetime
  ON patient_devices
  USING btree
  (unassignment_datetime );

患者 – ~1,000排

CREATE TABLE patients
(
  id serial NOT NULL,first_name character varying(50) NOT NULL,middle_name character varying(50),last_name character varying(50) NOT NULL,CONSTRAINT participants_pkey PRIMARY KEY (id )
)
WITH (
  OIDS=FALSE
);
ALTER TABLE patients
  OWNER TO postgres;

这是我的缩写查询.

SELECT device_patients.patient_id,serial_number FROM reads
INNER JOIN devices ON devices.id = reads.device_id
INNER JOIN patient_devices ON device_patients.device_id = devices.id
WHERE (reads.read_datetime BETWEEN '2012-01-01 10:30:01.000000' AND '2013-05-18 03:03:42')
AND (read_datetime > issuance_datetime) AND ((unassignment_datetime IS NOT NULL AND read_datetime < unassignment_datetime) OR
(unassignment_datetime IS NULL))
GROUP BY serial_number,patient_devices.patient_id LIMIT 10

最终这将是一个更大的查询的一小部分(没有LIMIT,我只添加限制来向自己证明长运行时不是由于返回一堆行),但是我做了一堆实验和确定这是较大查询的缓慢部分.当我在此查询上运行EXPLAIN ANALYZE时,我得到以下输出(也可以查看here)

Limit  (cost=156442.31..156442.41 rows=10 width=13) (actual time=2815.435..2815.441 rows=10 loops=1)
  ->  HashAggregate  (cost=156442.31..159114.89 rows=267258 width=13) (actual time=2815.432..2815.437 rows=10 loops=1)
        ->  Hash Join  (cost=1157.78..151455.79 rows=997304 width=13) (actual time=30.930..2739.164 rows=250150 loops=1)
              Hash Cond: (devices.device_id = devices.id)
              Join Filter: ((reads.read_datetime > patient_devices.issuance_datetime) AND (((patient_devices.unassignment_datetime IS NOT NULL) AND (reads.read_datetime < patient_devices.unassignment_datetime)) OR (patient_devices.unassignment_datetime IS NULL)))
              ->  Seq Scan on reads  (cost=0.00..7236.94 rows=255396 width=12) (actual time=0.035..64.433 rows=255450 loops=1)
                    Filter: ((read_datetime >= '2012-01-01 10:30:01'::timestamp without time zone) AND (read_datetime <= '2013-05-18 03:03:42'::timestamp without time zone))
              ->  Hash  (cost=900.78..900.78 rows=20560 width=37) (actual time=30.830..30.830 rows=25015 loops=1)
                    Buckets: 4096  Batches: 1  Memory Usage: 1755kB
                    ->  Hash Join  (cost=19.90..900.78 rows=20560 width=37) (actual time=0.776..20.551 rows=25015 loops=1)
                          Hash Cond: (patient_devices.device_id = devices.id)
                          ->  Seq Scan on patient_devices  (cost=0.00..581.93 rows=24893 width=24) (actual time=0.014..7.867 rows=25545 loops=1)
                                Filter: ((unassignment_datetime IS NOT NULL) OR (unassignment_datetime IS NULL))
                          ->  Hash  (cost=13.61..13.61 rows=503 width=13) (actual time=0.737..0.737 rows=503 loops=1)
                                Buckets: 1024  Batches: 1  Memory Usage: 24kB
                                ->  Seq Scan on devices  (cost=0.00..13.61 rows=503 width=13) (actual time=0.016..0.466 rows=503 loops=1)
                                      Filter: (entity_id = 2)
Total runtime: 2820.392 ms

我的问题是如何加快速度？现在我在我的Windows机器上运行它进行测试,但最终它将部署在Ubuntu上,这会有所作为吗？任何洞察为什么这需要2秒钟将非常感激.

谢谢

有人建议LIMIT可能正在改变查询计划. Here是没有LIMIT的相同查询.缓慢的部分似乎仍然是哈希加入.

此外,这是相关的调整参数.我现在只在Windows上测试它,我不知道这会对Linux机器产生什么影响

shared_buffers = 2GB
effective_cache_size = 4GB
work_mem = 256MB
random_page_cost = 2.0

以下是读取表的统计信息

Statistic   Value
Sequential Scans    130
Sequential Tuples Read  28865850
Index Scans 283630
Index Tuples Fetched    141421907
Tuples Inserted 255450
Tuples Updated  0
Tuples Deleted  0
Tuples HOT Updated  0
Live Tuples 255450
Dead Tuples 0
Heap Blocks Read    20441
Heap Blocks Hit 3493033
Index Blocks Read   8824
Index Blocks Hit    4840210
Toast Blocks Read   
Toast Blocks Hit    
Toast Index Blocks Read 
Toast Index Blocks Hit  
Last Vacuum 2013-05-20 09:23:03.782-07
Last Autovacuum 
Last Analyze    2013-05-20 09:23:03.91-07
Last Autoanalyze    2013-05-17 19:01:44.075-07
Vacuum counter  1
Autovacuum counter  0
Analyze counter 1
Autoanalyze counter 6
Table Size  27 MB
Toast Table Size    none
Indexes Size    34 MB

以下是设备表的统计信息

Statistic   Value
Sequential Scans    119
Sequential Tuples Read  63336
Index Scans 1053935
Index Tuples Fetched    1053693
Tuples Inserted 609
Tuples Updated  0
Tuples Deleted  0
Tuples HOT Updated  0
Live Tuples 609
Dead Tuples 0
Heap Blocks Read    32
Heap Blocks Hit 1054553
Index Blocks Read   32
Index Blocks Hit    2114305
Toast Blocks Read   
Toast Blocks Hit    
Toast Index Blocks Read 
Toast Index Blocks Hit  
Last Vacuum 
Last Autovacuum 
Last Analyze    
Last Autoanalyze    2013-05-17 19:02:49.692-07
Vacuum counter  0
Autovacuum counter  0
Analyze counter 0
Autoanalyze counter 2
Table Size  48 kB
Toast Table Size    none
Indexes Size    128 kB

以下是patient_devices表的统计信息

Statistic   Value
Sequential Scans    137
Sequential Tuples Read  3065400
Index Scans 853990
Index Tuples Fetched    46143763
Tuples Inserted 25545
Tuples Updated  24936
Tuples Deleted  0
Tuples HOT Updated  0
Live Tuples 25547
Dead Tuples 929
Heap Blocks Read    1959
Heap Blocks Hit 6099617
Index Blocks Read   1077
Index Blocks Hit    2462681
Toast Blocks Read   
Toast Blocks Hit    
Toast Index Blocks Read 
Toast Index Blocks Hit  
Last Vacuum 
Last Autovacuum 2013-05-17 19:01:44.576-07
Last Analyze    
Last Autoanalyze    2013-05-17 19:01:44.697-07
Vacuum counter  0
Autovacuum counter  6
Analyze counter 0
Autoanalyze counter 6
Table Size  2624 kB
Toast Table Size    none
Indexes Size    5312 kB

以下是我正在努力加快的完整查询.较小的查询确实更快,但我无法更快地完成我的完整查询,这将在下面重现.按照建议,我添加了4个新索引,UNIQUE(device_id,issuance_datetime),UNIQUE(patient_id,unassignment_datetime),unassignment_datetime)

SELECT 
first_name,last_name,MAX(max_read) AS read_datetime,SUM(value) AS value,serial_number
FROM (
    SELECT
    pa.first_name,pa.last_name,value,first_value(de.serial_number) OVER(PARTITION BY pa.id ORDER BY re.read_datetime DESC) AS serial_number -- I'm not sure if this is a good way to do this,but I don't know of another way,re.read_datetime,MAX(re.read_datetime) OVER (PARTITION BY pd.id) AS max_read
    FROM reads re
    INNER JOIN devices de ON de.id = re.device_id
    INNER JOIN patient_devices pd ON pd.device_id = de.id
        AND re.read_datetime >= pd.issuance_datetime
        AND re.read_datetime < COALESCE(pd.unassignment_datetime,'infinity'::timestamp)
    INNER JOIN patients pa ON pa.id = pd.patient_id
    WHERE re.read_datetime BETWEEN '2012-01-01 10:30:01' AND '2013-05-18 03:03:42'
) AS foo WHERE read_datetime = max_read
GROUP BY first_name,serial_number ORDER BY value desc
LIMIT 10

很抱歉没有发布这个,但我认为这个查询太复杂了,并试图简单地解决问题,但显然我仍然无法弄明白.如果我可以使用max_read变量限制嵌套选择返回的结果,似乎会更快,但根据众多消息来源,Postgres中不允许这样做.

解决方法

仅供参考：清理查询：

SELECT pd.patient_id,de.serial_number
FROM reads re
INNER JOIN devices de ON de.id = re.device_id
INNER JOIN patient_devices pd ON pd.device_id = de.id
        AND re.read_datetime >= pd.issuance_datetime -- changed this from '>' to '>='
        AND (re.read_datetime < pd.unissuance_datetime OR pd.unissuance_datetime IS NULL)
WHERE re.read_datetime BETWEEN '2012-01-01 10:30:01.000000' AND '2013-05-18 03:03:42'
GROUP BY de.serial_number,pd.patient_id 
LIMIT 10
   ;

更新：没有原始拼写错误：

EXPLAIN ANALYZE
SELECT pd.patient_id,de.serial_number
FROM reads re
INNER JOIN devices de ON de.id = re.device_id
INNER JOIN patient_devices pd ON pd.device_id = de.id
        AND re.read_datetime >= pd.issuance_datetime
        AND (re.read_datetime < pd.unassignment_datetime OR pd.unassignment_datetime IS NULL)
WHERE re.read_datetime BETWEEN '2012-01-01 10:30:01.000000' AND '2013-05-18 03:03:42'
GROUP BY de.serial_number,pd.patient_id
LIMIT 10
  ;

更新：这大约是此处的6倍(在合成数据上,并且数据模型略有改变)

-- Modified data model + synthetic data:
CREATE TABLE devices
( id serial NOT NULL,serial_number character varying(20) NOT NULL
--,created_at timestamp without time zone NOT NULL
--,CONSTRAINT devices_pkey PRIMARY KEY (id ),UNIQUE (serial_number)
) ;

CREATE TABLE reads
  -- ( id serial NOT NULL PRIMARY KEY -- You don't need this surrogate key
  (  device_id integer NOT NULL REFERENCES devices (id),read_datetime timestamp without time zone NOT NULL
  --,created_at timestamp without time zone NOT NULL
  --,PRIMARY KEY ( device_id,read_datetime)
  ) ;



CREATE TABLE patient_devices
-- ( id serial NOT NULL PRIMARY KEY -- You don't need this surrogate key
( patient_id integer NOT NULL -- REFERENCES patients (id),device_id integer NOT NULL REFERENCES devices(id),unassignment_datetime timestamp without time zone
--,PRIMARY KEY (device_id,UNIQUE (device_id,unassignment_datetime)
) ;


-- CREATE INDEX index_patient_devices_on_issuance_datetime ON patient_devices (device_id,unassignment_datetime );
-- may need some additional indices later

-- devices -- ~500 rows
INSERT INTO devices(serial_number) SELECT 'No_' || gs::text FROM generate_series(1,500) gs;

-- reads -- ~100K rows
INSERT INTO reads(device_id,read_datetime,value)
SELECT  de.id,gs,(random()*1000000)::bigint
FROM devices  de
JOIN generate_series('2012-01-01','2013-05-01','1 hour' ::interval) gs
        ON random() < 0.02;

-- patient_devices -- ~25,000 rows
INSERT INTO patient_devices(device_id,issuance_datetime,patient_id)
SELECT DISTINCT ON (re.device_id,read_datetime)
        re.device_id,pa
FROM generate_series(1,100) pa
JOIN reads re
        ON random() < 0.01;

        -- close the open intervals
UPDATE patient_devices dst
SET unassignment_datetime = src.issuance_datetime
FROM patient_devices src
WHERE src.device_id = dst.device_id
AND src.issuance_datetime > dst.issuance_datetime
AND NOT EXISTS ( SELECT *
        FROM patient_devices nx
        WHERE nx.device_id = src.device_id
        AND nx.issuance_datetime > dst.issuance_datetime
        AND nx.issuance_datetime < src.issuance_datetime
        )
        ;

VACUUM ANALYZE patient_devices;
VACUUM ANALYZE devices;
VACUUM ANALYZE reads;


-- EXPLAIN ANALYZE
SELECT pd.patient_id,de.serial_number
        --,COUNT (*) AS zcount
FROM reads re
INNER JOIN devices de ON de.id = re.device_id
INNER JOIN patient_devices pd ON pd.device_id = de.id
        AND re.read_datetime >= pd.issuance_datetime
        AND re.read_datetime < COALESCE(pd.unassignment_datetime,'infinity'::timestamp)
WHERE re.read_datetime BETWEEN '2012-01-01 10:30:01' AND '2013-05-18 03:03:42'
GROUP BY de.serial_number,pd.patient_id
LIMIT 10
        ;

（编辑：李大同）

【声明】本站内容均来自网络，其相关言论仅代表作者个人观点，不代表本站立场。若无意侵犯到您的权利，请及时与联系站长删除相关内容!