Hive SQL查询效率提升之Analyze方案的实施

0.简介

Analyze，分析表（也称为计算统计信息）是一种内置的Hive操作，可以执行该操作来收集表上的元数据信息。这可以极大的改善表上的查询时间，因为它收集构成表中数据的行计数，文件计数和文件大小（字节），并在执行之前将其提供给查询计划程序。

1.如何分析表？

基础分析语句

ANALYZE TABLE my_database_name.my_table_name COMPUTE STATISTICS;

这是一个基础分析语句，不限制是否存在表分区，如果你是分区表更应该定期执行。

分析特定分区

ANALYZE TABLE my_database_name.my_table_name PARTITION (YEAR=2019, MONTH=5, DAY=12) COMPUTE STATISTICS;

这是一个细粒度的分析语句。它收集指定的分区上的元数据，并将该信息存储在Hive Metastore中已进行查询优化。该信息包括每列，不同值的数量，NULL值的数量，列的平均大小，平均值或列中所有值的总和（如果类型为数字）和值的百分数。

分析列

ANALYZE TABLE my_database_name.my_table_name COMPUTE STATISTICS FOR column1, column2, column3;

它收集指定列上的元数据，并将该信息存储在Hive Metastore中以进行查询优化。该信息包括每列，不同值的数量，NULL值的数量，列的平均大小，平均值或列中所有值的总和（如果类型为数字）和值的百分数。

分析指定分区上的列

ANALYZE TABLE my_database_name.my_table_name PARTITION (YEAR=2019, MONTH=5, DAY=12, HOUR=0) COMPUTE STATISTICS for column1, column2, column3;ANALYZE TABLE my_database_name.my_table_name PARTITION (YEAR=2019, MONTH=5, DAY=12, HOUR) COMPUTE STATISTICS for column1, column2, column3;ANALYZE TABLE my_database_name.my_table_name PARTITION (YEAR=2019, MONTH=5, DAY=12, HOUR) COMPUTE STATISTICS FOR COLUMNS;

第一个SQL只会分析单小时分区上的列信息；
第二个SQL会分析单天分区上的列信息；
第三个SQL会分析单天分区上的所有列信息。

2.效果验证

测试案例1

数据准备
选取KS3线上数据集、TPC-DS基准测试数据集作为样本。结合Hive表分析操作，对多个文件格式以及压缩算法下的数据查询时间进行比对。

SELECT count(DISTINCT(uuid)) AS script_appentry_30day_uv
FROM test_hive.document_assistant
WHERE dt >= '2019-03-12'AND dt <= '2019-04-10'AND p3 = '14'AND p5 = 'script_appentry'

测试结果

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-Sn5VklqR-1587978901532)(https://github.com/buildupchao/ImgStore/blob/master/blog/analyze_table2.png?raw=true)]

测试案例2

数据准备（TPC-DS基础测试）
- 美国事务处理效能委员会(TPC,Transaction Processing Performance Council) ：是目前最知名的非赢利的数据管理系统评测基准标准化组织。它定义了多组标准测试集用于客观地、可重现地评测数据库的性能。
- TPC-DS测试基准是TPC组织推出的用于替代TPC-H的下一代决策支持系统测试基准：TPC-DS采用星型、雪花型等多维数据模式。它包含7张事实表，17张维度表，平均每张表有18列。
- TPC-DS 特点：
  - 共99个测试案例，遵循SQL’99和SQL 2003的语法标准，SQL案例比较复杂；
  - 分析的数据量大，并且测试案例是在回答真实的商业问题；
  - 测试案例中包含各种业务模型(如分析报告型，迭代式的联机分析型，数据挖掘型等)；
  - 几乎所有的测试案例都有很高的IO负载和CPU计算需求。

场景：单事实表、多个维表，复杂的Join
Store_Sales表记录数：2,879,987,999
事实表存储大小（GB）：Text：390， Parquet(Gzip)：116， Orc(Zlib):131

query27.sql:

-- start query 1 in stream 0 using template query27.tpl and seed 2017787633
select  i_item_id,s_state, grouping(s_state) g_state,avg(ss_quantity) agg1,avg(ss_list_price) agg2,avg(ss_coupon_amt) agg3,avg(ss_sales_price) agg4from store_sales, customer_demographics, date_dim, store, itemwhere ss_sold_date_sk = d_date_sk andss_item_sk = i_item_sk andss_store_sk = s_store_sk andss_cdemo_sk = cd_demo_sk andcd_gender = 'M' andcd_marital_status = 'U' andcd_education_status = '2 yr Degree' andd_year = 2001 ands_state in ('SD','FL', 'MI', 'LA', 'MO', 'SC')group by rollup (i_item_id, s_state)order by i_item_id,s_statelimit 100;-- end query 1 in stream 0 using template query27.tpl

query28.sql:

-- start query 1 in stream 0 using template query28.tpl and seed 444293455
select  *
from (select avg(ss_list_price) B1_LP,count(ss_list_price) B1_CNT,count(distinct ss_list_price) B1_CNTDfrom store_saleswhere ss_quantity between 0 and 5and (ss_list_price between 11 and 11+10or ss_coupon_amt between 460 and 460+1000or ss_wholesale_cost between 14 and 14+20)) B1,(select avg(ss_list_price) B2_LP,count(ss_list_price) B2_CNT,count(distinct ss_list_price) B2_CNTDfrom store_saleswhere ss_quantity between 6 and 10and (ss_list_price between 91 and 91+10or ss_coupon_amt between 1430 and 1430+1000or ss_wholesale_cost between 32 and 32+20)) B2,(select avg(ss_list_price) B3_LP,count(ss_list_price) B3_CNT,count(distinct ss_list_price) B3_CNTDfrom store_saleswhere ss_quantity between 11 and 15and (ss_list_price between 66 and 66+10or ss_coupon_amt between 920 and 920+1000or ss_wholesale_cost between 4 and 4+20)) B3,(select avg(ss_list_price) B4_LP,count(ss_list_price) B4_CNT,count(distinct ss_list_price) B4_CNTDfrom store_saleswhere ss_quantity between 16 and 20and (ss_list_price between 142 and 142+10or ss_coupon_amt between 3054 and 3054+1000or ss_wholesale_cost between 80 and 80+20)) B4,(select avg(ss_list_price) B5_LP,count(ss_list_price) B5_CNT,count(distinct ss_list_price) B5_CNTDfrom store_saleswhere ss_quantity between 21 and 25and (ss_list_price between 135 and 135+10or ss_coupon_amt between 14180 and 14180+1000or ss_wholesale_cost between 38 and 38+20)) B5,(select avg(ss_list_price) B6_LP,count(ss_list_price) B6_CNT,count(distinct ss_list_price) B6_CNTDfrom store_saleswhere ss_quantity between 26 and 30and (ss_list_price between 28 and 28+10or ss_coupon_amt between 2513 and 2513+1000or ss_wholesale_cost between 42 and 42+20)) B6
limit 100;-- end query 1 in stream 0 using template query28.tpl

query43.sql:

-- start query 1 in stream 0 using template query43.tpl and seed 1819994127
select  s_store_name, s_store_id,sum(case when (d_day_name='Sunday') then ss_sales_price else null end) sun_sales,sum(case when (d_day_name='Monday') then ss_sales_price else null end) mon_sales,sum(case when (d_day_name='Tuesday') then ss_sales_price else  null end) tue_sales,sum(case when (d_day_name='Wednesday') then ss_sales_price else null end) wed_sales,sum(case when (d_day_name='Thursday') then ss_sales_price else null end) thu_sales,sum(case when (d_day_name='Friday') then ss_sales_price else null end) fri_sales,sum(case when (d_day_name='Saturday') then ss_sales_price else null end) sat_salesfrom date_dim, store_sales, storewhere d_date_sk = ss_sold_date_sk ands_store_sk = ss_store_sk ands_gmt_offset = -6 andd_year = 1998group by s_store_name, s_store_idorder by s_store_name, s_store_id,sun_sales,mon_sales,tue_sales,wed_sales,thu_sales,fri_sales,sat_saleslimit 100;-- end query 1 in stream 0 using template query43.tpl

query67.sql:

-- start query 1 in stream 0 using template query67.tpl and seed 1819994127
select  *
from (select i_category,i_class,i_brand,i_product_name,d_year,d_qoy,d_moy,s_store_id,sumsales,rank() over (partition by i_category order by sumsales desc) rkfrom (select i_category,i_class,i_brand,i_product_name,d_year,d_qoy,d_moy,s_store_id,sum(coalesce(ss_sales_price*ss_quantity,0)) sumsalesfrom store_sales,date_dim,store,itemwhere  ss_sold_date_sk=d_date_skand ss_item_sk=i_item_skand ss_store_sk = s_store_skand d_month_seq between 1212 and 1212+11group by  rollup(i_category, i_class, i_brand, i_product_name, d_year, d_qoy, d_moy,s_store_id))dw1) dw2
where rk <= 100
order by i_category,i_class,i_brand,i_product_name,d_year,d_qoy,d_moy,s_store_id,sumsales,rk
limit 100;-- end query 1 in stream 0 using template query67.tpl

query46.sql:

-- start query 1 in stream 0 using template query46.tpl and seed 803547492
select  c_last_name,c_first_name,ca_city,bought_city,ss_ticket_number,amt,profitfrom(select ss_ticket_number,ss_customer_sk,ca_city bought_city,sum(ss_coupon_amt) amt,sum(ss_net_profit) profitfrom store_sales,date_dim,store,household_demographics,customer_addresswhere store_sales.ss_sold_date_sk = date_dim.d_date_skand store_sales.ss_store_sk = store.s_store_skand store_sales.ss_hdemo_sk = household_demographics.hd_demo_skand store_sales.ss_addr_sk = customer_address.ca_address_skand (household_demographics.hd_dep_count = 2 orhousehold_demographics.hd_vehicle_count= 1)and date_dim.d_dow in (6,0)and date_dim.d_year in (1998,1998+1,1998+2)and store.s_city in ('Cedar Grove','Wildwood','Union','Salem','Highland Park')group by ss_ticket_number,ss_customer_sk,ss_addr_sk,ca_city) dn,customer,customer_address current_addrwhere ss_customer_sk = c_customer_skand customer.c_current_addr_sk = current_addr.ca_address_skand current_addr.ca_city <> bought_cityorder by c_last_name,c_first_name,ca_city,bought_city,ss_ticket_numberlimit 100;-- end query 1 in stream 0 using template query46.tpl

query7.sql:

-- start query 1 in stream 0 using template query7.tpl and seed 1930872976
select  i_item_id,avg(ss_quantity) agg1,avg(ss_list_price) agg2,avg(ss_coupon_amt) agg3,avg(ss_sales_price) agg4from store_sales, customer_demographics, date_dim, item, promotionwhere ss_sold_date_sk = d_date_sk andss_item_sk = i_item_sk andss_cdemo_sk = cd_demo_sk andss_promo_sk = p_promo_sk andcd_gender = 'F' andcd_marital_status = 'W' andcd_education_status = 'Primary' and(p_channel_email = 'N' or p_channel_event = 'N') andd_year = 1998group by i_item_idorder by i_item_idlimit 100;-- end query 1 in stream 0 using template query7.tpl

query73.sql:

-- start query 1 in stream 0 using template query73.tpl and seed 1971067816
select c_last_name,c_first_name,c_salutation,c_preferred_cust_flag,ss_ticket_number,cnt from(select ss_ticket_number,ss_customer_sk,count(*) cntfrom store_sales,date_dim,store,household_demographicswhere store_sales.ss_sold_date_sk = date_dim.d_date_skand store_sales.ss_store_sk = store.s_store_skand store_sales.ss_hdemo_sk = household_demographics.hd_demo_skand date_dim.d_dom between 1 and 2and (household_demographics.hd_buy_potential = '>10000' orhousehold_demographics.hd_buy_potential = 'unknown')and household_demographics.hd_vehicle_count > 0and case when household_demographics.hd_vehicle_count > 0 thenhousehold_demographics.hd_dep_count/ household_demographics.hd_vehicle_count else null end > 1and date_dim.d_year in (2000,2000+1,2000+2)and store.s_county in ('Mobile County','Maverick County','Huron County','Kittitas County')group by ss_ticket_number,ss_customer_sk) dj,customerwhere ss_customer_sk = c_customer_skand cnt between 1 and 5order by cnt desc;-- end query 1 in stream 0 using template query73.tpl

测试结果

3.结论

Hive执行表分析后能大幅加速查询速度
- 查询耗时（压缩算法）：None > Snappy > Gzip/Zlib
- 查询耗时（文件格式）：Text > Parquet > Orc
当前测试场景下，ORC格式查询耗时最低
- Parquet与Orc查询耗时接近