doris实现数据聚合的三种方式--多明细聚合、物化视图与rollup

doris实现数据聚合的三种方式

假设以
①ds、hour为维度计算pv
②ds为维度计算pv

1、三种聚合方式
（1）聚合模型+数据源多次写入

flink写入kafka代码

insert into log_exp_pv
select ds
hour,
 device_id
from
  dwd_kafka_log
;

insert into log_exp_pv
select ds
'ALL' hour,
 device_id
from
  dwd_kafka_log
;

doris建表语句

create table log_exp_pv(
ds date,
hour varchar(1000),
pv bigint sum DEFAULT '0'
)
AGGREGATE KEY(ds,hour)
partition by range(ds)
(start('20220514') end ('20220520') every (INTERVAL 1 day))
distributed by hash(platform) buckets 32
PROPERTIES(
"dynamic_partition.enable" = "true",
"dynamic_partition.time_unit" = "DAY",
"dynamic_partition.start" = "-1",
"dynamic_partition.end" = "7",
"dynamic_partition.prefix" = "p",
"dynamic_partition.buckets" = "32",
"replication_num" = "1"
)
;

routine load 从kafka导入数据到doris

CREATE ROUTINE LOAD routine_load_log_exp_pv ON log_exp_pv
COLUMNS TERMINATED BY ",",
COLUMNS (
      ds
      ,hour
      ,pv = 1
      )
PROPERTIES
(
    "desired_concurrent_number"="3",
    "max_error_number"="0",
    "strict_mode" = "false",
    "format" = "json"
)
FROM KAFKA
(
    "kafka_broker_list"= "",
    "kafka_topic" = "log_exp_pv",
    "property.group.id"="routine_load_log_exp_pv"
);

（2）明细模型 + 物化视图
flink写入kafka代码

insert into log_exp_pv
select ds
hour,
 device_id
from
  dwd_kafka_log
;

doris建表语句

create table log_exp_detail(
ds date,
hour varchar(1000),
user_id bigint
)
duplicate key(ds,hour)
partition by range(ds)
(start('20220512') end ('20220520') every (INTERVAL 1 day))
distributed by hash(channel) buckets 32
PROPERTIES(
"dynamic_partition.enable" = "true",
"dynamic_partition.time_unit" = "DAY",
"dynamic_partition.start" = "-1",
"dynamic_partition.end" = "7",
"dynamic_partition.prefix" = "p",
"dynamic_partition.buckets" = "32",
"replication_num" = "1"
);

routine load 从kafka导入数据到doris

CREATE ROUTINE LOAD routine_load_log_exp_detail ON log_exp_detail
COLUMNS TERMINATED BY ",",
COLUMNS (
      ds
      ,hour
      ,user_id
      )
PROPERTIES
(
    "desired_concurrent_number"="3",
    "max_error_number"="0",
    "strict_mode" = "false",
    "format" = "json"
)
FROM KAFKA
(
    "kafka_broker_list"= "",
    "kafka_topic" = "log_exp_detail",
    "property.group.id"="routine_load_log_exp_detail"
);

创建物化视图

##ds维度
CREATE MATERIALIZED VIEW mv_ds AS 
SELECT ds
,count(user_id) pv
FROM log_exp_detail 
GROUP BY ds;

##ds、hour维度
CREATE MATERIALIZED VIEW mv_ds_hour AS 
SELECT ds
,hour
,count(user_id) pv
FROM log_exp_detail 
GROUP BY ds
,hour;

查询计划

explain
SELECT ds
,count(user_id) pv
FROM log_job_exp_detail 
GROUP BY ds;

（3）聚合模型 + rollup
doris建表语句

create table log_exp_pv(
ds date,
hour varchar(1000),
pv bigint sum DEFAULT '0'
)
AGGREGATE KEY(ds,hour)
partition by range(ds)
(start('20220514') end ('20220520') every (INTERVAL 1 day))
distributed by hash(platform) buckets 32
rollup (
r1(ds)
,r2(ds,hour)
)
PROPERTIES(
"dynamic_partition.enable" = "true",
"dynamic_partition.time_unit" = "DAY",
"dynamic_partition.start" = "-1",
"dynamic_partition.end" = "7",
"dynamic_partition.prefix" = "p",
"dynamic_partition.buckets" = "32",
"replication_num" = "1"
)
;

routine load 从kafka导入数据到doris

CREATE ROUTINE LOAD routine_load_log_exp_pv ON log_exp_pv
COLUMNS TERMINATED BY ",",
COLUMNS (
      ds
      ,hour
      ,user_id
      )
PROPERTIES
(
    "desired_concurrent_number"="3",
    "max_error_number"="0",
    "strict_mode" = "false",
    "format" = "json"
)
FROM KAFKA
(
    "kafka_broker_list"= "",
    "kafka_topic" = "log_exp_pv",
    "property.group.id"="routine_load_log_exp_pv"
);

查询计划

explain
SELECT ds
,count(user_id) pv
FROM log_exp_pv
GROUP BY ds;

2、三种导入方式对比

（1）聚合模型+数据源多次写入
优点：
①只需关心聚合模型即可
②存储为聚合模型，降低了数据量
缺点：
①假设维度为n，则数据量会膨胀2^n倍，导致数据导入压力增大
②维度过多时，明细数据开发复杂度增加
（2）明细模型 + 物化视图
优点：
①明细数据只需一份
②明细模型保留原始数据
缺点：
①存储为明细数据，导致存储数据量较大
②多维度时，物化视图需要建立多次
③物化视图语法有严格限制
（3）聚合模型 + rollup
优点：
①存储为聚合模型，降低了数据量
②建表时即可指定rollup，降低了开发复杂性
缺点：
①明细数据无法保留