Query Profile概覽

可視化Query Profile

StarRocks Manager支持對Profile進行可視化分析，您可以通過StarRocks Manager可視化您的Query Profile。

確認查詢瓶頸

Query Profile優(yōu)化案例

• Bitmap索引

• Bloom filter索引

• 優(yōu)化數(shù)據(jù)傾斜

• 單表物化視圖

• 確認Join左右表Plan是否合理

• 確認JoinRuntimeFilter是否生效

• Colocate Join

• 確認分桶或分區(qū)裁剪是否生效

Bitmap索引

創(chuàng)建索引

• 建表時創(chuàng)建Bitmap索引

  CREATE TABLE `student_info` (
    `s_stukey` bigint(20) NULL COMMENT "",
    `s_name` varchar(65533) NULL COMMENT "",
    `s_gender` varchar(65533) NULL COMMENT "",
    INDEX index1 (s_gender) USING BITMAP COMMENT 'index1'
  ) ENGINE=OLAP
  DUPLICATE KEY(`s_stukey`)
  COMMENT "OLAP"
  DISTRIBUTED BY HASH(`s_stukey`);
  
  INSERT INTO student_info
  VALUES
      (001,'student#000000019','male'),
      (002,'student#000000020','male'),
      (003,'student#000000021','male'),
      (004,'student#000000022','female');

• 建表后使用CREATE INDEX創(chuàng)建Bitmap索引

  CREATE INDEX index_name ON table_name (column_name) [USING BITMAP] [COMMENT ''];

查看創(chuàng)建索引進度

SHOW ALTER TABLE COLUMN [FROM db_name];

查看索引

SHOW {INDEX[ES] | KEY[S] } FROM [db_name.]table_name [FROM db_name];

刪除索引

DROP INDEX index_name ON [db_name.]table_name;

單列查詢測試

1. 執(zhí)行Query，過濾s_gender列。

   select * from student_info where s_gender='male';

2. 查看Profile。

   單擊OLAP_SCAN，然后單擊右側的節(jié)點詳情頁簽，過濾Bitmap可以看到Bitmap索引已經(jīng)生效。

Bloom filter索引

創(chuàng)建索引

CREATE TABLE table1
(
  k1 BIGINT,
  k2 LARGEINT,
  v1 VARCHAR(2048) REPLACE,
  v2 SMALLINT DEFAULT "10"
)
ENGINE = olap
PRIMARY KEY(k1, k2)
DISTRIBUTED BY HASH (k1, k2) BUCKETS 10
PROPERTIES("bloom_filter_columns" = "k1,k2"); --多個索引列之間需用逗號（,）隔開;

查看索引

SHOW CREATE TABLE table1;

修改索引

示例如下所示：

• 增加一個Bloom filter索引列v1。

  ALTER TABLE table1 SET ("bloom_filter_columns" = "k1,k2,v1");

• 減少一個Bloom filter索引列k2。

  ALTER TABLE table1 SET ("bloom_filter_columns" = "k1");

• 刪除table1的所有Bloom filter索引。

  ALTER TABLE table1 SET ("bloom_filter_columns" = "");

實例測試

1. 以TPC_H中的customer表為例，在基數(shù)較高的列且非排序Key的列上增加Bloom filter索引，例如c_phone列。

   ALTER TABLE tpc_h_sf100.customer SET ("bloom_filter_columns" = "c_custkey, c_phone");

2. 查看索引，確認Bloom filter索引已添加。

   SHOW CREATE TABLE tpc_h_sf100.customer;

   

3. 執(zhí)行Query，過濾c_phone列。

   select * from tpc_h_sf100.customer where c_phone = "10-334-921-5346";

4. 查看Profile。

   單擊OLAP_SCAN，然后單擊右側的節(jié)點詳情頁簽，找到BloomFilterFilterRows指標，確認BloomFilter Index生效。

優(yōu)化數(shù)據(jù)傾斜

以TPC_H中的lineitem表為例，選擇非均勻分布的key作為分桶鍵來測試數(shù)據(jù)傾斜問題。本示例中l_tag字段的value分布不均勻。

1. 創(chuàng)建測試數(shù)據(jù)，在TPC_H中的lineitem表中新增一列，作為DISTRIBUTED字段。

   CREATE TABLE `lineitem_tag` (
     `l_orderkey` bigint(20) NULL COMMENT "",
     `l_partkey` bigint(20) NULL COMMENT "",
     `l_suppkey` bigint(20) NULL COMMENT "",
     `l_linenumber` int(11) NULL COMMENT "",
     `l_quantity` double NULL COMMENT "",
     `l_extendedprice` double NULL COMMENT "",
     `l_discount` double NULL COMMENT "",
     `l_tax` double NULL COMMENT "",
     `l_returnflag` varchar(65533) NULL COMMENT "",
     `l_linestatus` varchar(65533) NULL COMMENT "",
     `l_shipdate` date NULL COMMENT "",
     `l_commitdate` date NULL COMMENT "",
     `l_receiptdate` date NULL COMMENT "",
     `l_shipinstruct` varchar(65533) NULL COMMENT "",
     `l_shipmode` varchar(65533) NULL COMMENT "",
     `l_comment` varchar(65533) NULL COMMENT "",
     `l_tag` varchar(65533)  default 'false' COMMENT ""
   ) ENGINE=OLAP
   DUPLICATE KEY(`l_orderkey`)
   COMMENT "OLAP"
   DISTRIBUTED BY HASH(`l_tag`) BUCKETS 96
   PROPERTIES (
   "replication_num" = "1",
   "in_memory" = "false",
   "storage_format" = "DEFAULT",
   "enable_persistent_index" = "false"
   );
   
   insert into lineitem_tag  select *, 'false' as l_tag  from tpc_h_sf100.lineitem;

2. 執(zhí)行Query，進行全表查詢。

   select count(1) from lineitem_tag;

3. 查看Profile。

   單擊OLAP_SCAN，在右側的節(jié)點頁簽，對比MaxTime和MinTime下的SCAN，發(fā)現(xiàn)時間相差好幾個量級，可能發(fā)生了數(shù)據(jù)傾斜。

4. 通過優(yōu)化表結構重新定義DISTRIBUTED字段。

5. 再次查看Profile，對比MaxTime和MinTime下的SCAN，可以發(fā)現(xiàn)數(shù)據(jù)傾斜問題得到優(yōu)化。

單表物化視圖

StarRocks中的單表物化視圖（Rollup）是一種特殊的索引，無法直接查詢。如果您的數(shù)據(jù)倉庫中存在大量復雜或重復的查詢，您可以通過創(chuàng)建單表物化視圖加速查詢。

查詢測試

1. 以TPC_H中l(wèi)ineitem表為例，執(zhí)行Query。

   select l_returnflag,l_linestatus,l_shipmode,sum(l_extendedprice) from lineitem group by l_returnflag,l_linestatus,l_shipmode;

2. 初次查詢，未創(chuàng)建物化視圖查詢耗時1115毫秒。

3. 查看Profile。

   單擊OLAP_SCAN，在右側的節(jié)點頁簽，可以看到Rollup掃描的是lineitem表本身。

創(chuàng)建物化視圖

CREATE MATERIALIZED VIEW material_test AS select l_returnflag,l_linestatus ,l_shipmode,sum(l_extendedprice) from lineitem group by l_returnflag,l_linestatus,l_shipmode;

驗證查詢是否命中單表物化視圖

1. 您可以使用EXPLAIN命令查看該查詢是否命中單表物化視圖。

   explain select l_returnflag,l_linestatus ,l_shipmode,sum(l_extendedprice) from lineitem group by l_returnflag,l_linestatus,l_shipmode;

   

2. 查看Profile。

   單擊OLAP_SCAN，在右側的節(jié)點頁簽，可以看到命中物化視，同時Query用時也更短，耗時0.05毫秒。

確認Join左右表Plan是否合理

通常StarRocks會選擇較小的表作為Join的右表。如果Query Profile顯示右表的數(shù)據(jù)量明顯大于左表，則該Join Plan異常。

1. 以TPC_DH中的query72.sql為例。

   select i_item_desc,
          w_warehouse_name,
          d1.d_week_seq,
          sum(case when p_promo_sk is null then 1 else 0 end) no_promo,
          sum(case when p_promo_sk is not null then 1 else 0 end) promo,
          count(*) total_cnt
    from inventory
    join catalog_sales on (cs_item_sk = inv_item_sk)
    join warehouse on (w_warehouse_sk=inv_warehouse_sk)
    join item on (i_item_sk = cs_item_sk)
    join customer_demographics on (cs_bill_cdemo_sk = cd_demo_sk)
    join household_demographics on (cs_bill_hdemo_sk = hd_demo_sk)
    join date_dim d1 on (cs_sold_date_sk = d1.d_date_sk)
    join date_dim d2 on (inv_date_sk = d2.d_date_sk)
    join date_dim d3 on (cs_ship_date_sk = d3.d_date_sk)
    left outer join promotion on (cs_promo_sk=p_promo_sk)
    join catalog_returns on (cr_item_sk = cs_item_sk and cr_order_number = cs_order_number)
    where d1.d_week_seq = d2.d_week_seq
      and inv_quantity_on_hand < cs_quantity
      and d3.d_date > (cast(d1.d_date AS DATE) + interval '5' day)
      and hd_buy_potential = '>10000'
      and d1.d_year = 1999
      and cd_marital_status = 'D'
    group by i_item_desc,w_warehouse_name,d1.d_week_seq
    order by total_cnt desc, i_item_desc, w_warehouse_name, d1.d_week_seq
    limit 100;

2. 查看Profile。

   查看HASH_JOIN的節(jié)點信息，可以看到右表的數(shù)據(jù)量明顯大于左表，說明該Join Plan異常。

3. 修改join catalog_returns方式為left outer join。

   select i_item_desc,
          w_warehouse_name,
          d1.d_week_seq,
          sum(case when p_promo_sk is null then 1 else 0 end) no_promo,
          sum(case when p_promo_sk is not null then 1 else 0 end) promo,
          count(*) total_cnt
    from inventory
    join catalog_sales on (cs_item_sk = inv_item_sk)
    join warehouse on (w_warehouse_sk=inv_warehouse_sk)
    join item on (i_item_sk = cs_item_sk)
    join customer_demographics on (cs_bill_cdemo_sk = cd_demo_sk)
    join household_demographics on (cs_bill_hdemo_sk = hd_demo_sk)
    join date_dim d1 on (cs_sold_date_sk = d1.d_date_sk)
    join date_dim d2 on (inv_date_sk = d2.d_date_sk)
    join date_dim d3 on (cs_ship_date_sk = d3.d_date_sk)
    left outer join promotion on (cs_promo_sk=p_promo_sk)
    left outer join catalog_returns on (cr_item_sk = cs_item_sk and cr_order_number = cs_order_number)
    where d1.d_week_seq = d2.d_week_seq
      and inv_quantity_on_hand < cs_quantity
      and d3.d_date > (cast(d1.d_date AS DATE) + interval '5' day)
      and hd_buy_potential = '>10000'
      and d1.d_year = 1999
      and cd_marital_status = 'D'
    group by i_item_desc,w_warehouse_name,d1.d_week_seq
    order by total_cnt desc, i_item_desc, w_warehouse_name, d1.d_week_seq
    limit 100;

4. 再次查看Profile，可以看到性能已得到優(yōu)化。

確認JoinRuntimeFilter是否生效

當Join的右表構建Hash Table時，會構建Runtime Filter，該Runtime Filter會被投遞到左子樹，并盡可能地下推到Scan Operator。您可以在Scan Operator的節(jié)點詳情選項卡上查看與JoinRuntimeFilter相關的指標。

1. 以TPC_DH中的query72.sql為例，。

   select i_item_desc,
          w_warehouse_name,
          d1.d_week_seq,
          sum(case when p_promo_sk is null then 1 else 0 end) no_promo,
          sum(case when p_promo_sk is not null then 1 else 0 end) promo,
          count(*) total_cnt
    from inventory
    join catalog_sales on (cs_item_sk = inv_item_sk)
    join warehouse on (w_warehouse_sk=inv_warehouse_sk)
    join item on (i_item_sk = cs_item_sk)
    join customer_demographics on (cs_bill_cdemo_sk = cd_demo_sk)
    join household_demographics on (cs_bill_hdemo_sk = hd_demo_sk)
    join date_dim d1 on (cs_sold_date_sk = d1.d_date_sk)
    join date_dim d2 on (inv_date_sk = d2.d_date_sk)
    join date_dim d3 on (cs_ship_date_sk = d3.d_date_sk)
    left outer join promotion on (cs_promo_sk=p_promo_sk)
    left outer join catalog_returns on (cr_item_sk = cs_item_sk and cr_order_number = cs_order_number)
    where d1.d_week_seq = d2.d_week_seq
      and inv_quantity_on_hand < cs_quantity
      and d3.d_date > (cast(d1.d_date AS DATE) + interval '5' day)
      and hd_buy_potential = '>10000'
      and d1.d_year = 1999
      and cd_marital_status = 'D'
    group by i_item_desc,w_warehouse_name,d1.d_week_seq
    order by total_cnt desc, i_item_desc, w_warehouse_name, d1.d_week_seq
    limit 100;

2. 查看Profile。

   單擊OLAP_SCAN，單擊右側的節(jié)點詳情頁簽，可以看到在Scan Operator算子掃描inventory表時觸發(fā)了JoinRuntimeFilter。

Colocate Join

在StarRocks中使用Colocate Join功能，需要在建表時為其指定一個Colocation Group（CG），同一CG內的表需遵循相同的Colocation Group Schema（CGS），可以保證同一CG內所有表的數(shù)據(jù)分布在相同一組BE節(jié)點上。當Join列為分桶鍵時，計算節(jié)點只需做本地Join，從而減少數(shù)據(jù)在節(jié)點間的傳輸耗時，提高查詢性能。因此，Colocate Join相對于其他Join，例如Shuffle Join和Broadcast Join，可以有效避免數(shù)據(jù)網(wǎng)絡傳輸開銷，提高查詢性能。

創(chuàng)建Colocation表

CREATE TABLE tbl (k1 int, v1 int sum)
DISTRIBUTED BY HASH(k1)
BUCKETS 8
PROPERTIES(
    "colocate_with" = "group1"
);

刪除Colocation Group

當Group中最后一張表徹底刪除后，該Group也會被自動刪除。徹底刪除是指從回收站中刪除。通常，一張表通過DROP TABLE命令被刪除后，會在回收站默認停留一天的時間后，再被徹底刪除。

查看Group信息

SHOW PROC '/colocation_group';

+-------------+--------------+----------+------------+----------------+----------+----------+
| GroupId     | GroupName    | TableIds | BucketsNum | ReplicationNum | DistCols | IsStable |
+-------------+--------------+----------+------------+----------------+----------+----------+
| 11912.11916 | 11912_group1 | 11914    | 8          | 3              | int(11)  | true     |
+-------------+--------------+----------+------------+----------------+----------+----------+

SHOW PROC '/colocation_group/GroupId';

SHOW PROC '/colocation_group/11912.11916';

+-------------+---------------------+
| BucketIndex | BackendIds          |
+-------------+---------------------+
| 0           | 10002, 10004, 10003 |
| 1           | 10002, 10004, 10003 |
| 2           | 10002, 10004, 10003 |
| 3           | 10002, 10004, 10003 |
| 4           | 10002, 10004, 10003 |
| 5           | 10002, 10004, 10003 |
| 6           | 10002, 10004, 10003 |
| 7           | 10002, 10004, 10003 |
+-------------+---------------------+
8 rows in set (0.00 sec)

修改Group屬性

ALTER TABLE tbl SET ("colocate_with" = "group_name");

1. 以TPC_H中數(shù)據(jù)為例，執(zhí)行以下操作。

   use tpc_h_sf100;
   ALTER TABLE orders SET ("colocate_with" = "cg_tpc_orders");
   ALTER TABLE lineitem SET ("colocate_with" = "cg_tpc_orders");

2. 執(zhí)行以下Query查詢。

   select count(1) from orders as o join lineitem as l on o.o_orderkey = l.l_orderkey;

3. 您可以在EMR StarRocks Manager頁面查詢的查詢計劃頁簽，查看Colocate Join是否生效。

   當colocate顯示為true時，表示Colocate Join生效。

確認分桶或分區(qū)裁剪是否生效