[DRILL-3269] Window function query takes too long to complete and return results

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 1.0.0
• Fix Version/s: 1.1.0
• Component/s: Execution - Flow
• Labels: window_function

Environment

1de6aed93efce8a524964371d96673b8ef192d89

• Assignee: Abdel Hakim Deneche
• Reporter: Khurram Faraaz
• Created: 10/Jun/15 05:52
• Updated: 15/Jun/15 23:41
• Resolved: 15/Jun/15 23:41

Description

Query that uses window functions takes too long to complete and return results. It returns close to a million records, for which it took 533.8 seconds ~8 minutes
Input CSV file has two columns, one integer and another varchar type column. Please take a look.

Size of the input CSV file
root@centos-01 ~]# hadoop fs -ls /tmp/manyDuplicates.csv
-rwxr-xr-x 3 root root 27889455 2015-06-10 01:26 /tmp/manyDuplicates.csv

select count(*) over(partition by cast(columns[1] as varchar(25)) order by cast(columns[0] as bigint)) from `manyDuplicates.csv`;
...
1,000,007 rows selected (533.857 seconds)

There are five distinct values in columns[1] in the CSV file. = [FIVE PARTITIONS]

0: jdbc:drill:schema=dfs.tmp> select distinct columns[1] from `manyDuplicates.csv`;
+-----------------------+
|        EXPR$0         |
+-----------------------+
| FFFFGGGGHHHHIIIIJJJJ  |
| PPPPQQQQRRRRSSSSTTTT  |
| AAAABBBBCCCCDDDDEEEE  |
| UUUUVVVVWWWWXXXXZZZZ  |
| KKKKLLLLMMMMNNNNOOOO  |
+-----------------------+
5 rows selected (1.906 seconds)

Here is the count for each of those values in columns[1]

0: jdbc:drill:schema=dfs.tmp> select count(columns[1]) from `manyDuplicates.csv` where columns[1] = 'FFFFGGGGHHHHIIIIJJJJ';
+---------+
| EXPR$0  |
+---------+
| 200484  |
+---------+
1 row selected (0.961 seconds)

0: jdbc:drill:schema=dfs.tmp> select count(columns[1]) from `manyDuplicates.csv` where columns[1] = 'PPPPQQQQRRRRSSSSTTTT';
+---------+
| EXPR$0  |
+---------+
| 199353  |
+---------+
1 row selected (0.86 seconds)

0: jdbc:drill:schema=dfs.tmp> select count(columns[1]) from `manyDuplicates.csv` where columns[1] = 'AAAABBBBCCCCDDDDEEEE';
+---------+
| EXPR$0  |
+---------+
| 200702  |
+---------+
1 row selected (0.826 seconds)

0: jdbc:drill:schema=dfs.tmp> select count(columns[1]) from `manyDuplicates.csv` where columns[1] = 'UUUUVVVVWWWWXXXXZZZZ';
+---------+
| EXPR$0  |
+---------+
| 199916  |
+---------+
1 row selected (0.851 seconds)

0: jdbc:drill:schema=dfs.tmp> select count(columns[1]) from `manyDuplicates.csv` where columns[1] = 'KKKKLLLLMMMMNNNNOOOO';
+---------+
| EXPR$0  |
+---------+
| 199552  |
+---------+
1 row selected (0.827 seconds)

Query plan for the long running query

| 00-00    Screen
00-01      UnionExchange
01-01        Project(EXPR$0=[$0])
01-02          Project($0=[$2])
01-03            Window(window#0=[window(partition {1} order by [0] range between UNBOUNDED PRECEDING and CURRENT ROW aggs [COUNT()])])
01-04              SelectionVectorRemover
01-05                Sort(sort0=[$1], sort1=[$0], dir0=[ASC], dir1=[ASC])
01-06                  Project($0=[$0], $1=[$1])
01-07                    HashToRandomExchange(dist0=[[$1]])
02-01                      UnorderedMuxExchange
03-01                        Project($0=[$0], $1=[$1], E_X_P_R_H_A_S_H_F_I_E_L_D=[castInt(hash64AsDouble($1))])
03-02                          Project($0=[CAST(ITEM($0, 0)):BIGINT], $1=[CAST(ITEM($0, 1)):VARCHAR(25) CHARACTER SET "ISO-8859-1" COLLATE "ISO-8859-1$en_US$primary"])
03-03                            Scan(groupscan=[EasyGroupScan [selectionRoot=/tmp/manyDuplicates.csv, numFiles=1, columns=[`columns`[0], `columns`[1]], files=[maprfs:///tmp/manyDuplicates.csv]]])

python script to generate data in CSV format

import random
f = open('/Users/kungfo/manyDuplicates.csv', 'a')
for i in range(1,000000):
    f.write(str(random.choice(xrange(1,1000000)))+','+str(random.choice(['AAAABBBBCCCCDDDDEEEE','FFFFGGGGHHHHIIIIJJJJ','KKKKLLLLMMMMNNNNOOOO','PPPPQQQQRRRRSSSSTTTT','UUUUVVVVWWWWXXXXZZZZ']))+'\n')
    f.flush()