[
https://issues.apache.org/jira/browse/DRILL-7117?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16799742#comment-16799742
]
Aman Sinha edited comment on DRILL-7117 at 3/23/19 6:16 PM:
------------------------------------------------------------
I have a work-in-progress implementation and initial results for the histogram
using the proposed design (see design doc in parent JIRA) are encouraging. I
analyzed a few columns from the TPC-DS table Web_Sales and in particular for a
skewed column ws_ext_tax I compared the output of the NTILE() function with
what is generated by the histogram implementation. They are very close.
Here's the NTILE() output (this is the baseline):
{noformat}
select x as ntile, min(ws_ext_tax) as min, max(ws_ext_tax) as max, count(*) as
count
from (select ws_ext_tax, ntile(10) over (order by ws_ext_tax) as x from
dfs.tmp.web_sales1)
group by x order by x;
|ntile | min | max | count |
| 1 | 0.0 | 0.0 | 71939 |
| 2 | 0.0 | 3.05 | 71939 |
| 3 | 3.05 | 9.28 | 71939 |
| 4 | 9.28 | 19.24 | 71939 |
| 5 | 19.24 | 34.51 | 71938 |
| 6 | 34.51 | 57.96 | 71938 |
| 7 | 57.96 | 95.01 | 71938 |
| 8 | 95.01 | 158.71 | 71938 |
| 9 | 158.71 | 292.44 | 71938 |
| 10 | 292.44 | 2433.41 | 71938 |
{noformat}
Here's the Histogram output after running ANALYZE and using 10 buckets:
{noformat}
{
"column" : "`ws_ext_tax`",
"majortype" : {
"type" : "FLOAT4",
"mode" : "OPTIONAL"
},
"schema" : 1.0,
"rowcount" : 719384.0,
"nonnullrowcount" : 719200.0,
"ndv" : 68400,
"avgwidth" : 4.0,
"histogram" : {
"category" : "numeric-histogram",
"numRowsPerBucket" : 71920,
"buckets" : [ 0.0, 0.0, 3.0012382450412556, 9.294927927208075,
19.451534502867155, 34.26953338960548, 57.30967536502495, 95.25307998787451,
158.88263313484484, 289.2295684860717, 24
33.409910105169 ]
}
}
{noformat}
Note that the bucket boundaries closely align with the baseline and there's
non-uniformity in the data distribution.
was (Author: amansinha100):
I have a work-in-progress implementation and initial results for the histogram
using the proposed design (see design doc in parent JIRA) are encouraging. I
analyzed a few columns from the TPC-DS table Web_Sales and in particular for a
skewed column ws_ext_tax I compared the output of the NTILE() function with
what is generated by the histogram implementation. They are very close.
Here's the NTILE() output (this is the baseline):
{noformat}
||Heading 1||Heading 2||
|Col A1|Col A2|
select x as ntile, min(ws_ext_tax) as min, max(ws_ext_tax) as max, count(*) as
count
from (select ws_ext_tax, ntile(10) over (order by ws_ext_tax) as x from
dfs.tmp.web_sales1)
group by x order by x;
||ntile || min || max || count ||
| 1 | 0.0 | 0.0 | 71939 |
| 2 | 0.0 | 3.05 | 71939 |
| 3 | 3.05 | 9.28 | 71939 |
| 4 | 9.28 | 19.24 | 71939 |
| 5 | 19.24 | 34.51 | 71938 |
| 6 | 34.51 | 57.96 | 71938 |
| 7 | 57.96 | 95.01 | 71938 |
| 8 | 95.01 | 158.71 | 71938 |
| 9 | 158.71 | 292.44 | 71938 |
| 10 | 292.44 | 2433.41 | 71938 |
{noformat}
Here's the Histogram output after running ANALYZE and using 10 buckets:
{noformat}
{
"column" : "`ws_ext_tax`",
"majortype" : {
"type" : "FLOAT4",
"mode" : "OPTIONAL"
},
"schema" : 1.0,
"rowcount" : 719384.0,
"nonnullrowcount" : 719200.0,
"ndv" : 68400,
"avgwidth" : 4.0,
"histogram" : {
"category" : "numeric-histogram",
"numRowsPerBucket" : 71920,
"buckets" : [ 0.0, 0.0, 3.0012382450412556, 9.294927927208075,
19.451534502867155, 34.26953338960548, 57.30967536502495, 95.25307998787451,
158.88263313484484, 289.2295684860717, 24
33.409910105169 ]
}
}
{noformat}
Note that the bucket boundaries closely align with the baseline and there's
non-uniformity in the data distribution.
> Support creation of histograms for numeric data types (except Decimal)
> ----------------------------------------------------------------------
>
> Key: DRILL-7117
> URL: https://issues.apache.org/jira/browse/DRILL-7117
> Project: Apache Drill
> Issue Type: Sub-task
> Components: Query Planning & Optimization
> Reporter: Aman Sinha
> Assignee: Aman Sinha
> Priority: Major
> Fix For: 1.16.0
>
>
> This JIRA is specific to creating histograms for numeric data types: INT,
> BIGINT, FLOAT4, FLOAT8 and their corresponding nullable/non-nullable
> versions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
