[
https://issues.apache.org/jira/browse/PHOENIX-4690?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16440461#comment-16440461
]
Advertising
chenglei edited comment on PHOENIX-4690 at 4/17/18 6:28 AM:
------------------------------------------------------------
[~jamestaylor] , thank you for the review.
bq.Orthogonal to this, but it'd be good if we had a way to know if it's better
to have the GROUP BY be order preserving followed by a client-side sort, or for
the GROUP BY to be non order preserving with a merge sort on the client. I
think we need histograms to make that determination.
In my opinion, I think GROUP BY be order preserving is always better than GROUP
BY to be non order preserving,because if GROUP BY is non order preserving, in
additional to a merge sort on the client, we also need a client-side sort for
the server-side hash aggregated results before the merge sort.
was (Author: comnetwork):
[~jamestaylor] , thank you for the review.
bq.Orthogonal to this, but it'd be good if we had a way to know if it's better
to have the GROUP BY be order preserving followed by a client-side sort, or for
the GROUP BY to be non order preserving with a merge sort on the client. I
think we need histograms to make that determination.
In my opinion, I think GROUP BY be order preserving is always better than GROUP
BY to be non order preserving,because if GROUP BY is non order preserving, in
additional to a merge sort on the client, we need a client-side sort on the
server-side hash aggregated results before the merge sort.
> GroupBy expressions should follow the order of PK Columns if GroupBy is
> orderPreserving
> ---------------------------------------------------------------------------------------
>
> Key: PHOENIX-4690
> URL: https://issues.apache.org/jira/browse/PHOENIX-4690
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.14.0, 4.13.2
> Reporter: chenglei
> Assignee: chenglei
> Priority: Critical
> Fix For: 4.14.0, 5.0.0
>
> Attachments: PHOENIX-4690_v1.patch
>
>
> Given a table :
> {code}
> create table test (
> pk1 integer not null ,
> pk2 integer not null,
> v integer,
> CONSTRAINT TEST_PK PRIMARY KEY (pk1,pk2))
> {code}
> and some data:
> {code}
> +------+------+-----+
> | PK1 | PK2 | V |
> +------+------+-----+
> | 1 | 8 | 10 |
> | 1 | 9 | 11 |
> | 2 | 3 | 13 |
> | 2 | 7 | 15 |
> | 3 | 2 | 17 |
> +------+------+-----+
> {code}
> for following sql :
> {code}
> select pk2,pk1,count(v) from test group by pk2,pk1 order by pk2,pk1
> {code}
> the expected result is :
> {code}
> +------+------+-----------+
> | PK2 | PK1 | COUNT(V) |
> +------+------+-----------+
> | 2 | 3 | 1 |
> | 3 | 2 | 1 |
> | 7 | 2 | 1 |
> | 8 | 1 | 1 |
> | 9 | 1 | 1 |
> +------+------+-----------+
> {code}
> but the actual result is :
> {code}
> +------+------+-----------+
> | PK2 | PK1 | COUNT(V) |
> +------+------+-----------+
> | 8 | 1 | 1 |
> | 9 | 1 | 1 |
> | 3 | 2 | 1 |
> | 7 | 2 | 1 |
> | 2 | 3 | 1 |
> +------+------+-----------+
> {code}
> The problem is caused by the {{GroupBy.compile}}, obviously, in line 154, for
> {{group by pk2,pk1}},
> {{OrderPreservingTracker.isOrderPreserving}} is true by reorder the
> {{pk2,pk1}} to {{pk1,pk2}} following the order of PK columns {{pk1,pk2}},
> but in line 158, for the new GroupBy, the GroupBy.expressions is still
> {{pk2,pk1}},not the actual {{pk1,pk2}}:
> {code}
> 141 public GroupBy compile(StatementContext context, TupleProjector
> tupleProjector) throws SQLException {
> 142 boolean isOrderPreserving = this.isOrderPreserving;
> 143 int orderPreservingColumnCount = 0;
> 144 if (isOrderPreserving) {
> 145 OrderPreservingTracker tracker = new
> OrderPreservingTracker(context, GroupBy.EMPTY_GROUP_BY, Ordering.UNORDERED,
> expressions.size(), tupleProjector);
> 146 for (int i = 0; i < expressions.size(); i++) {
> 147 Expression expression = expressions.get(i);
> 148 tracker.track(expression);
> 149 }
> 150
> 151 // This is true if the GROUP BY is composed of only PK
> columns. We further check here that
> 152 // there are no "gaps" in the PK columns positions used
> (i.e. we start with the first PK
> 153 // column and use each subsequent one in PK order).
> 154 isOrderPreserving = tracker.isOrderPreserving();
> 155 orderPreservingColumnCount =
> tracker.getOrderPreservingColumnCount();
> 156 }
> 157 if (isOrderPreserving || isUngroupedAggregate) {
> 158 return new
> GroupBy.GroupByBuilder(this).setIsOrderPreserving(isOrderPreserving)
> .setOrderPreservingColumnCount(orderPreservingColumnCount).build();
> 159 }
> {code}
> Then when we compile {{order by pk2,pk1}} in {{OrderByCompiler.compile}},
> because {{the GroupBy.isOrderPreserving}} is true, and {{order by pk2,pk1}}
> is consistent with the GroupBy.expressions {{group by pk2,pk1}} created in
> above {{GroupBy.compile}} method, so the result of
> {{OrderByCompiler.compile}} is {{OrderBy.FWD_ROW_KEY_ORDER_BY}}.
> But in fact,because the actual GroupBy.expressions is {{group by pk1,pk2}},so
> we need to execute {{order by pk2,pk1}} after the {{group by pk1,pk2}}.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
