[
https://issues.apache.org/jira/browse/CALCITE-3820?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ruben Q L updated CALCITE-3820:
-------------------------------
Description:
Currently, {{EnumerableDefaults#orderBy}} (which is the actual implementation
behind EnumerableSort) looks like this:
{code:java}
public static <TSource, TKey> Enumerable<TSource> orderBy(
Enumerable<TSource> source, Function1<TSource, TKey> keySelector,
Comparator<TKey> comparator) {
final Map<TKey, List<TSource>> map = new TreeMap<>(comparator);
LookupImpl<TKey, TSource> lookup = toLookup_(map, source, keySelector,
Functions.identitySelector());
return lookup.valuesEnumerable();
}
{code}
This implementation has the following disadvantages:
- 1) The TreeMap that performs the actual sorting is eagerly computed. This
operation can be quite heavy if we have a very big source. The fact that it is
eagerly computed might lead to bad performance in situations where the sort is
executed, but it is actually not needed. For example, in a NestedLoopJoin, if
the outer enumerator returns no results, the inner one is not even accessed. If
we had a huge orderBy as inner, today it would be computed in that scenario,
even though it is not actually required. For this reason, in terms of
performance it seems clearly preferable to delay the sort operation as much as
possible.
- 2) The Enumerable / Enumerator returned by {{EnumerableDefaults#orderBy}}
cannot be re-evaluated. Since the map, and the subsequent LookupImpl, the
Enumerable relies on is eagerly computed, every called to {{enumerator}} will
return the same values, *even if the source Enumerable has changed*. This is a
corner case, but we can see it if, for example, we have a MergeJoin, with
EnumerableSort (i.e. using {{EnumerableDefaults#orderBy}}) inside a RepeatUnion
(a.k.a. recursive union). In this situation, the RepeatUnion re-evaluates its
right child, so that the output of the iteration N-1 is used as input for the
iteration N. In this scenario, the {{EnumerableDefaults#orderBy}} will not work
as expected, since it will not be actually re-computed (see
{{EnumerableMethods#repeatUnion}})
was:
Currently, {{EnumerableDefaults#orderBy}} (which is the actual implementation
behind EnumerableSort) looks like this:
{code:java}
public static <TSource, TKey> Enumerable<TSource> orderBy(
Enumerable<TSource> source, Function1<TSource, TKey> keySelector,
Comparator<TKey> comparator) {
final Map<TKey, List<TSource>> map = new TreeMap<>(comparator);
LookupImpl<TKey, TSource> lookup = toLookup_(map, source, keySelector,
Functions.identitySelector());
return lookup.valuesEnumerable();
}
{code}
This implementation has the following disadvantages:
- 1) The TreeMap that performs the actual sorting is eagerly computed. This
operation can be quite heavy if we have a very big source. The fact that it is
eagerly computed might lead to bad performance in situations where the sort is
executed, but it is actually not needed. For example, in a NestedLoopJoin, if
the outer enumerator returns no results, the inner one is not even accessed. If
we had a huge orderBy as inner, today it would be computed in that scenario,
even though it is not actually required. For this reason, in terms of
performance it seems clearly preferable to delay the sort operation as much as
possible.
- 2) The Enumerable / Enumerator returned by {{EnumerableDefaults#orderBy}}
cannot be re-evaluated. Since the map, and the subsequent LookupImpl, the
Enumerable relies on is eagerly computed, every called to {{enumerator}} will
return the same values, *even if the source Enumerable has changed*. This is a
corner case, but we can see it if, for example, we have a MergeJoin, with
EnumerableSort (i.e. using {{EnumerableDefaults#orderBy}}) inside a RepeatUnion
(a.k.a. recursive union). In this situation, the RepeatUnion re-evaluates its
right child, so that the output of the iteration N-1 is used as input for the
iteration N. In this scenario, the {{EnumerableDefaults#orderBy}} will not work
as expected, since it will not be actually re-computed.
> EnumerableDefaults#orderBy should be lazily computed + support enumerator
> re-initialization
> -------------------------------------------------------------------------------------------
>
> Key: CALCITE-3820
> URL: https://issues.apache.org/jira/browse/CALCITE-3820
> Project: Calcite
> Issue Type: Improvement
> Components: core
> Affects Versions: 1.21.0
> Reporter: Ruben Q L
> Assignee: Ruben Q L
> Priority: Major
>
> Currently, {{EnumerableDefaults#orderBy}} (which is the actual implementation
> behind EnumerableSort) looks like this:
> {code:java}
> public static <TSource, TKey> Enumerable<TSource> orderBy(
> Enumerable<TSource> source, Function1<TSource, TKey> keySelector,
> Comparator<TKey> comparator) {
> final Map<TKey, List<TSource>> map = new TreeMap<>(comparator);
> LookupImpl<TKey, TSource> lookup = toLookup_(map, source, keySelector,
> Functions.identitySelector());
> return lookup.valuesEnumerable();
> }
> {code}
> This implementation has the following disadvantages:
> - 1) The TreeMap that performs the actual sorting is eagerly computed. This
> operation can be quite heavy if we have a very big source. The fact that it
> is eagerly computed might lead to bad performance in situations where the
> sort is executed, but it is actually not needed. For example, in a
> NestedLoopJoin, if the outer enumerator returns no results, the inner one is
> not even accessed. If we had a huge orderBy as inner, today it would be
> computed in that scenario, even though it is not actually required. For this
> reason, in terms of performance it seems clearly preferable to delay the sort
> operation as much as possible.
> - 2) The Enumerable / Enumerator returned by {{EnumerableDefaults#orderBy}}
> cannot be re-evaluated. Since the map, and the subsequent LookupImpl, the
> Enumerable relies on is eagerly computed, every called to {{enumerator}} will
> return the same values, *even if the source Enumerable has changed*. This is
> a corner case, but we can see it if, for example, we have a MergeJoin, with
> EnumerableSort (i.e. using {{EnumerableDefaults#orderBy}}) inside a
> RepeatUnion (a.k.a. recursive union). In this situation, the RepeatUnion
> re-evaluates its right child, so that the output of the iteration N-1 is used
> as input for the iteration N. In this scenario, the
> {{EnumerableDefaults#orderBy}} will not work as expected, since it will not
> be actually re-computed (see {{EnumerableMethods#repeatUnion}})
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
