[
https://issues.apache.org/jira/browse/CALCITE-7529?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18081038#comment-18081038
]
zzwqqq commented on CALCITE-7529:
---------------------------------
Hi [~mbudiu] Thanks for the references. I agree this is related to the
millisecond-based generated-code representation for TIME/TIMESTAMP.
This patch is narrower: it only preserves precision for values that are already
representable as RexLiteral values backed by TimeString/TimestampString. It
does not try to fix temporal operations such as EXTRACT, TIMESTAMPADD, or
TIMESTAMPDIFF.
Do you think this is a reasonable incremental fix, or should it wait for a
broader change to the runtime representation?
> RexExecutor constant reduction loses sub-millisecond precision for
> TIME/TIMESTAMP literals
> ------------------------------------------------------------------------------------------
>
> Key: CALCITE-7529
> URL: https://issues.apache.org/jira/browse/CALCITE-7529
> Project: Calcite
> Issue Type: Bug
> Components: core
> Reporter: zzwqqq
> Priority: Major
> Labels: pull-request-available
>
> When a custom RelDataTypeSystem allows datetime precision greater than
> Calcite's default maximum precision, RexExecutor constant reduction can lose
> sub-millisecond precision for TIME and TIMESTAMP values.
> Calcite's default RelDataTypeSystem limits TIME/TIMESTAMP precision to 3, but
> users can override RelDataTypeSystem#getMaxPrecision. With such a type
> system, RexBuilder can be configured to create TIME(6) and TIMESTAMP(6)
> literals backed by TimeString/TimestampString.
> However, RexExecutorImpl reduces constant expressions through generated Java
> code, where TIME and TIMESTAMP values are represented using millisecond-based
> runtime values. When RexExecutable.reduce rebuilds a RexLiteral from those
> values, digits beyond milliseconds are lost.
> Minimal reproduction:
> {code:java}
> @Test void testReduceTimeCastWithMicros() {
> final RelDataTypeFactory typeFactory =
> new JavaTypeFactoryImpl(
> new RelDataTypeSystemImpl() {
> @Override public int getMaxPrecision(SqlTypeName typeName) {
> switch (typeName) {
> case TIME:
> case TIMESTAMP:
> return 6;
> default:
> return super.getMaxPrecision(typeName);
> }
> }
> });
> final RexBuilder rexBuilder = new RexBuilder(typeFactory);
> final RexExecutorImpl executor =
> new RexExecutorImpl(
> DataContexts.of(
> ImmutableMap.of(
> DataContext.Variable.TIME_ZONE.camelName,
> TimeZone.getTimeZone("GMT"),
> DataContext.Variable.LOCALE.camelName, Locale.US)));
> final RexNode cast =
> rexBuilder.makeCast(
> typeFactory.createSqlType(SqlTypeName.TIME, 6),
> rexBuilder.makeLiteral("12:34:56.123456"));
> final List<RexNode> reducedValues = new ArrayList<>();
> executor.reduce(rexBuilder, ImmutableList.of(cast), reducedValues);
> assertThat(
> ((RexLiteral)
> reducedValues.get(0)).getValueAs(TimeString.class).toString(6),
> equalTo("12:34:56.123456"));
> }
> {code}
> Expected:
> {code}
> 12:34:56.123456
> {code}
> Actual before the fix:
> {code}
> 12:34:56.123000
> {code}
> The same issue can occur for TIMESTAMP(6). It can also affect already-created
> high precision TimeString/TimestampString RexLiterals if they are passed
> through RexExecutorImpl.reduce.
> A possible fix is to preserve high-precision temporal literal casts as Rex
> literals when they can be represented directly by TimeString/TimestampString,
> and to let RexExecutorImpl return existing RexLiteral values without
> compiling and executing them again.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
