[
https://issues.apache.org/jira/browse/CALCITE-4419?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17242027#comment-17242027
]
Julian Hyde commented on CALCITE-4419:
--------------------------------------
A couple of minor things:
* I'd rename 'testPosixRegex' to 'checkPosixRegex'. (Conventionally only test
methods are prefixed with 'test'.)
* Consider removing the {{boolean harmonize}} argument to
{{PosixRegexMethodImplementor}} since it's always {{true}}.
+1 when you've done those. Nice work.
As a future change, consider introducing a Guava cache for compiled patterns.
Compiling the pattern once per row is probably extremely slow.
> Posix regex operators cannot be used within RelBuilder
> ------------------------------------------------------
>
> Key: CALCITE-4419
> URL: https://issues.apache.org/jira/browse/CALCITE-4419
> Project: Calcite
> Issue Type: Bug
> Components: core
> Reporter: Ruben Q L
> Assignee: Ruben Q L
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.27.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Posix regex operators are special, they are defined as binary ({{public class
> SqlPosixRegexOperator extends SqlBinaryOperator}}), but their actual
> implementation takes three arguments, the third being a caseSensitive boolean
> flag ({{SqlFunctions#posixRegex(String s, String regex, Boolean
> caseSensitive)}}).
> Using them in SQL works fine. The magic happens in
> {{SqlPosixRegexOperator#createCall}}, which adds an extra boolean parameter
> (the caseSensitive flag) into the call:
> {code:java}
> @Override public SqlCall createCall(SqlLiteral functionQualifier,
> SqlParserPos pos, SqlNode... operands) {
> pos = pos.plusAll(Arrays.asList(operands));
> operands = Arrays.copyOf(operands, operands.length + 1);
> operands[operands.length - 1] = SqlLiteral.createBoolean(caseSensitive,
> SqlParserPos.ZERO);
> return new SqlBasicCall(this, operands, pos, false, functionQualifier);
> }
> {code}
> However, if we try to use them in a plan created by RelBuilder, problems
> occur.
> 1) The following plan:
> {code:java}
> RelBuilder builder = ...
> builder
> .scan("s", "emps")
> .filter(
> builder.call(
> SqlStdOperatorTable.POSIX_REGEX_CASE_SENSITIVE,
> builder.field("name"),
> builder.literal("E...")))
> ...
> {code}
> Fails because it cannot find the implementor method
> {noformat}
> Caused by: java.lang.IllegalStateException: Unable to implement
> EnumerableCalc(expr#0..4=[{inputs}], expr#5=['E...'], expr#6=[POSIX REGEX
> CASE SENSITIVE($t2, $t5)]...
> ...
> Caused by: java.lang.NoSuchMethodException:
> org.apache.calcite.runtime.SqlFunctions.posixRegex(java.lang.String,
> java.lang.String)
> {noformat}
> 2) The solution seems either adding a (somehow redundant, but necessary due
> to posix regex operators design) third boolean parameter
> {code:java}
> .filter(
> builder.call(
> SqlStdOperatorTable.POSIX_REGEX_CASE_SENSITIVE,
> builder.field("name"),
> builder.literal("E..."),
> builder.literal(true)))
> {code}
> or maybe add this parameter transparently inside {{RelBuilder#call}} code:
> {code}
> private @Nonnull RexCall call(SqlOperator operator, List<RexNode>
> operandList) {
> switch (operator.getKind()) {
> ...
> case POSIX_REGEX_CASE_SENSITIVE:
> if (operandList.size() == 2) {
> return (RexCall) call(operator, operandList.get(0),
> operandList.get(1), literal(true));
> }
> break;
> case POSIX_REGEX_CASE_INSENSITIVE:
> if (operandList.size() == 2) {
> return (RexCall) call(operator, operandList.get(0),
> operandList.get(1), literal(false));
> }
> break;
> ...
> {code}
> Either way, it will now fail validating the operands:
> {noformat}
> wrong operand count 3 for POSIX REGEX CASE SENSITIVE
> java.lang.AssertionError: wrong operand count 3 for POSIX REGEX CASE SENSITIVE
> at org.apache.calcite.util.Litmus$1.fail(Litmus.java:31)
> at
> org.apache.calcite.sql.SqlBinaryOperator.validRexOperands(SqlBinaryOperator.java:200)
> {noformat}
> 3) So, it would seem {{SqlPosixRegexOperator}} should override
> {{validRexOperands}} to support 2 *and 3* parameters (which in fact would
> seem correct and aligned to the current overridden behavior of
> {{SqlPosixRegexOperator#getOperandCountRange}} and
> {{SqlPosixRegexOperator#checkOperandTypes}}:
> {code:java}
> @Override public boolean validRexOperands(int count, Litmus litmus) {
> if (count != 2 && count != 3) {
> return litmus.fail("wrong operand count {} for {}", count, this);
> }
> return litmus.succeed();
> }
> {code}
> But now the test fails with almost the same message as the very first one:
> {noformat}
> Caused by: java.lang.IllegalStateException: Unable to implement
> EnumerableCalc(expr#0..4=[{inputs}], expr#5=['E...'], expr#6=[true],
> expr#7=[POSIX REGEX CASE SENSITIVE($t2, $t5, $t6)]...
> ...
> Caused by: java.lang.NoSuchMethodException:
> org.apache.calcite.runtime.SqlFunctions.posixRegex(java.lang.String,
> java.lang.String, boolean)
> {noformat}
> 4) The problem seems to be a mismatch between our third argument (boolean)
> and the implementor's third argument (Boolean). Looking at the code, it seems
> clear that a primitive boolean should / must be used since having a null
> Boolean object makes no sense (and will cause a NPE anyway). So if we change
> it:
> {code:java}
> // SqlFunctions.java
> ...
> public static boolean posixRegex(String s, String regex, boolean
> caseSensitive) // instead of Boolean
> { ... }
> // ---------
> // BuiltInMethod.java
> ...
> POSIX_REGEX(SqlFunctions.class, "posixRegex", String.class, String.class,
> boolean.class), // instead of Boolean.class
> {code}
> Now our test finally succeeds.
> Therefore, I propose to implement these sequence of changes in order to solve
> this issue.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
