Richard Zowalla created OPENJPA-3010:
----------------------------------------
Summary: Bulk delete and map key tests assert Derby-shaped SQL and
fail on MySQL
Key: OPENJPA-3010
URL: https://issues.apache.org/jira/browse/OPENJPA-3010
Project: OpenJPA
Issue Type: Bug
Components: jpa
Reporter: Richard Zowalla
Six tests added by the OPENJPA-2990 / OPENJPA-2994 / OPENJPA-2965 work assert
SQL shapes that only hold on Derby and H2, so a full MySQL run fails. Reported
by Maxim Solodovnik on the dev list while running the MySQL suite.
Reproduced on master (b4bfa57f0) against mysql:8.0 via -Dtest-mysql-docker, 6
failures:
TestBulkDeleteJoinTable.testBulkDeleteWithoutCriteriaEmptiesOwnedTables
TestBulkDeleteSubclassJoinTable.testBulkDeleteThroughBaseClass
TestBulkDeleteCompositeId.testCompositeKeyCandidateWithoutCriteria
TestBulkDeleteDependentFields.testUnfilteredDeleteCleansOwnedTablesOfDependentFields
TestBiDirectionalJoinTable.testBulkDelete
TestContainerSpecCompatibilityOptions.testLegacyDefaultMapKeyColumnName
Two distinct causes.
1. The criteria-less bulk delete renders differently on MySQL. The assertions
expect "DELETE FROM <table>" with an optional alias appended, i.e. the regular
expression "DELETE FROM .*BULK_OWNER( t[0-9]+)?", while MySQL emits the alias
before the FROM:
DELETE t0 FROM BULK_OWNER t0
Derby and H2 emit "DELETE FROM BULK_OWNER" and "DELETE FROM BULK_OWNER t0"
respectively, which is what the patterns were written against. Five of the six
failures are this.
2. The legacy map key column name is dictionary specific.
testLegacyDefaultMapKeyColumnName asserts the generated DDL contains KEY0,
which is what Derby produces because KEY is in its invalid-column-word set and
OpenJPA renames it. MySQL instead delimits the reserved word, so the DDL
contains a delimited lower case key:
CREATE TABLE Uni_Map_KeyCol_attrs (UniMapKeyCol_ID BIGINT, `key`
VARCHAR(255), `value` VARCHAR(255)) ENGINE = innodb
Both are test defects, not product defects: the product behaviour is correct on
each database, the assertions are written for one dialect. Fix the assertions
to be dictionary agnostic (tolerate the MySQL delete form, and derive the
expected column name from the dictionary rather than hard coding KEY0).
Note why CI did not catch it: the GitHub Actions matrix runs only the
test-derby and test-h2 profiles, so no MySQL, MariaDB or PostgreSQL run happens
on a pull request. Any assertion on emitted SQL therefore needs to be checked
against the other dictionaries by hand before merging, or the matrix needs
extending.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
