Revision history for DBIO-Oracle

0.900007  2026-11-17
    * Storage
        - Numbers are bound and read with a decimal point in every NLS
          territory: a connection sets NLS_NUMERIC_CHARACTERS to '.,' for its
          session, so a NUMBER, BINARY_DOUBLE or BINARY_FLOAT with decimals
          makes the round trip under a client such as
          NLS_LANG=GERMAN_GERMANY.AL32UTF8, where a bound 1.5 was refused
          with ORA-01722. The environment variable of that name does not
          change it; an on_connect_do or on_connect_call of the application
          runs after and wins.
        - An error text carries no bound value whatever language the session
          reports in: the rules that replace the value (ORA-03302, ORA-03301,
          ORA-01722, ORA-01438, ORA-61800, ORA-01428) read the error code and
          the structure of the message, not the English sentence, so a short
          value no longer stands in the clear in a German message. Where the
          structure of a message is not known, the text after the code is
          replaced by ***; the code and the columns stay. French is read
          too (the French details line and help line of ORA-00001, the help
          label with a blank before the colon, and a link line that follows
          the language of the client and not of the session), so the
          columns of a key stay in a French environment, in an American
          session as well
          (docs/adr/0012-oracle-error-values-masking.md).
    * Introspect
        - The model lists its foreign keys of a table and the views a view
          rests on in the same order in every session language: names are
          read in binary order (NLSSORT with NLS_SORT=BINARY), not in the
          NLS_SORT of the session, under which names such as a_x, aB and Ab
          came in another order. The session keeps its own NLS_SORT.
    * Tests
        - The error masking is tested with the German texts of the server and
          with made-up texts of an unknown language; the live test runs every
          case in an English and in a German session of its own. The live
          tests that read an English sentence of the server put their
          connection on English messages.
        - The live tests that list the data dictionary sort names binary, so
          their expectations hold in a session with another NLS_SORT. The
          order of the introspection is pinned offline by the text of its
          ORDER BY clauses and live under NLS_SORT=GERMAN.
        - The temporary-user deploy test replaces its subs under
          no warnings 'redefine', so it passes on older perls too.
    * Documentation
        - The dbio-oracle and dbio-oracle-database skills state the LOB
          comparison parts as the code has them (text 1000 characters,
          binary 2000 bytes), list the capability declarations (transactional
          DDL, set operator ALL family, exists query, upsert) and the live
          test path.
    * Dependencies
        - Requires Math::Base36 >= 0.14; earlier versions warn "Negative
          repeat count does nothing" on every shortened identifier.

0.900004  2026-11-08
    * Storage
        - An error text names the error code, the column, the table and the
          constraint but no value of the statement: Oracle 23ai writes the
          value into ORA-01722 (the first character), ORA-03302, ORA-03301
          (the key of the row of ORA-00001), ORA-01438, ORA-61800 and
          ORA-01428, bound or not, and the storage replaces it with ***
          whatever its length, also in the errors of populate and
          insert_bulk. The statement in the message keeps its literals
          (docs/adr/0012-oracle-error-values-masking.md).
        - The DSN is built from connection parameters without a DSN
          (DBIO::AccessBroker::Static, a connect-info hash): without host
          dbi:Oracle:<dbname>, with host Easy Connect
          dbi:Oracle://<host>[:<port>][/<dbname>], dbname being the service
          name. port without host throws.
        - A DBIO::Replicated replicant (HostBound view) over an Easy Connect
          DSN connects to the view's host and port and keeps the service. An
          alias, a SID and a TNS descriptor are refused, naming the form.
        - A connection sets LongReadLen to 16 MiB (a CLOB or BLOB is read in
          full; a cut value is an error) and ora_charset and ora_ncharset to
          AL32UTF8, unless connect_info names them. The ISO NLS date and
          timestamp formats are the default of every connection, the async
          pool included.
        - A one-byte Perl string bound to a character column is upgraded to
          characters on a copy, so "\x{fc}" is stored as itself. A binary
          column and a bind without a data type are sent unchanged.
        - INSERT and UPDATE binding a BLOB or CLOB work on Oracle 23ai; an
          UPDATE returns the row count. A LOB is compared by value in parts
          that fit a SQL string (text, and binary as RAW), the empty value
          as an empty LOB. populate of a LOB column inserts row by row.
        - create({}) is INSERT ... (col) VALUES (DEFAULT). The sequence
          behind an auto-increment column is remembered by the storage, so a
          table dropped and created again in one process works.
        - DBD::Oracle is loaded where it is needed, so a box without the
          Oracle client loads the driver and fails at the first LOB bind.
        - transactional_ddl is 0: the first DDL statement commits the open
          transaction. DBIO::Oracle::Deploy documents the half-run apply.
        - A populate in void context turns PrintWarn off on its statement
          handle, so DBD::Oracle stays quiet on a partial failure.

    * SQL
        - $rs->except renders MINUS. except_all and intersect_all are refused
          at query-build time, naming the version boundary (before 21c);
          DBIO::Oracle::Storage->_use_set_operator_all_family(1) lifts it.
        - results_exist and its async twins work (FROM DUAL); $rs->rand()
          orders by DBMS_RANDOM.VALUE; $rs->explain runs EXPLAIN PLAN and
          SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY) on one connection and needs
          PLAN_TABLE visible to the session.
        - prior is a word operator, so { parentid => { -prior => $col } }
          passes core's check of operator names.
        - The DDL dialect is oracle: the semicolon after END of a CREATE
          TRIGGER or a PL/SQL block survives core's statement splitter.

    * DDL
        - Every identifier is double-quoted as the result class spells it.
          Unique constraints sit inside CREATE TABLE; index declarations of
          the result source are written as CREATE [UNIQUE] INDEX, the oracle
          block taking only expression (function-based index).
        - An auto-increment column gets its sequence and the BEFORE INSERT
          trigger; belongs_to becomes a FOREIGN KEY with on_delete, and
          is_deferrable. Oracle has no ON UPDATE; one other than NO ACTION or
          RESTRICT is refused.
        - A missing is_nullable deploys as NOT NULL. DEFAULT is written before
          NOT NULL; a default_value is a SQL literal.
        - A table named schema.table is created in that schema, with its
          sequence, trigger, indexes and foreign keys (see Deploy).
        - What Oracle cannot hold is refused with a DBIO exception naming
          table and column before any SQL: enum, set and bit columns, names
          with a double quote or NUL, an index declaration that meets a unique
          constraint, unsupported actions (see DBIO::Oracle::DDL, REFUSALS).
          declared_objects names what an install writes.

    * Types
        - What a Result class declares, the DDL deploys and the introspection
          reads back is the declared type, and the second column diff is
          empty: sizes of VARCHAR2, NVARCHAR2, NCHAR, TIMESTAMP(n) and RAW,
          character semantics ("20 CHAR"), FLOAT(n), TIMESTAMP WITH LOCAL TIME
          ZONE, INTERVAL YEAR(n) TO MONTH; integer types are INTEGER, double
          is BINARY_DOUBLE, long is CLOB.
        - The column diff does not report what the server cannot keep apart
          (varchar and varchar2, default precisions, every spelling of
          "20 CHAR"); CHAR against BYTE stays a change.

    * Introspect
        - A live Oracle schema is read: rows by lower-case keys, foreign keys
          of several columns paired by position, long defaults and trigger
          bodies. Key columns report is_nullable => 0, and primary key
          columns carry is_pk and pk_position.
        - IDENTITY columns and columns filled by the install's trigger are
          is_auto_increment. A string default is read back as declared.
        - The attribute schemas reads further schemas; their tables are in
          the model as schema.table. The foreign-key name is constraint_name.
        - A view carries its definition (all_views.text) and its column names
          in the model, a view's dependencies in depends_on.

    * Deploy
        - install, diff, upgrade and apply work. diff and upgrade deploy the
          desired state into a temporary Oracle user DBIO_TMP_<pid>_<n>,
          introspect it there and drop it, also on failure; the connected
          user needs CREATE USER and DROP USER. A table in another schema
          gets a temporary user of its own. See DBIO::Oracle::Deploy.
        - install refuses a table in a schema the server does not have,
          naming user and table; it never creates a user.
        - apply and upgrade come from DBIO::Deploy::Base. A diff that drops
          something or may alter stored values runs only with
          allow_destructive => 1; the rows of the live tables are checked
          first (NOT NULL column without default on a table with rows, NOT
          NULL over a NULL, a datatype Oracle does not convert).
        - $schema->deploy installs through DBIO::Oracle::Deploy alone.

    * Diff
        - Column changes by size, type or default are written (removing a
          default is DEFAULT NULL); a changed index is dropped and created.
          Column statements quote table and column.
        - Foreign keys and unique constraints of a table both models have
          are compared and written (DROP CONSTRAINT, ADD CONSTRAINT); a
          foreign key over a unique constraint the target changes is
          re-created, one with nothing to point to is refused before the
          first statement. on_delete RESTRICT is NO ACTION.
        - Views are created or replaced (CREATE OR REPLACE VIEW with the
          column list) and dropped, ordered by their dependencies; a cycle or
          a view without text is refused before the first statement.
        - A table only the target has is created whole as the install writes
          it. DROP TABLE drops the install's sequences, DROP COLUMN cascades
          constraints. A renamed column is dropped and added again.

    * Test support
        - DBIO::Oracle::Test::Conformance answers the shared DDL, Diff and
          Bind case tables of core; deviations and refusals are declared in
          the class. The cleanup drops only what a case created.

    * Dependencies
        - Requires DBIO >= 0.900004.

0.900001  2026-07-12
    * Fixes
        - SQLMaker's _quote() now allowlists Oracle's hierarchical/row
          pseudo-columns (LEVEL, ROWNUM, ROWID, CONNECT_BY_ISCYCLE,
          CONNECT_BY_ISLEAF) and emits them bare instead of quoting them;
          quoting any of them raises ORA-00904 ("invalid identifier") now
          that identifier quoting defaults to on. Caught both as a bare
          string and as the single-segment arrayref shape SQL::Abstract v2
          uses for an already-split identifier (e.g. order_by).
          CONNECT_BY_ROOT is a prefix operator, not a standalone
          identifier, so it never reaches _quote() and is intentionally
          not in the allowlist. karr #17.
        - Type::is_lob_type / is_text_lob_type took a Data::Type object,
          but LOBSupport's _prep_for_execute passes a plain data-type
          string (matching sqlt_datatype), so the predicate silently
          returned false for every bind and LOB bind attrs were never
          applied. Now takes the string, mirroring DBIx::Class::Storage::
          DBI::_is_lob_type semantics.
        - dump_value's diagnostic in the "Partially failed populate
          throws" test pinned Data::Dumper's exact quoting form
          (name => "pop_art_1"), coupling a behavioural assertion (which
          row failed) to a serialization detail. Relaxed to assert the
          diagnostic header by pattern and the dumped failing row
          structurally (eval'd back to a hashref and compared with
          is_deeply), so either quoting form passes. ADR core/0027,
          karr #18.

    * Storage
        - DBIO::Oracle::Test::SequenceTest switched from
          is_auto_increment => sequence => 'name' to auto_nextval => 1
          with the sequence as a scalar ref, matching the API documented
          in DBIO::Oracle::Storage.
        - Added a DBIO::PK::Auto::Oracle deprecated compat shim: the
          generic DBIO::PK::Auto component now covers Oracle auto-PK
          handling directly, but load_components('PK::Auto::Oracle')
          still needs a class to exist to load. The shim exists only to
          satisfy that call and pulls in the generic component; new code
          should use DBIO::PK::Auto directly. karr #16.

    * Tests
        - t/10/t/12/t/20/t/22/t/32: DDL and literal identifiers used in
          the test fixtures are now quoted to match identifier quoting
          being on by default (the unquoted `cd` table stays unquoted by
          design, per its table(\'cd') declaration); the LEVEL order_by
          expectation is corrected to bare LEVEL, with a new assertion
          that LEVEL/ROWNUM stay bare under both quote_char states.

    * Documentation
        - Added SECURITY.md (CPANSec author guidelines v1.5.0): email-first
          private reporting with CPANSec escalation; explicit notice that
          Codeberg has no confidential issue channel.

    * Dependencies
        - DBD::Oracle is now a hard requires (was recommends): a database-
          specific driver is useless without its DBD. Dropped the runtime
          require DBD::Oracle workaround in Type, LOBSupport, and
          Introspect::Columns in favour of plain use.
        - Requires core DBIO >= 0.900001 (family-wide version alignment)

0.900000  2026-06-23
    First release. Oracle database driver for DBIO.

    * Storage
        - Oracle storage auto-detected for dbi:Oracle: DSNs, with the -ora
          shortcut, load_components('Oracle'), and DBIO::Base subclassing
        - Sequence-based auto-increment (auto_nextval / SELECT seq.nextval)
        - LOB binding and chunking for CLOB/NCLOB/BLOB
        - Savepoints, deferrable FK constraint handling, and NLS session
          setup on connect
        - Pre-9.0 fallback to WHERE-clause joins when ANSI joins are
          unavailable

    * SQLMaker
        - ROWNUM-based limit/offset via the _RowNum dialect (apply_limit)
        - RETURNING ... INTO ? insert-returning syntax
        - CONNECT BY / START WITH / ORDER SIBLINGS BY / PRIOR hierarchical
          query support
        - Automatic identifier shortening to Oracle's 30-character limit
          across SELECT/FROM and WHERE/HAVING paths and relation aliases

    * Deploy
        - Desired-state deployment via test-deploy-and-compare
        - install_ddl creates a fresh schema; upgrade diffs live vs. desired

    * Introspect
        - Native introspection of tables, columns, keys, indexes, and
          foreign keys via USER_*/ALL_* views

    * Diff
        - Native diff of two introspected models at table, column, and
          index level

    * DDL
        - Native Oracle DDL generation from DBIO Result classes
        - Dedupe sources sharing a physical table, skip subquery and view
          sources, dereference literal table names, and emit CREATE VIEW
          for non-virtual views
