Skip to content

v1.11.5

Choose a tag to compare

@github-actions github-actions released this 24 Oct 13:00
· 227 commits to main since this release
d0748d1

This release fixes SQL compilation and connection-handling issues, improves inline-literal safety, and strengthens development and test workflows. Versions 1.11.1–1.11.4 were not published; upgrade directly from v1.11.0 and review the notes below.

Upgrade notes

  • Unpublished release candidates (1.11.1–1.11.4; #838, #910, #1005, #1011) — None has a tag, GitHub Release, or PyPI version. Their gates failed on Python 3.15, CUBRIDdb integration (#1003), cookbook verification, and make integration (#1035), respectively. All their changes are included here; upgrade directly from v1.11.0 to v1.11.5.

  • Development test and lint commands changed (#1009) — make test now runs all offline tests (previously make test-offline); the former fast suite is make test-fast. make lint checks Ruff lint only; use make format-check or make check for formatting. make check also runs type checking, security, and documentation checks. The old make test-offline and make check-all aliases still work (#1010) by @yeongseon.

  • Local Docker cleanup and broker access changed (#1008) — make docker-down now preserves test data; use make docker-clean to remove volumes (the old docker-down behavior). The broker binds to 127.0.0.1 instead of all host interfaces. CUBRID_TEST_PORT (default 33000) controls the host port for docker-up, Compose, and integration; Compose may use .env when not explicitly set. make integration still cleans up its own test project (#1017) by @yeongseon.

  • MERGE assignment keys from other tables now raise CompileError (#847) — A bound key in when_matched_then_update() or when_not_matched_then_insert() must be a column of the target (or its alias); a source-table column with the same name, previously resolved as the target column, now raises CompileError naming that table. Use the target's column (target.c.name) or string key "name". String keys, unbound sa.column() keys and the column-list shorthand when_not_matched_then_insert([source.c.id, ...]) are unchanged, as are valid SQL and bind order (#895) by @yeongseon.

  • Other changes that can need action — Alembic alter_column() nullability and comment changes now require existing_type (#861). Passing display_width or values to the types listed under Deprecated emits a DeprecationWarning (#870). Unless create_engine(no_backslash_escapes=...) is passed, the first connection probes the server's backslash mode and fails on an unexpected result (#863). Reflected string and date server defaults now come back as quoted SQL literals (#865). CUBRID_PORT is deprecated in favor of CUBRID_TEST_PORT (#1008).

Added

  • Offline tests now save results (#1019) — make test writes junit.xml, coverage.xml and summary.json (PASS/FAIL/INCOMPLETE) to the gitignored test-results/test/, which make clean removes. Failed suites still exit nonzero; test selection and the 95% coverage threshold are unchanged (#1022) by @yeongseon.

  • Integration tests now save diagnostics (#1020) — make integration writes JUnit results, a summary, redacted environment details and, on failure, service.log under test-results/integration/ (or INTEGRATION_RESULTS). Missing or all-skipped results report INCOMPLETE and exit nonzero; cleanup failures and rejected runs report FAIL, not a stale success. Credentials are masked in recorded commands and logs (#1032) by @yeongseon.

Changed

  • Development commands are clearer and interrupted integration runs clean up predictably (#1009) — an interrupt during docker compose up -d in make integration waits one INTEGRATION_STOP_GRACE interval, then sends SIGTERM and allows a second interval before SIGKILL; other interrupted steps get SIGTERM, then SIGKILL after the grace period. Ordinary stalled Compose calls do not have a timeout. New test-fast, format-check and offline-only verify targets are added; make doctor checks Python 3.11+, missing tools and the Ruff/Mypy pins, and make format verifies tool versions. Tool commands remain overrideable through Make variables (#1010) by @yeongseon.

Deprecated

  • make check-all (#1009) — now an alias of make check that prints a deprecation notice on stderr; it will be removed after the next minor release. make test-offline remains an alias of make test (#1010) by @yeongseon.

  • CUBRID_PORT (#1008) — use CUBRID_TEST_PORT, which wins when both are set. When it is unset, CUBRID_PORT is still honoured by docker-compose.yml, make docker-up and make integration (the last two print a deprecation warning); other targets ignore it, so a value exported for another CUBRID tool causes no warning (#1017) by @yeongseon.

  • Unused display_width and values type arguments (#870) — display_width (SMALLINT, BIGINT) and values (CHAR, VARCHAR, NCHAR, NVARCHAR, STRING) never changed the DDL or any behavior; passing them now emits a DeprecationWarning. They are still accepted and ignored, and will be removed in a future major release. The unused private CubridTypeCompiler._get helper is removed (#891) by @yeongseon.

Fixed

  • aggregate_order_by() on a filtered GROUP_CONCAT names FILTER in its error (#995) — on SQLAlchemy 2.1, aggregate_order_by(func.group_concat(col).filter(cond), col) still raises CompileError, but now says CUBRID has no FILTER (WHERE ...) clause instead of claiming the call is not a GROUP_CONCAT. Every other form renders as before (#996) by @yeongseon.

  • aggregate_order_by() no longer drops its ORDER BY (#991) — on SQLAlchemy 2.1, func.group_concat(col).aggregate_order_by(col.desc()) silently rendered GROUP_CONCAT(col), returning values in an unspecified order; it now renders GROUP_CONCAT(col ORDER BY col DESC), also with DISTINCT and in HAVING, with boolean items grouped (#970). Any other function, including aggregate_strings(), raises CompileError, as no other CUBRID aggregate takes ORDER BY (verified for JSON_ARRAYAGG, COUNT, MAX, SUM and MEDIAN). within_group() and SQLAlchemy 2.0 are unchanged. Verified on CUBRID 11.4; not verified on 10.2, 11.0 or 11.2 (#994) by @yeongseon.

  • GROUP_CONCAT DISTINCT, ORDER BY and SEPARATOR no longer fail (#990, #984) — an extra pair of argument parentheses made CUBRID reject the modifiers (-493). func.group_concat(col.distinct()) and literal_column("col ORDER BY col DESC SEPARATOR ';'") now run, and GROUP_CONCAT((x)) renders as GROUP_CONCAT(x). Boolean arguments are grouped once (#970); values are unchanged. Still unsupported (see FEATURE_SUPPORT): a separate SEPARATOR argument, more than one argument, and SQLAlchemy 2.1's aggregate_strings() (aggregate_order_by() is fixed by #991). Verified on CUBRID 11.4; not verified on 10.2, 11.0 or 11.2 (#989) by @yeongseon.

  • EXISTS and AND/OR as a simple CASE value no longer fail (#987) — a bare AND/OR/NOT/EXISTS value= operand of case(), such as value=and_(...) or value=exists(), is now parenthesized as CUBRID requires, like THEN/ELSE values (#982); columns, bind parameters, comparisons and every other rendering are unchanged. The documentation adds three CUBRID limits: a boolean-result CASE without a non-NULL ELSE (such as a single WHEN) can fail with -495 with or without grouping, so pass else_=; CUBRID 11.4 evaluates a CASE window ORDER BY or PARTITION BY item as a constant, without an error; and a boolean-valued CASE fails as a GROUP BY item (-493). Verified on CUBRID 11.4; not verified on 10.2, 11.0 or 11.2 (#988) by @yeongseon.

  • EXISTS and AND/OR as a CASE THEN or ELSE value no longer fail (#982) — a bare AND/OR/NOT/EXISTS THEN or ELSE value of a searched or simple case(), such as and_(...) or exists(), is now parenthesized as CUBRID requires. WHEN conditions and other renderings are unchanged. Verified on CUBRID 11.4; not verified on 10.2, 11.0 or 11.2 (#985) by @yeongseon.

  • Boolean expressions now compile in CAST, function, and some window arguments (#970) — a bare EXISTS, NOT, AND or OR used as a CAST or function argument (in SELECT, filters, joins or updates), and EXISTS or NOT in window and WITHIN GROUP ordering or partitioning, are now parenthesized as CUBRID requires. SQLAlchemy splits a window and_() or or_() into separate items, which can change the meaning; keep it with a list such as over(order_by=[and_(...)]) or .self_group(). WHERE/HAVING/ON are unchanged, and CUBRID still rejects Boolean GROUP BY items, even function-wrapped; docs/FEATURE_SUPPORT.md replaces the earlier guidance (#971). Verified on CUBRID 11.4; not verified on 10.2, 11.0 or 11.2 (#972) by @paikend.

  • EXISTS and AND/OR inside a CAST, a function or ORDER BY no longer fail (#968) — a bare AND/OR/NOT/EXISTS expression used as an ORDER BY item, or inside a CAST or function that is a SELECT column or ORDER BY item (also through type_coerce(), labels and sort modifiers), is now parenthesized instead of failing with -493; for example select(cast(and_(...), Integer)), select(func.coalesce(exists(), False)), a Boolean TypeDecorator with column_expression(), and order_by(exists()). WHERE, HAVING, ON, GROUP BY, CASE conditions and window ORDER BY are unchanged. CUBRID rejects a boolean GROUP BY item even in parentheses; group by a SELECT-list label instead (see docs/FEATURE_SUPPORT.md) (#969) by @yeongseon.

  • Boolean expressions in SELECT lists now compile (#966) — CUBRID rejects unparenthesized EXISTS, NOT, AND, and OR expressions in SELECT results. The dialect now groups these expressions, including labeled columns, subqueries, and CTEs; select(exists().where(...)) and session.query(q.exists()).scalar() work. Their use in WHERE, HAVING, join conditions, and CASE conditions remains unchanged (#967) by @paikend.

  • Repeated isolation re-apply failures no longer lose the retry marker (#937) — after a successful commit or rollback, do_begin() retries a failed isolation-level re-apply. If that retry also failed, the pending marker was cleared too early, so a later direct dialect do_begin() could start work without restoring the level. The marker now stays pending until a re-apply succeeds or a new level is set. SQLAlchemy's normal Connection rollback had masked this gap in the observed reproduction. The preceding commit/rollback still succeeds and no work is replayed. Applies to cubrid+pycubrid:// and cubrid+aiopycubrid:// (#947) by @yeongseon.

  • Server errors no longer cause false disconnects (#916) — With pycubrid 1.13's error_source == "server", sync and async connections no longer treat unrelated server error codes (-1000 to -1200) or misleading error text as CAS connection failures. Actual CAS errors, unknown origins, and pycubrid versions before 1.13 retain previous behavior; the supported pycubrid range is unchanged (#938) by @yeongseon.

  • bool JSON path indices are rejected (#914) — because bool is an int subclass, json_col[True] or json_col[("items", False)] rendered paths such as $[True], which CUBRID 10.2 and 11.4 reject (-494) only after the statement is sent. JSONIndexType and JSONPathType now raise TypeError before any SQL is sent (sqlalchemy.exc.StatementError on execution, sqlalchemy.exc.CompileError with literal_binds). Integer (including 0, 1 and negative) and string components render byte for byte as before on SQLAlchemy 2.0 and 2.1 (#917) by @yeongseon.

  • INSERT / REPLACE ... FROM SELECT over a CTE renders the WITH clause after the column list (#897) — insert(t).from_select([...], select(cte)) and the dialect's replace() equivalent always failed, because CUBRID 10.2 and 11.4 reject a leading WITH c AS (...) INSERT ... (-493). They now render INSERT INTO t (...) WITH c AS (...) SELECT ...; recursive CTEs and bind order are unchanged, and live INSERT, recursive INSERT and REPLACE cases pass on CUBRID 10.2 and 11.4 (#899) by @yeongseon.

  • Unique-constraint reflection fails closed on a catalog mismatch (#848) — get_unique_constraints() silently omitted a unique index that db_index lists but SHOW INDEXES does not, which Alembic autogenerate would read as a constraint to create. Such a mismatch now raises InvalidRequestError naming the unresolved and resolved indexes; a verified-empty catalog still returns []. No such mismatch has been seen on a live CUBRID server; this hardens the reflection invariant (#896) by @yeongseon.

  • Garbage-collected async connections are terminated (#872) — on SQLAlchemy 2.0.44+, when a checked-out cubrid+aiopycubrid:// connection is garbage-collected, its socket is now closed and its pycubrid session retired, even after the connection's event loop has closed, instead of the pool logging "cannot be safely terminated" and the stream raising a ResourceWarning. An invalidated connection closes with CLOSE_DATABASE. Earlier SQLAlchemy 2.0 releases keep the old behavior (#894) by @yeongseon.

  • on_duplicate_key_update() is cacheable (#871) — INSERT ... ON DUPLICATE KEY UPDATE no longer emits SQLAlchemy's "will not make use of SQL compilation caching" SAWarning or recompiles on every execution; a cached statement binds each execution's own values, on SQLAlchemy 2.0 and 2.1 (#893) by @yeongseon.

  • Reflection of names containing :name, collection type classes and join linting (#870) — reflecting a table named like x :y or w (:z) no longer fails with "A value is required for bind parameter 'y'" (the SQL sent is unchanged); SET(Integer) with a type class no longer fails create_all; and SQLAlchemy's join linter no longer reports a joined select as a cartesian product or misses a real one (#891) by @yeongseon.

  • Expression, descending, filtered and prefix indexes reflect and autogenerate (#869) — LOWER(a), b DESC, WHERE c > 3 and the prefix a(5) now reflect as expressions, column_sorting and the new cubrid_where / cubrid_length index options; before, an expression index was dropped with "Skipping index" and re-proposed by autogenerate on every run, and the others reflected as plain columns. CREATE INDEX now renders the two new options, previously ignored silently, and Alembic compares them after normalization, so an unchanged index is not proposed again; detecting a changed sort order, filter or prefix length needs Alembic 1.12+ (#885) by @yeongseon.

  • Alembic renders CUBRID collection member types with the dialect prefix (#868) — autogenerate wrote members such as cubrid_types.MULTISET(sa.MONETARY()), which raised AttributeError when the migration file was loaded. Members from sqlalchemy_cubrid.types now render as cubrid_types.MONETARY() and so on; generic SQLAlchemy members keep sa. (#884) by @yeongseon.

  • Select.fetch() renders as LIMIT (#867) — select(...).fetch(2) and .offset(1).fetch(2) emitted OFFSET ... ROWS FETCH FIRST n ROWS ONLY, which CUBRID rejects (-493); they now render as LIMIT 2 and LIMIT 1, 2. percent=True and with_ties=True raise CompileError, since CUBRID has no LIMIT form for them (#880) by @yeongseon.

  • JSON path keys use JSON string escaping (#866) — a key containing " failed with -1200 "Invalid JSON path", and one containing \ silently returned None. Keys are now built with json.dumps(), so " and \ are backslash-escaped and non-ASCII characters become \uXXXX escapes that do not depend on the connection charset (#879) by @yeongseon.

  • String and date server defaults reflect as quoted SQL literals (#865) — get_columns() dropped string defaults' quotes ('it''s' reflected as it''s; '', the string 'NULL', dates and 'CURRENT_USER' were also misreported), breaking reflected CreateTable SQL and Alembic's existing_server_default. Non-numeric defaults are now read from SHOW CREATE TABLE (one extra query per table that has one) and reflect as 'it''s', '', 'NULL', date '01/02/2020' or the function itself; a string column's '5' stays quoted, and numeric defaults of numeric columns are unchanged (#878) by @yeongseon.

  • Alembic alter_column emits CUBRID-compatible SQL for nullability, dropped defaults and comments (#861) — nullable= and server_default=None produced syntax errors and comment= could not be rendered, so autogenerated nullability changes failed at upgrade. Nullability and comment changes now use MODIFY / CHANGE and require existing_type (otherwise ValueError); a dropped default emits ALTER COLUMN c SET DEFAULT NULL, or two MODIFY statements for a NOT NULL column. Because CUBRID 10.2/11.0 reset attributes that MODIFY omits while 11.2+ keep them, NULL, DEFAULT NULL and COMMENT '' are stated explicitly where needed, and an existing column without a default no longer gets DEFAULT NULL. In offline --sql mode, a UNIQUE create_index raises the documented "live connection" CompileError instead of NoInspectionAvailable (#877) by @yeongseon.

  • literal_binds and other compile options reach CAST operands, ON DUPLICATE KEY UPDATE values and collection members (#862) — these positions dropped the compile options, so literal_binds and Alembic --sql output kept ? placeholders, a functional Index over a CAST rendered a ?, and server_default=cast(literal(5), String(10)) failed create_all with "wrong number of parameters". CUBRID itself still rejects CAST in a function-based index on 10.2 and 11.4 (#875) by @yeongseon.

  • has_table(), get_table_names() and get_view_names() see only the current owner's classes on CUBRID 11.2 and later (#864) — they also matched other owners' classes, although an unqualified name resolves only in the current user's schema: another user's own_t made create_all() skip dba.own_t (so the next INSERT failed), MetaData.reflect() abort with NoSuchTableError, and same-named tables appear twice. They now filter on owner_name = CURRENT_USER. CUBRID 10.2 and 11.0 are unchanged (#876) by @yeongseon.

Security

  • Inline literals follow the server's backslash mode (#863) — on a server with no_backslash_escapes=no, literal_binds rendered the value \' OR 1=1 -- as '\'' OR 1=1 -- ', producing SQL that matched every row, because the dialect assumed no_backslash_escapes=yes. Unless create_engine(no_backslash_escapes=...) pins the mode (and skips the probe), the first connection now probes the server with SELECT CHAR_LENGTH('\\'), and literal rendering doubles backslashes when the server treats them as escapes; an unexpected probe result fails the connection. Native ENUM labels use the same escaping. A dialect that has not connected still assumes the server default (#874) by @yeongseon.

Documentation

  • The architecture pages show the end-to-end driver communication flow (#1013) — docs/ARCHITECTURE.md (English and Korean) adds a Mermaid diagram tracing the three URL variants through their drivers to the Broker, CAS and database server, and the site now renders Mermaid diagrams instead of showing code. Documentation only (#1016) by @yeongseon.
  • Two CUBRIDdb result differences are documented (#1003) — because CUBRIDdb binds values on the server, a CASE whose THEN/ELSE values are all bound parameters returns a string ('10', not 10), and bound parameters in a GROUP_CONCAT argument and aggregate_order_by() item fail with -494. The Driver Compatibility guide gives workarounds (cast(..., Integer); aggregate_order_by(literal_column("1"))). The dialect is unchanged (#1004) by @yeongseon.

  • Boolean expressions that still failed on CUBRID (#968) — the docs/FEATURE_SUPPORT.md note and its case() workaround are replaced, since the #970 fix makes those forms run. Do not use that workaround inside a window: CUBRID 11.4 evaluates a CASE (or if()) in a window ORDER BY or PARTITION BY as a constant (#971) by @yeongseon.

  • The first-connection course has one page per lesson (#943) — Lessons 3 (save and read back a row), 4 (roll back a transaction) and 5 (cleanup and docker rm -f -v) move from the Quick Start to their own English and Korean pages, and its old anchors point to them. Every command was rerun against cubrid/cubrid:11.4 with the published SQLAlchemy 2.1.4, sqlalchemy-cubrid 1.11.0 and pycubrid 1.13.0 (#956) by @yeongseon.

  • The first-connection course starts with two lesson pages (#943) — preparing CUBRID and Python (docs/LESSON_PREPARE.md) and running the first query (docs/LESSON_FIRST_QUERY.md) are separate English and Korean pages, and the site shows previous/next links. Every command was rerun as written against cubrid/cubrid:11.4 with the published SQLAlchemy 2.1.4, sqlalchemy-cubrid 1.11.0 and pycubrid 1.13.0 (#955) by @yeongseon.

  • The documentation site is organized by reader question (#943) — the English and Korean navigation is now Start here, How-to guides, Reference, Help, and Project and internals. No page moved and no URL changed. Two accidental headings in the Connection guide and Feature Support are plain paragraphs again (#954) by @yeongseon.

  • Development documentation now identifies unreleased content (#946) — the documentation site says it is built from main and may describe candidate or abandoned changes, and records its exact source commit. RELEASING.md distinguishes candidates, partial publication, abandoned versions and verified published packages, explains bot-PR CI states such as action_required (observed on #910), and lists the evidence to record; a __version__ value on main is not proof of publication (#951) by @yeongseon.

  • The quickstart removes the lab's Docker volume too (#948) — the English and Korean Quick Start now remove the lab with docker rm -f -v sqlalchemy-cubrid-quickstart, deleting the anonymous volume with the tutorial's testdb that docker rm -f left behind (#949) by @yeongseon.

  • The README async examples are complete programs (#941) — the Async examples used module-level await, a SyntaxError in a script; they now run through asyncio.run(main()), always dispose their engine, and the insert example creates and drops its own sqlalchemy_async_example_notes table. Both were run against CUBRID 11.4. The Korean README carries both programs and the German, Hindi, Russian and Chinese READMEs the corrected query example; an offline test in the always-running lint job checks them (#944) by @yeongseon.

  • A first SQLAlchemy connection for CUBRID beginners (#940) — the Quick Start (English and Korean) is one recommended path from installing sqlalchemy-cubrid[pycubrid] to a first query, a committed row, a rollback, cleanup and the real first-connection errors, run as written against CUBRID 11.4; the README and documentation home link to it. The CUBRIDdb C-extension driver built from source stays supported; only the [cubrid] and [cubriddb] extras are deprecated (#942) by @yeongseon.

  • CAS error -4 is described as an interrupted statement (#873) — pycubrid 1.13.0 reports -4 (ER_INTERRUPTED, a KILL QUERY) as Interrupted with SQLSTATE HY008 (earlier releases: 08S01), and the guides now say so. The dialect still treats -4 as a non-disconnect, as since 1.9.0 (#572); the Driver Compatibility guide also explains why -671 (ER_CSS_RECV_OR_SEND) is not a disconnect code (#939) by @yeongseon.

  • A maintainer's review is the standalone-agent alternative to Copilot (#926) — AGENTS.md and the contributor guides say a standalone agent PR waits for a substantive GitHub Copilot review or a maintainer's review recorded on the PR (maintainer decision, 2026-10-15); a reviewer agent run by the implementer does not count (#927) by @yeongseon.

  • Agent guidance split into task guides (#907) — the root AGENTS.md keeps every rule and gains a task routing table, while detailed context moved unchanged to docs/agents/; scripts/check_agents_docs.py checks its links, commands and version pins, and with --retained-from REF that no earlier line was dropped. No code, CI or release behavior changes (#918) by @yeongseon.

  • Python 3.15 limitation of garbage-collected connection cleanup (#872) — with SQLAlchemy 2.1 (2.1.0 to 2.1.4 checked) on CPython 3.15, a pooled connection reclaimed by the cycle collector is never terminated: its socket closes only through Python's stream finalizer with a ResourceWarning, and its pool slot stays checked out (python/cpython#136189). Every dialect is affected, sync pools included; SQLAlchemy 2.0 and Python 3.14 and earlier are not. The Connection Guide and Support Matrix describe it (#909) by @yeongseon.

  • The README and the documentation site show the CUBRID Lab repository icon (#886) — the README, its translations and the site's logo and favicon use the sqlalchemy-cubrid icon from one immutable cubrid-lab/brand commit (#900) by @yeongseon.

  • ROADMAP.md no longer presents shipped releases as plans (#833) — the Python 3.10 retirement (1.10.0) and Python 3.15 support (1.11.0) moved to Completed, and version numbers are replaced by links to CHANGELOG.md, PyPI and docs/SUPPORT_MATRIX.md (#844) by @yeongseon.

  • docs/PERFORMANCE.md has a current same-CUBRID baseline (#830) — raw pycubrid, SQLAlchemy Core and ORM timings on one CUBRID 11.4 server at main a570ea4 (dialect source as in v1.11.0; pycubrid 1.12.0, SQLAlchemy 2.1.4 and 2.0.54, Python 3.12; 5 trials each) from the new scripts/bench_dialect_baseline.py, with raw data in benchmarks/2026-10-12-dialect-baseline/. No dialect hotspot or regression was found; runtime behavior is unchanged (#843) by @yeongseon.

  • The dialect baseline now covers CUBRID 10.2, 11.0 and 11.2, CUBRIDdb, the async driver and concurrent connections (#830 follow-up) — a 2026-10-13 run through pycubrid, CUBRIDdb (built from cubrid-python v11.3.0.51) and the async driver on 1, 4 and 8 connections (benchmarks/2026-10-13-dialect-baseline-followup/). Reads differ little; pycubrid and the async driver scale with connections, while CUBRIDdb, whose C extension never releases the GIL, does not scale across threads. Runtime behavior is unchanged (#860) by @yeongseon.

  • The feature support table no longer marks CUBRID BLOB / CLOB as fully supported (#829) — they now show ⚠️: DDL, type mapping, reflection, binding bytes / str and NULL round trips work, but an ordinary Core/ORM read of a non-NULL value returns the driver's LOB locator on cubrid://, cubrid+pycubrid:// and cubrid+aiopycubrid:// (Known Issue 6 in docs/DRIVER_COMPAT.md). Documentation only (#842) by @yeongseon.

CI

  • CI publishes the offline test summary and uploads results on success and failure (#1021) — offline jobs write JUnit, coverage and summary.json, add the summary to the job summary and upload test-results/, and integration jobs upload their pytest logs, all with if: always(); the live-lane summary follows #1020. Test selection, the 95% coverage floor and permissions are unchanged (#1027) by @yeongseon.

  • The make-integration jobs also upload test-results/integration/ (#1020) — the live results of make integration join make-integration.log in the existing pytest-logs-make-integration-* artifact; nothing else in the workflows changes (#1032) by @yeongseon.

  • A change to docker-compose.yml runs the test that pins it (#997) — the tooling path filter now lists docker-compose.yml, so a pull request changing only that file runs test/test_make_integration.py (#999) by @yeongseon.

  • The CUBRID 11.4 local Docker test database is kept in a named volume (#1000) — with CUBRID_VERSION=11.4, whose image keeps databases in /home/cubrid/CUBRID/databases, docker compose down lost the data. A second named volume, cubrid-databases, now holds that path, make docker-down deletes both, and make integration refuses a pre-existing <project>_cubrid-databases. Checked on CUBRID 10.2, 11.2 and 11.4; opening a database created by another CUBRID version is not verified, so reset before switching CUBRID_VERSION. Same fix as cubrid-lab/pycubrid#1004 (#1001) by @yeongseon.

  • The local Docker test database keeps working after its container is recreated (#997) — the database was registered under the old container ID, so every connection failed with -353 after recreation. docker-compose.yml now sets CUBRID_DB_HOST: localhost (for 11.4 see #1000 above); for an older volume, make docker-fix-db-host rewrites only testdb's host and keeps the data and a databases.txt.before-fix-db-host backup, or make docker-down (docker compose down -v) deletes it. make integration, CI and the dialect were not affected. Same fix as cubrid-lab/pycubrid#1001 (#998) by @yeongseon.

  • Repository-tooling tests run when only their inputs change (#964) — the tooling path filter now lists agent-evals/**, ten repository-tooling test modules and test/conftest.py, and test_workflow_path_impact.py checks that every REPO_TOOLING_MODULES entry selects the tooling lane (#965) by @yeongseon.

  • Manual agent runs are recorded and summarized as distributions (#906) — scripts/agent_eval.py gains prompt, record and summarize, which record manual runs and summarize them per task and model with a 95% Wilson interval, never a bare pass percentage. Runs are manual-only, from an isolated clone without publish credentials; nothing runs an agent in CI, and no score gates a pull request (#963) by @yeongseon.

  • Agent evaluation tasks for reflection and Alembic (#906) — agent-evals/ adds tasks from the #865 and #861 fixes. Neither has recorded live evidence; the Alembic task passes only a candidate that reports the live migration check as not run and says live behavior is not verified. No dialect source or CI lane changes (#962) by @yeongseon.

  • Offline agent evaluation tasks and grader (#906) — agent-evals/ and scripts/agent_eval.py grade a coding agent's patch for a frozen historical fix (the first task is the #914 fix) against hidden tests, allowed files, documentation, the CHANGELOG and its claims; a live CUBRID claim without a recorded live run, compiled or bound SQL reported as live evidence, or calling the supported CUBRIDdb driver deprecated fails. make agent-eval-check re-grades recorded fixtures without a model or database, in the repo lane; no agent score gates a pull request (#961) by @yeongseon.

  • The documentation check's shell scanner matches pycubrid again (#959) — scripts/_markdown.py takes pycubrid's final _substitution_end() byte for byte (cubrid-lab/pycubrid#978), fixing case/esac, ;&/;;& and comment handling inside substitutions. No other checker rule changes (#960) by @yeongseon.

  • The documentation check allocates site anchors as the site does (#957) — the shared scripts/_markdown.py uses the slugifier configured in mkdocs.yml and the toc extension's id-collision rules; ordinary headings keep their anchors. The change matches cubrid-lab/cubrid-cookbook-python#402 and cubrid-lab/pycubrid#976 (#958) by @yeongseon.

  • The site build runs when the MkDocs provenance hook changes (#952) — scripts/docs_provenance_hook.py, loaded by mkdocs.yml since #951, is now a site path-filter input, so a pull request changing only the hook runs mkdocs build --strict (#953) by @yeongseon.

  • Prepare release checks the last-release-sha baseline override (#945) — preparation stops, asking for removal in a normal PR, when the override set after abandoning 1.11.1 is obsolete (a later release containing it passed the publication proof), malformed, unrelated or unreadable, so a stale override cannot make a recovery silently relist shipped commits. An API error is reported as unresolved, nothing is edited automatically, and RELEASING.md documents the rule; the current 1.11.1 override is active (#950) by @yeongseon.

  • The documentation check handles shell substitutions as pycubrid does (#934) — scripts/_markdown.py takes pycubrid's substitution fixes (cubrid-lab/pycubrid#947): $(...), <(...) and >(...) belong to their word and their bodies are checked as commands, a # after a substitution is not a comment, and an unclosed quote is prose. Review added fixes for case patterns, quoted substitutions, arithmetic expansion and newlines, which pycubrid receives next (#935) by @yeongseon.

  • agent-review no longer counts empty reply-only Copilot records as review rounds (#933) — such records could push a PR past the two-round budget and reject valid evidence; a record that opens a thread still counts. Ported unchanged from cubrid-lab/cubrid-mcp-server#349 (#936) by @yeongseon.

  • Advisory review-evidence check added (#905) — agent-review.yml reads GitHub's actual review records with a read-only token and the base-branch scripts/agent_review_gate.py, not PR checkboxes or bot summaries. It accepts a substantive Copilot review within the two-round budget, a maintainer APPROVED review, or a maintainer COMMENTED review opening an inline thread; changes to literal rendering, identifiers, dialect connections/disconnects, Alembic, workflows, release configuration or gate scripts require maintainer approval. Stale reviews, incomplete API data, missing commit identities and unresolved or author-resolved critical/security threads do not pass: only maintainer resolution satisfies a critical/security finding. Release-please-only and Dependabot PRs are exempt. Adapted from cubrid-lab/cubrid-mcp-server#341, it remains advisory: PRs can spoof its check name, so it must not be required until a trusted identity posts it (#929) (#1015) by @yeongseon.

  • Issue-number lines in block quotes inside list items are reported (#925) — the #<digit> line check now covers quotes such as - > #750, which Python-Markdown renders as headings, and parses container prefixes in linear time (#930) by @yeongseon.

  • The documentation check takes link lines from the source and masks inline HTML (#923) — link line numbers now come from each link's own source, and only what markdown-it actually parsed (inline HTML, autolinks, link tails) is masked before references are checked, so undefined references are no longer hidden or misplaced (#928) by @yeongseon.

  • The documentation check reports lines that the site renders as headings (#920) — a docs/ line starting with # and a digit is plain text on GitHub but an <h1> on the MkDocs site; scripts/check_docs_links.py now reports it and asks for \#, and the 11 existing cases are escaped (#924) by @yeongseon.

  • Documentation link checks handle more Markdown, HTML, Makefile and shell cases (#921) — line numbers across multi-line HTML, grouped and double-colon Makefile targets, references hidden by malformed links or comments, shell continuations and redirections, and HTML <source src> links, matching the reviewed pycubrid parser (cubrid-lab/pycubrid#940, cubrid-lab/pycubrid#941) (#922) by @yeongseon.

  • Deterministic documentation defects block the merge (#908) — a new docs-check job, required by matrix-result when Markdown, scripts/ or the Makefile change, runs scripts/check_docs_links.py and scripts/check_agents_docs.py and fails with file:line on broken link targets or anchors, undefined references, unclosed code fences and nonexistent make goals or scripts/ paths. It parses CommonMark with the pinned markdown-it-py, as pycubrid's gate does (cubrid-lab/pycubrid#939), and never fetches external links; the shared doc-lint workflow stays advisory. <!-- docs-check: ignore --> marks a reviewed false positive (#919) by @yeongseon.

  • Release preparation starts after the abandoned 1.11.1 — release-please-config.json sets last-release-sha to the 1.11.1 merge, so release-please no longer proposes 1.12.0 with every commit since 1.8.0; RELEASING.md documents setting that override and removing it after the next release is tagged (#912) by @yeongseon.

  • A release commit's main push no longer repeats the release's runtime lanes (#737) — on a release commit, ci.yml skips the typecheck, offline, Alembic, packaging and integration lanes that integration-full.yml already runs behind the fail-closed full-matrix-result, and integration-full.yml gains the two pycubrid endpoint steps it lacked; lint, documentation and tooling checks still run. On the 1.11.0 release commit this would have saved 14.1 runner minutes for about 1.4 added; docs/CI_POLICY.md records runner minutes after the #760 and #797 consolidation (#841) by @yeongseon.

Tests

  • The live test for garbage-collected async connections passes in make integration again (#872) — after #1017 bound the Compose broker to 127.0.0.1, hosts resolving localhost to ::1 first (GitHub runners) kept the test's connection alive in an asyncio reference cycle, failing every make integration lane and stopping the 1.11.4 release matrix. The test now reuses a pre-opened pooled connection and still requires the leaked connection to be terminated, with no ResourceWarning (#1035) by @yeongseon.

  • The CASE value and GROUP_CONCAT ORDER BY live tests pass with CUBRIDdb (#1003) — the tests added for #987 and #991 asserted pycubrid's results, so the release matrix's make integration (cubriddb, …) lanes failed on CUBRID 10.2 and 11.4 and 1.11.2 was not published. They now pin the CUBRIDdb results above and keep the exact pycubrid assertions (#1004) by @yeongseon.

  • ORDER BY output matches SQLAlchemy's for plain items (#968) — the dialect renders ORDER BY one item at a time, so a guard test compares its output with SQLAlchemy's own order_by_clause for ORDER BY items that need no grouping (#969) by @yeongseon.

  • SHOW TRACE read failures release both result handles (#854) — a new mocked offline test checks that when reading the SHOW TRACE rows raises, trace_query() closes both results, still attempts SET TRACE OFF and passes the read error to the caller; it does not replace integration coverage on a real server. No production code changes: the behaviour already held, so no failure could be demonstrated (#931) by @tnalbant (first contribution 🎉).

  • Garbage-collection tests no longer depend on the cycle collector (#872) — the live async terminate and closed-event-loop tests now commit before losing the connection, so the pool terminates it when the last reference goes on every supported Python, including 3.15. The reference-cycle case is a separate expected failure where the pool cannot clean it up (SQLAlchemy 2.1 on CPython 3.15). Both had failed on all four Python 3.15 cells of the full compatibility matrix (#894) by @yeongseon.

  • Benchmark artifacts verify the measured dialect and carry their provenance in the report (#903) — scripts/bench_dialect_baseline.py stops if the installed sqlalchemy_cubrid does not match --source-sha, and records the package SHA-256, verification method, image digest, host label and comparable group. Contract sqlalchemy-cubrid-benchmark-artifact/2 rejects an unverified install such as a PyPI wheel and in-place edits of those fields; /1 manifests are rejected but kept unchanged, and the pair is re-recorded in benchmarks/2026-10-14-artifact-pair-v2. Benchmark tooling only (#904) by @yeongseon.

  • Benchmark results can now be archived and compared with provenance (#901) — scripts/bench_artifact.py exports and validates a sqlalchemy-cubrid-benchmark-artifact/1 manifest from a complete dialect benchmark report and compares baseline and candidate runs. It records dialect SHA, driver and SQLAlchemy versions, CUBRID image/server, Python/host, a password-redacted connection target (database user and URL query parameters are still recorded), pool settings, workload identity and every trial. Missing files, corrupt data or mixed source revisions are rejected, and differences other than the dialect make results not comparable. Operator-supplied SHA, image digest and host labels are not independently verified (#903). The pinned v1.11.0-to-main CUBRID 11.4 pair is in benchmarks/2026-10-14-artifact-pair; the older #830 archives are unchanged. Benchmark tooling only: no dialect change and no new CI gate (#902) by @yeongseon.

  • Isolated live CTE coverage (#849) — SQLAlchemy's 14 CTETest compliance nodes are strict xfails because their fixture cleanup fails on CUBRID's per-row foreign-key check, so TestIsolatedCTE849 replays their scenarios on fresh tables against CUBRID 10.2 to 11.4: SELECT, recursive, bound-parameter, UPDATE/DELETE with a CTE subquery and values-via-CTE cases pass, UPDATE/DELETE ... FROM are asserted as documented limits, and INSERT FROM SELECT over a CTE passes since #897 (#898) by @yeongseon.

  • Offline tests for identifier quoting, preflight engine disposal and the reflection probe report (#859, #858, #857) — test/test_base.py covers escaped and reserved identifier quoting, test/test_check_integration_connection.py pins disposal of the preflight sync engine on success and on errors, and test/test_reflection_probe_shape.py pins the reflection_probe report shape with a fake inspector (#978, #979, #981) by @CloudCorpRecords.

  • Offline tests for select attribute accessors, compile benchmark iterations and version-report fallbacks (#855, #850, #851) — test/test_compiler.py covers select attribute accessors, scripts/bench_compile.py validates its iteration count (test/test_bench_compile.py), and test/test_report_driver_versions.py covers the version report fallbacks (#1025, #1026, #1031) by @WilliamK112 (first contribution 🎉).

Full Changelog: v1.11.0...v1.11.5