[rad-pat]

Summary

Adds an asyncio dialect, so Databend works with create_async_engine, AsyncConnection.stream() and AsyncSession:

from sqlalchemy.ext.asyncio import create_async_engine

engine = create_async_engine("databend+async://user:password@host:8000/db?sslmode=disable")

create_async_engine given a plain databend:// URL selects the async dialect too. Install with pip install "databend-sqlalchemy[asyncio]", which brings in greenlet. SQLAlchemy 2.1 no longer installs greenlet by default.

This PR also fixes CI, which has been failing since mid-October for reasons unrelated to the open PRs (see below), and moves packaging to pyproject.toml.

How it works

• databend_sqlalchemy/async_connector.py: the driver's AsyncDatabendClient has no async cursor, so this adds a PEP 249-style async connection and cursor.
  • It renders parameters exactly as the blocking cursor does, through an OperationRenderer mixin pulled out of connector.py. Sync behaviour is unchanged: across about 2,000 parameter and statement shapes, the SQL handed to the driver is byte-identical to before.
  • All cursors of a connection share one driver connection, so temporary tables and SET persist across a connection's statements. The blocking connector opens a session per cursor.
  • Batch fetches (fetchmany/fetchall) drain the driver's row iterator in a worker thread, one hop per batch. Awaiting the iterator once per row costs a round trip through the driver's runtime each time, which made 200k rows take about 24 s; it now takes about 0.4 s.
• databend_sqlalchemy/async_dialect.py: builds on SQLAlchemy's asyncio adapter (sqlalchemy.connectors.asyncio).
  • Results are buffered by default and stream with stream() or stream_results.
  • The cursor also buffers the rows executemany leaves behind, because post_exec reads the inserted row count back from them.

Tests

• The full SQLAlchemy compliance suite runs against the async dialect with pytest --db async, added as a second CI step.
• New tests/test_async.py covers the asyncio API itself:
  • buffered results that outlive the connection;
  • streaming, checking that rows are still pending in the driver after the first partition;
  • single and executemany rowcounts;
  • session state shared across statements;
  • errors raised as DBAPIError;
  • async_creator;
  • an ORM AsyncSession round trip.
• ServerSideCursorsTest overrides now assert real behaviour. test_ss_cursor_status used to call a bare super() and so never ran. It now checks buffered and streamed cursors under the async dialect.

Python	SQLAlchemy	databend-driver	sync	async
3.11 (as CI)	2.1.6	0.34.2	744 passed	756 passed
3.13 (Windows)	2.0.48	0.27.4	711 passed	723 passed
3.10	1.4.54	0.34.2	35 passed (compliance suite not collected)	n/a

All runs had 0 failures. Before rebasing onto current main, the same changes also passed on:

• Python 3.10 with SQLAlchemy 2.0.55;
• Python 3.14 and 3.15rc1 with SQLAlchemy 2.1.6;
• the minimum versions: Python 3.10, SQLAlchemy 2.0.42 and databend-driver 0.26.2.

CI fixes

The open PRs fail CI because the job dies before any tests run:

• MinIO: the databend:nightly image's MinIO download now returns HTTP 410. MinIO never starts, so Databend fails with StorageUnavailable and the health check gets a connection reset. No test needs object storage (stage COPY INTO works on the default fs storage), so MINIO_ENABLED and port 9000 are removed.
• SQLAlchemy 2.1 (installed by pipenv --skip-lock) removed DeprecatedCompoundSelectTest, so tests/test_sqlalchemy.py failed to import. That import and override are now guarded to < 2.1.0.
• Sequence tests race with -n4: Databend sequences belong to the tenant, not a database, so parallel workers raced on SequenceTest's tab_id_seq (SequenceAlreadyExists / UnknownSequence).
  • The three sequence test classes now share one xdist_group, and pytest runs with --dist loadgroup, so they run on a single worker while everything else stays parallel.
  • conftest.py runs SQLAlchemy's collection hook first; otherwise xdist tags the items before SQLAlchemy replaces them with the per-backend classes.
• SQLAlchemy 1.4: a new CI step runs the tests on SQLAlchemy 1.4 so Superset compatibility doesn't regress. The compliance suite needs SQLAlchemy 2.0.42's test requirements (nvarchar_types) and isn't collected below that, nor are the async tests below 2.0.28. New tests/test_basic_usage.py therefore covers what Superset relies on, on every SQLAlchemy release: reflection, views, executemany, LIMIT/OFFSET, text() with %, and type round trips.

Packaging

• pyproject.toml: metadata moves there from setup.cfg (PEP 621, with a PEP 639 licence expression). setup.cfg keeps only the flake8 and SQLAlchemy test-plugin sections, which have no pyproject.toml support.
• asyncio extra added.
• Python: requires-python >= 3.10, which the package already needed (types.NoneType, runtime X | Y unions). Classifiers cover 3.10 to 3.15.
• SQLAlchemy: still >= 1.4 for the sync dialect, because Superset (up to its current 6.1 release) requires SQLAlchemy < 2. The asyncio extra requires >= 2.0.28, and the async dialect raises an ImportError saying so on anything older. 2.0.28 is the first release where ExecutionContext.rowcount returns the _rowcount the dialect sets in post_exec. Before it, rowcount is -1 for INSERT/UPDATE/DELETE/COPY INTO.
• README is published as text/x-rst (it was declared as markdown), and one short heading underline is fixed so it renders. twine check --strict passes.
• superset extra now points at apache-superset.

Known limitations

• Pre-existing: the executemany multi-row INSERT path doesn't undo the escaper's % doubling, so '50%' is stored as '50%%'. This affects sync and async alike; single-row inserts are fine.
• Abandoned streams on older drivers: with databend-driver < 0.34, a stream abandoned before it's exhausted leaves its query parked on the server until it times out. This happens with the sync dialect too; 0.34 kills the query when the iterator is dropped.
• Cancellation: a cancelled async statement isn't killed on the server.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-23T13:09:43.367803Z	37fbfed	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 37fbfed601

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[rad-pat]

@hantmac, please take a look if you get chance 👍