Every ODBC driver, an Arrow driver.

On PyPI, and on the release page. The library rides inside a py3-none-<platform> wheel for Linux x86_64 and aarch64, macOS arm64 and Windows x64 — one wheel for every Python version, no extension module.

The crate carries the C sources; cargo build compiles the driver on the spot (unixODBC headers needed), and a test keeps the bundled copy in sync with the repository.

netstandard2.0 and net8.0; the library ships as runtimes/<rid>/native assets and loads through Apache.Arrow.Adbc.

Over adbc-driver-jni; the natives live inside the jar and are extracted once per JVM. Sources and javadoc jars included.

Over drivermgr (cgo). A Go module cannot carry a prebuilt library, so it comes from install.sh, a release binary or the manifest — or a build tag embeds one.

A fourth SQL backend for the dlt-based loader, built on its maintainer's request: SQLAlchemy reflects the table, the database's ODBC driver reads it, Arrow batches come out, and incremental append and merge work as before. Tested in omniload's own harness on PostgreSQL, CrateDB and SQL Server, and inside its container image. How it was built and tested.

🎉 Merged and released on 2026-09-17: panodata/omniload#362, reviewed by both maintainers, shipped the same day in omniload v0.17.0.

“Perfect. Thank you so much!” “Your contribution was included in v0.17.0, thank you again.”Andreas Motl, omniload maintainer, on the pull request

Two shapes, no change to dlt: a plain resource yielding record batches from adbcbridge.connect(), or a registered sql_table backend so incremental loading works as usual. One million PostgreSQL rows into DuckDB in 2.1 s, and the backend that keeps a timestamp(0) column at second precision. the measurements.

A Go command line that moves tables between databases over ADBC. adbcBridge plugs in as a custom driver, with the ODBC connection string as the source URI: one million rows from PostgreSQL into SQLite in 1.3 s, and 200,000 rows out of CrateDB, where the native PostgreSQL driver stops at its first COPY, in 0.23 s. the measurements.

The driver manager overwrites its own stack and heap, and the process aborts, on the first SQL error from a driver whose SQLWCHAR is 4 bytes. Found on macOS with two vendor drivers; reproduced on Linux with a 40-line fake driver; affects 2.3.12 and 2.3.14.

Affects: OpenLink Virtuoso, sqlflite, Dremio, InfluxDB 3 (the two macOS drivers built to the 4-byte width) · Read the note

The macOS driver is built to iODBC's 4-byte SQLWCHAR and nothing says so — through unixODBC every application dies on its first SQL error. On Linux the ANSI driver treats SQL_C_WCHAR as wchar_t on both sides, stores and reads SQL_C_SBIGINT as 0, steps date parameter arrays by ColumnSize, and aborts on any Charset= but the literal UTF-8 — four programs. Two of them are now sent as fixes (2026-09-18): the abort turned out to be a double free of the charset name, and the date stride comes from the ODBC 3 type names missing from one size switch.

Affects: OpenLink Virtuoso · Read the note

The Apple Silicon build has the same undocumented width, a logger that breaks SQLAllocHandle when enabled, and docs that contradict each other on Apple Silicon support.

Affects: sqlflite, Dremio, InfluxDB 3 · Read the note

Column-wise parameter arrays step fixed-length C types at the BufferLength stride, so every row receives row 1's values; a SQL_BIGINT parameter once bound as SQL_C_DEFAULT keeps writing NULL; SQLPrepare discards the rows-fetched pointer set before it. Three plain-ODBC programs, reproduced on 3.0.1 and 3.5.0-rc1.

Affects: Firebird · Read the note

Timestamp columns cannot be bound (SQLGetData converts them, SQLBindCol does not), timestamp parameters and most BOOL parameter types are refused, and a DECIMAL column fails the whole SELECT — each with the driver's own diagnostic and a program.

Affects: TDengine · Read the note

Server-side prepare refuses a string parameter sent with MySQL type BLOB — what MySQL Connector/ODBC sends for every SQL_C_WCHAR parameter — so Unicode ODBC clients cannot insert through prepared statements: a NullPointerException on 2.1.0, a clean refusal on 4.1.3.

Affects: Apache Doris · Read the note

A bound SQL_NULL_DATA is ignored whenever a value buffer is attached, and parameter arrays stop after set 1 — the driver's one-set-per-SQLMoreResults protocol answers SQL_NO_DATA on the first call. Two programs; the affected-row count thread from 2021 re-measured on 1.5.5. The ignored SQL_NULL_DATA is now sent as a fix (2026-09-18): two lines in the driver decided nullability from the value pointer alone.

Affects: ClickHouse · Read the note

An out-of-range int8 is narrowed to SQL_C_SLONG silently — 9223372036854775807 reads as −1 under SQL_SUCCESS; and with Protocol=7.4-0 a statement freed inside a transaction still sends SAVEPOINT;DEALLOCATE;RELEASE, whose failure on a server without savepoints is swallowed while COMMIT reports success. Wire captures included.

Affects: PostgreSQL, QuestDB, Materialize · Read the note

A BINARY column in a simple-query result is sent with format code 1 where PostgreSQL always sends text, so libpq and psqlodbc both cut the value at its first 0x00; and DEALLOCATE ALL throws a NullPointerException even on a fresh connection.

Affects: QuestDB · Read the note

A multi-statement simple query returns every CommandComplete tagged with the whole string's leading text — BEGIN;INSERT 0, BEGIN;INSERT 1 — so tag-parsing clients get no row count inside a transaction. Side by side with PostgreSQL at the wire level.

Affects: CrateDB · Read the note

Any character value converted to SQL_C_TYPE_TIMESTAMP comes back as today's date under SQL_SUCCESS; and SQLErrorW never clears the connection error, so unixODBC loops forever on any failed connect. Both traced to the source line.

Affects: OpenSearch · Read the note

Column-wise parameter arrays test row 0's NULL indicator for every row: a NULL below it stores an empty string, or segfaults the client on a binary column when WriteInt8Array is handed −1 as a length. The offending lines in parameter.cpp, with a fix.

Affects: Apache Ignite · Read the note

SQLDescribeCol describes a DATETIME(6) as 19 characters with 0 digits while returning 26; and a MySQL-wire server reporting a zero charset width makes SHOW VARIABLES kill the client with SIGFPE in get_column_size().

Affects: MySQL, SingleStore, MatrixOne · Read the note

DEALLOCATE "Name" keeps the quotes in its lookup, so a statement prepared at the protocol level with a mixed-case name — psqlodbc's _PLAN0x… — can never be released, and inside a transaction the failure aborts it. The server half of the psqlodbc row-loss pair; filed with a psycopg side-by-side against PostgreSQL 16, which releases it.

Affects: Materialize · Read the note

BEGIN, COMMIT and the savepoint statements were answered with a zero-field RowDescription before the command tag, so libpq reported PGRES_TUPLES_OK and the Arrow PostgreSQL ADBC driver stopped at Failed to begin transaction with autocommit off. Sent as a fix with a wire-level test and merged by the maintainer the same day (2026-09-17), who re-ran the module's integration suite and extended the fix to prefixed statements such as {sql}BEGIN.

Written up by ArcadeDB on 2026-09-21: ArcadeDB in the Apache Arrow Ecosystem, Through Its PostgreSQL Wire Protocol — their own account of all four findings, the two Arrow routes and the language benchmark.

Affects: ArcadeDB · Read the note

SQLGetInfo(SQL_DBMS_VER) segfaults under unixODBC: inside the driver the call to SQLAllocHandle binds to the driver manager, fails, and the error path frees an uninitialized handle. Linked directly it answers SQL_SUCCESS with an empty string, read into one buffer and returned from another. Sent as a fix rather than a report, and merged by a DuckDB maintainer the same day (2026-09-17): the version now comes straight from the engine, with a test against pragma_version(). A second fix: a parameter array stored set 0's value in every row for every fixed-length type under SQL_SUCCESS; each set now gets its own value, with a test. Merged the same day too.

Affects: DuckDB · Read the note

SQLGetTypeInfo fails with 42S22 Unknown column 'json' whenever sql_mode contains ANSI_QUOTES: the type-info result set is built from a SQL template with double-quoted string literals, which that mode turns into identifiers. And SQLColumns with a NULL catalog name segfaults the client on a connection opened without Database= — strdup(NULL) on the client library's empty current-database field. Both reproduced through pyodbc alone, with the source lines named.

Affects: SingleStore · Read the note