Client compatibility¶
Validated locally on 2026-10-20 against Redis 8.6.2. The server uses Arrow Go 18.8.0, Apache ADBC 1.12.0 and Redis ADBC revision a66467bc5e6a. The local run used macOS arm64 for Go/Python and a disposable Temurin Java 25 Docker image. Hosted workflow configuration is not itself evidence of a passing run.
| Client | Version | Automated coverage | Local evidence / limitations |
|---|---|---|---|
| Go Flight SQL / Flight RPC | arrow-go 18.8.0 | Metadata schemas/types/capabilities, prepared parameters, ingestion, scans, sessions, cancellation, datasets, errors and cleanup; standalone/Cluster race suites | Passed. GetXdbcTypeInfo queried directly, including type filters and real DDL |
| Python ADBC Flight SQL | driver/manager 1.12.0 | quickstart.py and dataset_to_sql.py through Test workflow / make example | Passed: metadata, parameters, queries/updates, ingestion and explicit dataset projection/import |
| Python PyArrow Flight | 26.0.0 | Same Python scripts | Passed: nested/dictionary datasets, multiple batches, projections/ranges, discovery and drop |
| Flight SQL JDBC | 19.0.0 | JdbcSmoke.java; workflow runs Java 17/21/25 | Local Java 25 passed metadata, parameters, update, session state and concurrent statement/read checks. Java 17/21 were not rerun locally |
| pandas | 3.0.6 | Python tour converts a query result to a dataframe | Passed conversion; no claim of arbitrary pandas SQL API coverage |
| DBeaver | Not manually validated on 2026-10-20 | Underlying JDBC smoke only at that date | Later manual connection, navigator and SELECT passed; see 2026-10-30 evidence |
| Tableau | Not validated | None | Requires generic JDBC setup and actual generated-SQL/metadata validation |
| Power BI | Not validated | None | Requires a custom ADBC Power Query connector; none is shipped |
| Polars, DuckDB, R | Not validated | None | Arrow/ADBC consumption examples are not a compatibility guarantee |
Driver v0.0.17 follow-up validation¶
On 2026-10-22, Redis ADBC v0.0.17 (revision 9cc8e25c0f01) passed the Go
standalone and Cluster race suites against disposable Redis 8.6.2, plus both
Python smoke scripts against the rebuilt server. Python pins were ADBC
driver/manager 1.12.0, PyArrow 26.0.0 and pandas 3.1.0. JDBC and manual BI
clients were not rerun. See the validation record.
The earlier matrix retains its original versions and evidence.
For current Superset connection recommendations, see the version-specific guide: direct Redis ADBC for 6.1 and a Flight SQL candidate for 7. The historical results below do not establish a Superset 7 application pass.
Flight client validation on 2026-10-29¶
Source: 84ea021 plus the client smoke scripts added with this validation.
The community server was built on macOS arm64 against its pinned Redis ADBC
revision 9cc8e25c0f01. Tests used a new disposable Redis 8.6.2 container
on port 16389 and a Flight endpoint on port 18815. Clients ran in Linux arm64
containers. These results supplement the historical matrix above.
| Client | Versions exercised | Result and scope |
|---|---|---|
pandas pd.read_sql |
Python 3.11, pandas 3.1.0, ADBC driver/manager 1.12.0, PyArrow 26.0.0 | PASS ordered results, NULLs and empty results through Flight SQL. params= raises NotImplementedError in pandas's ADBC path; direct ADBC binding followed by Arrow-to-pandas conversion passes |
Polars pl.read_database |
Polars 2.1.0, same ADBC/PyArrow versions | PASS ordered results, NULLs, empty results and batch iteration through an autocommit Flight SQL connection |
| DuckDB | DuckDB 2.0.0, same PyArrow version | PASS local aggregates over Flight SQL Arrow results and ordered/NULL results over a Flight RPC dataset uploaded in multiple batches and downloaded from every endpoint. Uses Arrow transfer into local DuckDB; no native DuckDB Flight extension tested |
| R | R 4.5.2, adbcdrivermanager 0.25.0.9000, nanoarrow 0.9.0.9000, Flight SQL shared driver 1.12.0 | PASS DDL, INSERT, ordered query, NULLs, empty results and SUM through the driver manager. DBI/adbi not tested |
| DBeaver / its JDBC driver | DBeaver Community 26.2.3, bundled Temurin 25.0.4.1; Flight SQL JDBC 19.0.0 | JDBC smoke PASS: metadata, parameters, updates, sessions and 100,000-row concurrent read. Application test BLOCKED: repeated macOS accessibility serialization crashes on the custom connection form before a connection was attempted; browsing/generated SQL remain unvalidated |
| Apache Superset | Superset 6.1.0, sqlalchemy-adbc 0.1.0, 0.2.0 and 0.3.0 | BLOCKED before application execution: Superset requires SQLAlchemy >=1.4,<2; all three adapter versions require >=2.0. pip returns ResolutionImpossible. No dashboard or SQL Lab pass is claimed |
The R packages came from https://apache.r-universe.dev with CRAN as fallback.
These are development versions, not release pins. adbcflightsql was unavailable
from those repositories for this runtime. Instead, the R test loaded the official
Linux arm64 libadbc_driver_flightsql.so from the Python 1.12.0 wheel through
adbc_driver(path, entrypoint = "AdbcDriverFlightSQLInit"). Supply the driver
library built for the R runtime's OS and architecture; the Python installation
is only a source of that shared library and is not used to execute R queries.
Reproduce dataframe and R checks¶
Against a disposable server, run:
make clients-smoke FLIGHT_URI=grpc://localhost:8815
# Or install requirements-clients.txt into an isolated Python environment:
python -m pip install -r examples/python/requirements-clients.txt
python examples/python/client_compatibility.py --uri grpc://localhost:8815
# Install R's adbcdrivermanager and nanoarrow first, then provide the library:
make r-smoke FLIGHT_URI=grpc://localhost:8815 FLIGHT_SQL_DRIVER=/path/to/libadbc_driver_flightsql.so
Scripts create unique tables/datasets and clean up their fixtures. Their scope is an unauthenticated local endpoint; TLS/authentication and arbitrary generated BI SQL are not covered. The dataframe script reports the pandas parameter limitation explicitly and verifies the direct-ADBC workaround. Its remaining assertions must all pass. The Test workflow now invokes it; hosted CI was not run during this local validation. R remains an explicit local target.
Reproduce the Superset installation failure in a fresh Python 3.11 environment:
python -m pip install 'apache-superset==6.1.0' 'sqlalchemy-adbc[flightsql]==0.3.0'
Do not force incompatible dependencies with --no-deps and count that as a
supported installation. This blocks this specific adapter route, not every
possible future Superset Flight integration.
DBeaver application attempt (Superset excluded)¶
After the user requested continuing without Superset, DBeaver Community 26.2.3 was downloaded from the official site into a temporary macOS arm64 installation with an isolated workspace. Its bundled Temurin 25.0.4.1 already included the README's two Arrow JVM options. The connection picker had no Arrow entry. Driver Manager successfully saved a custom generic driver with these settings:
Name: Redis Flight SQL
Class: org.apache.arrow.driver.jdbc.ArrowFlightJdbcDriver
URL template: jdbc:arrow-flight-sql://{host}:{port}/?useEncryption=false
Default port: 18815 (isolated test endpoint; normal project default is 8815)
Library: flight-sql-jdbc-driver-19.0.0.jar
Opening the custom connection form repeatedly crashed DBeaver during UI
automation, including an attempt using screenshots instead of accessibility
tree inspection. Three local macOS crash reports recorded EXC_BAD_ACCESS /
SIGABRT; the inspected stack included CFStringGetLength,
cfAttributedStringSerialize, AXSerializeCFType and
_AXXMIGCopyParameterizedAttributeValue. This is a local application/automation
blocker, not a demonstrated Flight SQL protocol failure. No successful Test
Connection, schema browsing or query execution was verified in that attempt.
See the later manual pass.
Reported upstream as DBeaver issue #42640.
Manual validation remains: connect with autocommit, expand redis/public,
inspect columns, open a table's data tab, and run a filtered aggregate. Record
the generated SQL and results separately from the successful JDBC smoke test.
Superset follow-up is deferred under the user's instruction; it is tracked in
the endpoint repository
and the adapter repository.
DBeaver rerun and manual validation on 2026-10-30¶
Rebuilt source 4a26c34 and reran against a fresh disposable Redis 8.6.2
on 127.0.0.1:16389, with Flight on 127.0.0.1:18815. JDBC 19.0.0
in Temurin 25 passed every JdbcSmoke.java check, including metadata,
prepared statements, updates, session state and the 100,000-row concurrent
read. The known empty JDBC getTypeInfo() result remains unchanged.
Retried the installed macOS arm64 DBeaver Community 26.2.3 with the custom
driver configuration above. Two application attempts crashed: opening the
connection form, then a retry that clicked Test Connection and reached
“Set connection defaults.” No successful connection result, schema browsing,
data tab or generated SQL was verified. Crash reports at 08:00:22 and
08:00:58 PDT reproduce EXC_BAD_ACCESS / SIGABRT; the inspected stack
again includes CFStringGetLength, cfAttributedStringSerialize,
AXSerializeCFType and _AXXMIGCopyParameterizedAttributeValue.
These automation attempts remained blocked by the local accessibility crash.
A later manual test on the same date succeeded with the installed DBeaver
Community 26.2.3, bundled Temurin 25.0.4.1 and Flight SQL JDBC 19.0.0, using
Generic driver type and the same local ports. The user confirmed connection;
subsequent screen inspection showed redis → public → Tables → dbeaver_sample
and a result grid for:
SELECT "id", "name", "qty"
FROM "redis"."public"."dbeaver_sample";
The grid contained (1, Apples, 10), (2, Oranges, 20) and (3, Pears, 30).
This establishes manual connection, catalog/schema/table navigation and a quoted
SELECT with correct results. The fixture was created through JDBC in a separate
connection and found through JDBC metadata before the manual test. It does not
establish DDL or writes from DBeaver, expanded column metadata, a table Data tab,
filtered aggregates or arbitrary generated SQL. TLS/authentication and large-result
UI paging were not tested. The automation crash is separate from this manual pass.
Reproduce the workflow with DBeaver setup and manual tests, which links the screenshot guide and the exact 19.0.0 JAR. The direct Maven Central download was reported faster for manual setup; no download benchmark was run. Built-in driver support is proposed in DBeaver PR #42645, linked to feature request #15535. The tested application used a custom driver; the proposed packaged driver entry has not been validated in a rebuilt DBeaver product.
Core endpoint validation on 2026-10-29¶
A separate run against source b84b1e2a1693a0269dd503ae84f2a353dc724a39
passed both standalone and three-shard Cluster Go race suites (80 tests passed
and one enterprise-only admission test skipped in each). Python 3.14 with the
pinned ADBC 1.12.0, PyArrow 26.0.0 and pandas 3.1.0 passed the quickstart and
dataset-to-SQL scripts; Flight SQL JDBC 19.0.0 on Temurin Java 25 passed its
smoke test. Community metrics returned the expected empty HTTP 204 response.
The validation record and raw logs identify the tested revision, commands, versions and limitations. This run did not exercise DBeaver, Superset or production deployment behavior, and does not supersede the separate application attempts above. JDBC's empty type information and the enterprise-only skip remain explicit limitations.
Reproducing the standard client checks¶
Use a disposable Redis with Search and a running server. Pin Python dependencies
with examples/python/requirements-smoke.txt, then run:
python examples/python/quickstart.py --uri grpc://localhost:8815
python examples/python/dataset_to_sql.py --uri grpc://localhost:8815
make jdbc-smoke
The scripts accept --user and --password for ACL-authenticated test servers.
Use ephemeral test credentials; do not embed production credentials in commands.
Current client validation requires a qualified enterprise Flight build backed by
a Redis Enterprise cluster with verified TLS and username/password authentication.
See integration setup and private CI. Public CI runs unit/build
checks only. Go backend tests require an authenticated rediss:// REDIS_URI;
make test-integration fails if it is absent. The JDBC Make target supplies the
required JVM memory flags.
JDBC type information and schema caveats¶
JDBC 19's DatabaseMetaData.getTypeInfo() returns an empty result: its
ArrowFlightMetaImpl
inherits Avatica's default implementation rather than forwarding this call to
Flight SQL GetXdbcTypeInfo. JdbcSmoke records that client limitation explicitly;
the Go wire-level tests validate the endpoint's type information separately.
This is independent of table schemas returned through GetTables/GetSchema.
#17 covers JDBC 19 classifying prepared statements with no dataset schema as updates. Use separate statements for scripts whose final schema depends on earlier writes, and explicit casts/types for parameter-only projections. Do not claim a newer JDBC release fixes this without checking and retesting it.
JDBC and Metabase validation on 2026-11-01¶
Against community source f8a2723, disposable Redis 8.6.2 and the pinned Redis
ADBC revision 2f053d3affad, Flight SQL JDBC 19.0.0 on Java 25 passed the
existing read/write smoke suite. Metabase 0.64.2.9 with community Flight SQL
plugin 0.8.0 passed native reads (including NULLs/empty results), table/column
sync, generated filters/aggregates and custom SQL INSERT/UPDATE/DELETE Actions
with SELECT read-back after each write. Metabase native DDL via its read-query
endpoint failed with Select statement did not produce a ResultSet for native query.
No browser/dashboard rendering or implicit model actions were tested. Tableau
was not installed: application reads and writes remain blocked, independently
of the passing JDBC driver test. See versions, reproduction and logs.
Tableau read validation on 2026-11-01¶
After the Metabase run above, Tableau Desktop Free Edition 2026.3 (Apple silicon)
was installed and manually tested with Arrow Flight SQL JDBC 19.0.0 against
community source f8a2723, Redis ADBC 2f053d3affad and disposable Redis 8.6.2.
Other Databases (JDBC), with SQL92 dialect, discovered database redis, schema
public, the fixture table and its three columns. Its generated data-preview
SELECT displayed all three expected rows, including a NULL amount. No Tableau
JVM options, TDC overrides or custom connector were needed for this test.
Acceptance is read-only at the requester's direction. Worksheets, filters, aggregates, dashboards, extracts, joins and TLS/authentication were not tested. This supersedes the earlier installation blocker for the recorded read scope. See setup, expected values and evidence limits.
All 27 inventoried Flight SQL routes — November 3, 2026¶
The saved integration sweep exercised all 21 previously candidate routes and reran the six previously exercised routes. All 27 have scoped passes: 24 automated tool workflows, RStudio and Tableau native UI checks, and Metabase API checks. The final run recorded 28 automated passes and two strict Datasets 5.1.0 expected failures; the one-shard empty-save workaround passed. Expected failures are not passes.
The repeatable runner uses seven isolated dependency
profiles, persists JUnit/logs/versions/artifacts, isolates native ML runtimes,
and keeps application evidence explicit. The local run used community source
c019afc, standalone Redis Open Source 8.6.2 with Search, and ADBC Flight SQL
1.13.0. It does not establish Software/QPF, TLS/authentication, distributed
execution or full application certification. Operation-specific scopes and
remaining limitations are recorded with every route.
Live guide recipes — November 3, 2026¶
The guide recipe validation executed the documented client flows against the retained Enterprise-backed endpoint after the historical OSS sweep above. All 27 inventoried routes have scoped recipe passes: output writes and exact read-back through ADBC for library workflows, native RStudio console execution, Tableau external ingestion followed by native preview, and Metabase Actions plus native API reads. Java, Go, DuckDB, driver-level JDBC SQL and Superset 7.0.0 development SQL Lab APIs were also exercised. This is not a claim of native database writers in every tool.
The notebook, Airflow and DuckDB recipes were corrected and rerun successfully. Backend checks verified TLS/hostname against a Docker-sourced pinned proxy certificate, ACL identity, rejection of plaintext/unauthenticated/untrusted connections, three active primary shards and active 4x QPF. Flight clients used local plaintext gRPC. The old host bind mount disappeared, so the server's binary checksum and full source fingerprint could not be recovered. These client checks do not qualify current-source builds or a release. Consult the per-guide results for partial Power BI/Redis Insight checks and unchanged application UI limitations.