Live integration guide recipes — November 3, 2026

The new guide recipes were executed against the retained Enterprise-backed Flight endpoint at grpc://localhost:18816. All 27 inventoried routes passed their stated recipe scope, including remote reads, output writes and read-back. Tableau's write is external ADBC ingestion followed by a native Tableau preview; it is not a native Tableau writer. Go, Java, DuckDB, driver-level JDBC SQL and Superset SQL Lab recipes also passed. See the per-guide summary and machine-readable results.

Three defects in the new documentation were corrected and retested:

Recipe Initial failure Correction and successful retest
Jupyter pandas produced Arrow large_string, while SQL returned string; exact table equality failed despite equal values Pass schema=source.schema to Table.from_pandas; execute the complete cell in a separate real kernel
Airflow The example DAG was not discoverable at module scope; dag.test() could not serialize it Bind dag_instance at module scope and keep the example in an isolated DAG directory; local TaskFlow run finished successfully
DuckDB DuckDB 2.0 removed DuckDBPyConnection.fetch_arrow_table() Use local.sql(...).to_arrow_table(); documented 1.12 ADBC client environment passed read/write and cleanup

Initial failure evidence is retained. Terminal color escapes and trailing whitespace were normalized in the saved text transcripts. An early extraction harness also used package names as Python filenames and put harness files in Airflow's discovery directory. Those harness errors were fixed before the final runs; they are not client failures. The documented filenames use _example.py.

Environment and provenance

  • Existing local Enterprise Flight container: redis-flight-enterprise-test-flight-1; no new Redis/Flight deployment was built.
  • Redis Software container redislabs/redis:8.2.0-78.18, database flight-integration, three active primary shards, active 4x QPF.
  • Verified the Redis TLS hostname against its proxy certificate retrieved through the local Docker control channel and used as a pinned partial-chain trust anchor. Verified ACL username/password authentication and rejection of unauthenticated TLS, plaintext and an untrusted certificate. No global trust configuration changed. Flight clients used local plaintext gRPC.
  • The original host bind mount and executable are no longer readable. The environment receipt records the running container identity, but its binary checksum and full source fingerprint could not be recovered. This is live client recipe evidence, not current-source or release qualification. The source-fingerprint gate was not bypassed to build anything.
  • Guide baseline: d5158fcd93c466412336a8528d1ddafc85a30921, plus the three corrections described above. Guide hashes identify the final reviewed documents; the archive retains the executed code.
  • Reused seven previously prepared dependency profiles from /tmp/flight-repeatable-5068/. Native model runtimes ran in separate processes. Created a fresh clients environment from examples/python/requirements-clients.txt for pandas, Polars, PyArrow, DuckDB and external application writes. Those final client runs used ADBC 1.12.0; the new profile-based examples used 1.13.0. Every environment has a *-versions.json snapshot. This was not a fresh installation test of every package or of JupyterLab's browser interface.
  • JDK 21 with Flight SQL JDBC 19.0.0; R 4.6.2 with adbcdrivermanager 0.25.0 and nanoarrow 0.9.0; RStudio 2026.10.1+239; Quarto bundled with RStudio.
  • Fresh disposable Metabase 0.64.2.9, checksummed community plugin 0.8.0, bundled JDBC 19.0.0. The API transcript records the guide's exact INSERT/UPDATE/DELETE statements and native read-back after each.
  • Retained Superset development runtime reporting package version 7.0.0, sqlalchemy-adbc 0.3.0 and SQLAlchemy 2.0.55. Used a separate copy of its test metadata database and the actual SQL Lab API via Flask's in-process test client. This does not verify a published Superset 7 release installation, Superset 6.1, browser rendering or metadata discovery.

Execution and evidence

Python snippets were extracted from each guide and executed in guide order. Only the endpoint and input fixture name were substituted: the standard sample became the test-owned guide_live_20261103_d2f23049d00d, seeded with the guide's three rows and NULL label. No arbitrary user table was replaced.

  • Library/dataframe/model/plotting recipes read SQL, produce tool-specific output, ingest a unique output table through Flight SQL, verify exact ordered Arrow read-back, then drop the output. Plot/image files remain local artifacts.
  • The Jupyter cell ran in a separate real kernel and its executed notebook was saved. Streamlit ran with AppTest, including the actual write-button click and assertions; browser rendering was not tested.
  • Spark used its JDBC reader, then the documented small driver-side collection and ADBC write bridge. No distributed JDBC writer support is claimed.
  • Airflow used dag.test(), Kubeflow used the local SubprocessRunner and Metaflow used its local datastore. These are real local framework executions, not deployed schedulers or clusters.
  • Quarto rendered HTML containing the successfully executed read/write cell.
  • The combined Go program and both Java guide programs executed against the endpoint. Generic JDBC guide DML also passed via the real JDBC driver; arbitrary JDBC applications remain unqualified by that check.
  • R ran both standalone and in RStudio's actual R Console. The RStudio transcript contains the harness PASS and session data.
  • Tableau created a fresh JDBC data source, selected the test table and clicked Update Now after the external write. Its UI showed four rows, including 99 / guide / 9.50000. The UI observation log records the contemporaneous screenshot/AX observation; no screenshot file or workbook was published. The external DELETE/read-back then passed.

Artifacts include executed source fragments, the notebook, Quarto HTML and plot outputs. Extract with tar -xzf artifacts.tar.gz. Runtime paths and the saved fixture name describe this run; adapt them to your own installed runtimes and a new owned fixture when reproducing. Follow the qualified environment setup, the individual guides and the recorded package snapshots. Jupyter, Streamlit, Quarto and native applications require their documented execution contexts; running every block as an ordinary Python script is not equivalent to those checks.

Remaining application limits and cleanup

Power BI and a compatible connector were unavailable. Its external ADBC write/read-back/cleanup passed, but no Power BI read or refresh is claimed. Redis Insight was not installed: its exact disposable Redis command sequence passed directly over verified TLS, but its UI was not tested. DBeaver had no new SQL snippet in this change; its UI retest could not be completed because its navigator lacked accessible controls and coordinate targeting failed. Existing historical DBeaver results remain unchanged. Superset 6.1's Flight adapter installation blocker remains unchanged.

No native Tableau or Power BI writes, R DBI/adbi, browser dashboards, full ML training service, distributed execution, client-facing TLS/authentication, production retries, throughput or release qualification are claimed. No expected failure is counted as a recipe pass; the historical Datasets persistence defect was not rerun by these new read/write examples.

The cleanup transcript confirms the test-owned input table and row 99 were removed and no guide_ output tables remained. The disposable Metabase container created for this run was removed. Existing Redis/Flight containers and preexisting tables were left running and untouched. Tableau's unsaved preview may retain cached rows after fixture cleanup; nothing was published. Test artifacts and isolated framework metadata remain in /tmp.