Client Compatibility

Which BI and ELT clients are checked against Arneb, how, and what does not work yet. The checks run on every pull request in the Cluster smoke CI job (scripts/client-smoke/run.sh): a coordinator and two workers on TPC-H SF1 Parquet, with the client versions pinned in scripts/client-smoke/requirements.txt.

Checked in CI

Client	Protocol	What is checked
psycopg2 2.9	pgwire	Default (non-autocommit) mode: statements inside BEGIN … COMMIT, a bound parameter, version()
psycopg 3.3	pgwire	Server-side parameters (integers, floats, booleans, dates and timestamps sent in binary)
trino-python-client 0.340 (DB-API)	Trino	SHOW CATALOGS, current_catalog / current_schema, an upper-case table name
SQLAlchemy 2.1 Trino dialect (Superset's path)	Trino	Inspector: schemas, tables, columns, keys, indexes, table comment; a reflected table queried with GROUP BY / ORDER BY / LIMIT
dbt-trino 1.10	Trino	dbt debug, dbt test (not-null, unique, relationships, accepted-values tests on sources and an ephemeral model), dbt show, dbt docs generate
SQLAlchemy 2.1 postgresql dialect (psycopg2 and psycopg 3)	pgwire	Inspector: schemas, tables, has_table, views; a query (reflecting a table's columns is not supported yet, see below)
dbt-postgres 1.11	pgwire	dbt debug, dbt test, dbt show, dbt docs generate (the same project as dbt-trino)

The same job checks the 22 TPC-H queries against DuckDB and runs 5 rounds of 20 concurrent queries with the admission cap lifted, failing on any error.

Superset itself and Metabase are not started in CI. Superset reaches Trino through the SQLAlchemy dialect checked above. Metabase's Trino driver uses the Trino JDBC driver; the system.jdbc tables its metadata queries read are covered by the protocol test suite, but Metabase itself has not been run against Arneb yet.

Not supported yet

• Read-only. dbt models can be ephemeral only: table, view and incremental need CREATE TABLE AS / CREATE VIEW on a writable connector. dbt seed and snapshots need writes too.
• PostgreSQL catalog introspection is partial.
  • Over pgwire, pg_catalog reads run in the engine over tables that describe the session's catalog: pg_namespace, pg_class, pg_attribute, pg_type, pg_database, pg_roles, pg_am and pg_tablespace. pg_depend, pg_description, pg_attrdef and pg_constraint are there too, always empty.
  • These work: joins, filters, unqualified names, 'name'::regclass, = ANY (ARRAY[...]), and the catalog functions.
  • Arneb has no array type, so a query that needs arrays (unnest, array_agg, indkey) gets the old fixed answer: the whole first pg_catalog table it names, whatever it asks for. So does a query that needs json_build_object or a correlated subquery in CASE. Each such answer is logged at warn on arneb::protocol with its SQL and counted in arneb_pg_catalog_fallbacks_total.
  • That is how SQLAlchemy's postgresql dialect reflects columns, keys and indexes, so Table(..., autoload_with=engine) and get_columns do not work over pgwire yet. Metabase's PostgreSQL driver has not been tried.
  • Use the Trino protocol for these for now.
  • A query that mixes pg_catalog tables with your own tables cannot see your tables.
• Several statements in one simple-query message (BEGIN; SELECT 1; COMMIT) are refused; drivers send them one at a time.
• JOIN … USING works with INNER and LEFT joins. RIGHT / FULL joins with USING, and NATURAL JOIN, are refused.
• Binary parameters other than integers, floats, booleans, text, dates and timestamps (for example bytea, uuid, numeric) are refused with SQLSTATE 0A000; send them in text format.

Running it locally

python3 -m venv .venv && .venv/bin/pip install -r scripts/client-smoke/requirements.txt
.venv/bin/python scripts/k8s-smoke/gen_tpch.py /tmp/tpch 1
cargo build --release --bin arneb
PYTHON=.venv/bin/python scripts/client-smoke/run.sh /tmp/tpch

The cluster listens on pgwire 15432, Trino 18080 and the Web UI 16432 (PG_PORT / TRINO_PORT change them), and stops when the script exits. scripts/client-smoke/cluster.sh start DATA_DIR RUN_DIR starts the same cluster on its own for manual testing.