analitiq-cdk

The Connector Development Kit: a vendor-neutral, transport-neutral toolbox of reusable building blocks a connector uses. It is consumed both for bulk streaming (the OSS engine reads/writes Arrow batches) and for a synchronous control-plane (discover schemas / create tables) that needs neither the columnar nor the HTTP weight.

The one rule

The dependency points engine → CDK, never back. No CDK module may import anything engine-side (src/grpc, src/state, src/models, src/source, src/engine/{engine,orchestrator,pipeline}, server.py, runner.py, main.py). Anything that must cross from engine to CDK crosses as a plain value or a CDK-owned type — never an engine object. A connectivity capability whose implementation is engine- or deployment-specific is expressed as a CDK-owned seam (Protocol / ABC) the other side implements (SecretsResolver, CheckpointStore).

What lives here

• cdk/types.py — CDK-native value types (AckStatus, Cursor, SchemaSpec, BatchWriteResult, EndpointScope, CheckpointStore).
• cdk/contract.py — capability Protocols (Readable, Writable, Discoverable, TableCreator, ColumnDef).
• transports + ConnectionRuntime, the ADBC sub-registry, shared DB helpers, the query builder, the rate limiter.
• the expression resolver + derived functions.
• the type-map engine (type_map/).
• the secrets seam + local/in-memory resolvers (secrets/).
• the abstract destination base (base_handler.py).

Dependency tiers (core + extras)

The CDK ships a small required core and opt-in extras, so a consumer installs only what its role needs. The split is enforced by lazy imports: cdk.sql and cdk.type_map resolve the Arrow helpers (and the HTTP transport its aiohttp session) only on first use, so importing the core control-plane surface never pulls pyarrow/aiohttp.

Plus the per-driver DB package a given connector needs (asyncpg, adbc-driver-*, …).

The thin-surface entry points raise cdk.MissingExtraError (naming the extra, e.g. analitiq-cdk[arrow]) when the extra is absent, rather than a bare No module named 'pyarrow': the package-level lazy reexports (cdk.sql.AdbcReader / open_adbc_reader, cdk.type_map.parse_arrow_type / resolve_arrow_type) and the HTTP transport (build_http_transport). A genuinely-absent extra is relabelled; an unrelated or broken-install import failure (a partial build failing on a submodule, a missing transitive dep) is re-raised untouched so the real cause survives.

Directly importing a streaming module that belongs to an extra (cdk.schema_contract, cdk.sql_types, cdk.sql.adbc_reader, cdk.type_map.arrow, cdk.sql.generic, cdk.base_handler) without that extra raises a plain ImportError naming pyarrow on the import line — those modules are the extra and require it by design; a thin control-plane consumer reaches the Arrow surface only through the guarded lazy reexports above, never by importing these modules directly.

Install

Published on PyPI as analitiq-cdk. Releases are currently pre-releases, so pip needs --pre (or an exact pin).

Control-plane only (discover + create-table, no pyarrow/aiohttp):

pip install --pre analitiq-cdk

Full streaming surface:

pip install --pre "analitiq-cdk[streaming]"

The engine consumes it in-tree (it is on PYTHONPATH alongside src/) and declares pyarrow/aiohttp itself, so the bundled engine image always has the full surface regardless of the CDK's own extras.