erpl-dlt

Licence notice, above the fold. This package — the Python glue — is Apache-2.0. The ERPL DuckDB extension binaries it loads are separately licensed under BUSL-1.1 by DataZoo GmbH. Installing erpl-dlt grants you no licence to them. See NOTICE.

dlt sources for SAP, built on the ERPL DuckDB extensions.

pip install erpl-dlt

# or pull the container image
docker pull ghcr.io/datazoode/erpl-dlt:2026.9.4

import dlt
from erpl_dlt import erpl_rfc_source

pipeline = dlt.pipeline(pipeline_name="sap", destination="duckdb", dataset_name="sap_raw")
pipeline.run(erpl_rfc_source(table_names=["SFLIGHT", "SBOOK"]))

Five SAP read paths, one connection layer:

Source	Reads
erpl_rfc_source	Any table or CDS view over RFC, with projection and filter push-down
erpl_odata_source	Any OData v2/v4 service
erpl_odp_source	ODP delta — real server-side change data, over RFC or the SAP Gateway
erpl_bics_source	BW InfoProviders and BEx queries
erpl_invoke_source	Remote-enabled function modules and BAPIs

Configuration

Credentials resolve through dlt, so nothing needs to be passed in code:

# .dlt/secrets.toml
[sources.erpl_dlt.credentials]
ashost = "sap.example.com"
sysnr = "00"
client = "100"
user = "SVC_DLT"
password = "..."

They are bound as parameters to DuckDB's CREATE SECRET, never formatted into SQL text — query text reaches logs.

What the sources do

from erpl_dlt import erpl_rfc_source

erpl_rfc_source(
    table_names=["SFLIGHT", "SBOOK"],
    # Pushed into SAP. On a 55-column table, naming two columns measured 3.8x
    # faster than reading all of them.
    columns={"SBOOK": ["CARRID", "CONNID", "FLDATE", "LOCCURAM"]},
    # Becomes a WHERE clause that DuckDB pushes into the RFC call, so SAP does
    # the filtering. `EXPLAIN` shows it as `Filters:` inside SAP_READ_TABLE.
    cursor_columns={"SBOOK": "FLDATE"},
    primary_keys={"SBOOK": ["CARRID", "CONNID", "FLDATE", "BOOKID"]},
)

Extraction is Arrow-native: batches come from DuckDB's to_arrow_reader and go straight to dlt, so no row becomes a Python dict. SAP's types arrive correct without casting — DATS as date32[day], CURR as decimal128, an empty SAP date as NULL. dlt's naming normaliser will rename columns containing digits: AS4DATE lands as as4_date.

Each resource takes its own DuckDB cursor, so parallelized=True is safe.

erpl_odp_source uses SAP's ODQ subscriber process as the delta identity. By default that name is stable for the dlt pipeline name and ODP provider, and it is capped at SAP's 32-character field width with a hash suffix so long names do not collide by truncation. If you need a pre-existing SAP subscriber, pass subscriber_processes={provider_name: "SUBSCRIBER"}; after a delta run records a subscriber in dlt state, a later run that would use a different subscriber is refused instead of silently starting a new delta queue.

ODP delta reads are not transactional with respect to the destination load. SAP's pointer can advance while dlt commits nothing; what can be recovered then depends on ODQ confirmation semantics. ERPL's source shows a recover mode for the last unconfirmed packet, but whether a particular installed ERPL build exposes that mode is checked from duckdb_functions() at runtime.

Development

uv venv && uv pip install -e ".[dev]"
just check            # ruff, mypy --strict, unit tests
ERPL_IT=1 just it     # integration tests against a real SAP system

Integration tests are skipped unless ERPL_IT=1 and credentials are set. They use no mocks: they assert that a projection reaches SAP, that EXPLAIN shows the predicate inside the scan, that ODP's delta token survives a fresh process, and that DECIMAL/DATE/TIME survive into the destination schema.

Status

Tested against ERPL and ERPL Web v2026.09.04, DuckDB 1.5.5, dlt 1.30, Python 3.10–3.13, on Linux x86-64 — the only platform ERPL publishes these extensions for. Releases follow CalVer matching the supported ERPL release (e.g. 2026.9.4).