For SAP data on Databricks, zero-copy (SAP Business Data Cloud Connect over Delta Sharing) should be your default for SAP-managed data products consumed for analytics and AI. Replication still wins where you need full physical control of the data: custom ABAP objects outside the data-product catalog, deep history and SCD2, heavy repeated scans, and strict data-residency or multi-platform governance. Most mature SAP estates will end up running both, deliberately.
For two decades, getting SAP data into an analytics platform meant one thing: extract it. ABAP reports, BW extractors, SLT triggers, ODP-based CDC tools. Every pattern produced a copy, and every copy brought pipelines to maintain, lost business semantics, and a second governance boundary.
That model is now being challenged from two directions at once. SAP has tightened what third-party tools may use to pull data out. And SAP and Databricks have shipped a supported zero-copy path between their platforms.
Why the SAP-to-Databricks decision changed in 2026
Two events turned a tooling preference into an architecture decision.
1. The ODP-RFC restriction is now enforced.
SAP Note 3255746 has been tightening since 2022. Version 11, released on April 21, 2026, states that ODP-RFC modules are designed only for SAP-to-SAP transfer, and that any use by customer or third-party applications against on-premise or private-cloud ABAP systems is prohibited. SAP began technically blocking unauthorized ODP-RFC calls with a security patch on Patch Day, June 9, 2026, and offers a time-limited opt-out through the end of 2026. The note now points customers toward SAP Business Data Cloud or the ODP-OData API as alternatives.
The nuance matters for engineers: the note does not ban RFC as a protocol. Table reads, CDS view extraction, BAPIs and DeltaQ remain usable. What is closed is the ODP Data Replication API over RFC, which happened to be the backbone of many CDC connectors, including the Azure Data Factory SAP CDC connector.
2. Zero-copy became production-grade.
SAP Business Data Cloud Connect for Databricks is generally available on all clouds. It lets an existing Databricks workspace consume SAP data products live and bi-directionally over Delta Sharing. At TechEd Berlin in late October 2026, the partnership expanded further, with Lakeflow, Unity Gateway and Lakebase planned for SAP Databricks. Databricks cites Adidas, Zalando, Mondi and CAF as production users.
The strategic signal is clear. SAP wants SAP data to stay in SAP-governed storage and be shared out, not extracted.
How zero-copy actually works for SAP data
Zero-copy here means one physical copy of a data product, read by more than one engine. It does not mean no copy of SAP data ever exists outside the ERP.
The path runs as follows:
- Materialization inside BDC. SAP-managed data products are generated from source line-of-business definitions into the BDC object store as Delta tables, modeled in bronze and silver layers. For S/4HANA, that load is itself a replication, run and governed by SAP.
- Publication as a share. SAP BDC acts as a Delta Sharing provider. A Databricks admin creates an SAP BDC connection from an invitation link issued by the BDC admin, which registers the BDC account as a provider (and as a recipient, for write-back). The trust is secured with mTLS and OpenID Connect.
- Consumption in Unity Catalog. Shared data products surface as Delta tables your SQL warehouses, notebooks and pipelines query directly. Databricks compute reads SAP-hosted Parquet files; nothing is landed in your storage.
- Write-back. Tables you build in Databricks, such as a demand forecast or an enriched customer view, can be shared back to BDC for use in Datasphere or SAP Analytics Cloud.
Two deployment shapes exist, and they are often confused. SAP Databricks is a Databricks workspace embedded in BDC and billed through SAP. BDC Connect links BDC to the Databricks estate you already run. Sources cite that SAP Databricks bills in SAP Capacity Units with lower cost transparency, while Enterprise Databricks with BDC Connect keeps full bi-directional sharing and DBU-level visibility. For an organization with an existing lakehouse, BDC Connect is usually the relevant comparison.
The real value is semantic, not just economic. Data products arrive with SAP business context intact: the joins, the currency and unit conversions, and the hierarchy logic that extraction pipelines have historically reimplemented, badly, in SQL.
What replication looks like in 2026
Replication is still very much alive. It is just narrower, and compliance now shapes tool selection as much as throughput does. The supported patterns today fall into four groups:
- SAP Datasphere Replication Flows to hyperscaler object storage (S3, ADLS Gen2, GCS), which Databricks then ingests. This is SAP’s documented route to external stores. Outbound flows to non-SAP targets carry premium outbound integration charges.
- ODP over OData through SAP Gateway, the API SAP now names as the compliant alternative. It is supported, but it has lower throughput ceilings than RFC for large initial loads.
- Non-ODP third-party extraction: table-level CDC, CDS view extraction, DeltaQ and BW cube readers from vendors that have re-platformed away from ODP-RFC.
- Database-level replication, with caveats. SAP does not support log-based replication from HANA (Note 2971304), and trigger-based approaches need a full-use HANA license for ERP.
Whatever the tool, replication lands data you own, in your storage, under your Unity Catalog policies. Your team controls layout, clustering, retention and history. That control is exactly what you pay for in pipelines, reconciliation effort and semantic rework.
Zero-copy vs. Replication: Head-to-Head
| Dimension | Zero-copy (BDC Connect + Delta Sharing) | Replication (into your lakehouse) |
| Freshness | Live read of the current data-product version | Bounded by CDC or batch interval |
| Semantic fidelity | SAP business context preserved in the data product | Rebuilt by your team; drift risk |
| Scope | Data-product catalog plus extended or custom products | Anything you can legally extract |
| Physical layout | Owned by SAP; recipient cannot re-cluster | Fully yours: clustering, Z-order, partitioning |
| History / SCD2 | Depends on what the data product exposes | Fully under your control |
| Governance | Two catalogs; authorization not unified by the protocol | Single Unity Catalog boundary |
| Compliance exposure | SAP-endorsed path | Tool-dependent; ODP-RFC tools are non-compliant |
| Cost model | BDC licensing plus recipient compute on every read | Pipeline compute, storage, integration licenses |
| Pipeline effort | Minimal for standard products | High: build, monitor and reconcile |
Neither column is a superset of the other. The trade is semantic fidelity and operational simplicity against physical control.
Where zero-copy breaks down
Zero-copy is the right default, but it is not free of engineering consequences. These are the failure modes we see senior teams underestimate.
- Catalog coverage is the real boundary. You can share only what exists as a data product. Custom Z-fields flow through if the underlying CDS view is extended and the data product is then activated or reinstalled. Custom data products can be built from S/4HANA using replication flows into the BDC object store. Loading local Datasphere tables into the object store was originally not possible, but SAP has since added a route: share the table from a standard HANA space to a file space and load it from there. The route has constraints, such as primary keys being required, so the gap is narrowing rather than closed. Heavily customized ECC estates with bespoke Z-tables will hit this wall first.
- You cannot tune what you do not own. A Delta Sharing recipient reads the provider’s files as laid out. You cannot run OPTIMIZE, apply liquid clustering or change file sizes on a shared SAP table. For a dashboard that scans a 2-billion-row ACDOCA-derived product dozens of times an hour, that matters. The usual fix is a curated, materialized gold table in your own storage, which is replication by another name. Be honest about it in the architecture.
- Governance spans two catalogs. Delta Sharing grants read access to a table; it does not carry row-level or column-level policy semantics across systems. Your SAP authorization model and your Unity Catalog policies must be reconciled explicitly, especially for HR and finance data.
- Every query is a remote read. Zero-copy moves cost from storage and pipelines to query-time compute and network. Keep the BDC tenant and the consuming workspace in the same cloud region. If they cannot be colocated, model the cross-region read volume before you commit.
- History is the provider’s choice. If an ML feature pipeline needs point-in-time correctness over five years of material movements, confirm what history the data product retains and whether change data is exposed. Otherwise you will be snapshotting it yourself.
Also Read: End of Support for SAP BW 7.5
A decision framework
Decide per data domain, not per platform. Run each SAP domain (finance, procurement, supply chain, HR) through five questions, in order:
- Is it covered by a standard or extensible data product? If yes, start with zero-copy. If not, ask whether a custom data product is feasible before you build an extractor.
- What is the read pattern? Exploratory analytics, ML feature engineering and periodic reporting suit zero-copy. High-concurrency BI with tight latency SLAs on very large tables usually justifies a materialized gold layer.
- What history do you need? If regulatory or model requirements exceed what the data product retains, plan a snapshot or replication path for that domain only.
- Where do the authorization rules live? If SAP authorizations must be mirrored at row level, budget explicit policy mapping into Unity Catalog. That mapping is work in either pattern.
- Is your current extraction compliant? Audit every pipeline for ODP-RFC usage. Anything that depends on it is a migration item regardless of which pattern you prefer.
The hybrid reference architecture
In practice, the pattern that holds up looks like this:
- Bronze (shared, not copied): SAP data products mounted in Unity Catalog through BDC Connect. No ingestion jobs, no landed files.
- Silver (computed in Databricks): joins of SAP data with non-SAP sources such as CRM, IoT and clickstream, ingested through Lakeflow Connect or your existing tooling. Databricks has published this exact SAP-plus-Salesforce pattern.
- Gold (materialized where justified): performance-critical aggregates written to your own Delta tables, with clustering you control. These are documented, deliberate copies, not shadow pipelines.
- Write-back: model outputs shared back to BDC so planners in SAP Analytics Cloud act on them without another integration.
- Exception lane: a compliant replication path (Replication Flows, ODP-OData or non-ODP CDC) for custom objects and deep history outside the catalog.
This keeps one source of truth for SAP semantics. It also accepts that some physical copies are an engineering choice, not a failure of the strategy.
The Bottom Line
Zero-copy has moved SAP-on-Databricks from a pipeline problem to a product-catalog problem. The questions for 2027 roadmaps are no longer which extractor to license. They are which domains the data-product catalog covers, where you need physical control for performance or history, and how you unify authorization across two catalogs.
Start with an ODP-RFC audit of every existing pipeline. Then map each SAP domain against the five questions above. The result will rarely be all zero-copy or all replication. It will be a deliberate hybrid, and that is the architecture worth defending to your board.
Frequently Asked Questions
What is zero-copy SAP data sharing with Databricks?
It is live, governed access to SAP Business Data Cloud data products from Databricks over the open Delta Sharing protocol. Databricks reads the data where SAP stores it, so no extraction pipeline or second physical copy is needed.
Is zero-copy cheaper than replication?
Not automatically. It removes pipeline, storage and reconciliation costs, but adds BDC licensing and query-time compute on every read. For read-light analytics and AI it is usually cheaper. For heavy, repeated scans, a materialized gold layer often wins.
Does SAP Note 3255746 stop us from replicating SAP data into Databricks?
No. It prohibits third-party use of the ODP Data Replication API over RFC. Replication through SAP Datasphere, ODP-OData, or extraction methods that avoid ODP-RFC remains possible.
Do we need SAP Databricks, or can we use our existing Databricks workspace?
BDC Connect links SAP Business Data Cloud to an existing Databricks workspace on any major cloud. SAP Databricks is a separate, SAP-embedded and SAP-billed option. Teams with a mature lakehouse typically evaluate BDC Connect first.
Can custom Z-fields be shared via zero-copy?
Yes, if the underlying CDS view is extended and the data product is activated or reinstalled afterwards. Fully custom tables need a custom data product or a separate replication path.
Can Databricks write results back to SAP?
Yes. BDC Connect is bi-directional, so Databricks tables can be shared back to Business Data Cloud for use in Datasphere and SAP Analytics Cloud.



