Analytics Legends logoAI & Analytics Legends Sign inSign up
Guide

SAP Datasphere replication flow: how to run one you can explain

As of 2026-11-17

A SAP Datasphere replication flow copies data from a source into Datasphere (or another target) and keeps it current through a delta load. Building one takes minutes; keeping it trustworthy depends on a layer the flow does not own, because on SAP sources the changes are published by the operational data provisioning framework and held in a delta queue. A flow that reports success has only proved that it ran, not that the source published every change.

What a replication flow is, and what it is not

A replication flow is a Datasphere object that moves data from a source connection to a target connection in bulk and, if you ask for it, keeps following changes. SAP documents three load types: initial only (copy what exists at the start, then stop), initial and delta (copy, then keep replicating changes at an interval or on a schedule), and delta only (skip the starting data and follow changes). Which of those you can choose depends on the connection: some sources, such as object stores and Snowflake, are listed by SAP as initial-load-only sources, so check the connection table before you promise a customer a near-real-time feed.

It is not a modelling tool. A flow lands tables; views, analytic models and access rules are built on top of them afterwards. It is also not the same thing as a remote table with replication switched on. A replication flow is built for large volumes and many objects in one run, with delta handling and the option to write to non-SAP targets. A remote table is a single-object convenience, and mixing the two for the same source is how teams end up with two copies and two sets of numbers.

Targets are the other half of the design. Inside Datasphere the flow writes to the HANA database or, for cheaper bulk storage, to the object-store file layer. Outside Datasphere it can write to SAP and non-SAP targets, which is the reason many programmes adopt it in the first place: one tool to feed the lakehouse as well as the semantic layer.

What sits underneath the flow on an SAP source

Two components, and they do different jobs. Operational data provisioning is the publishing surface. Source systems expose data through enabled providers, in this order of preference for new work: CDS views annotated for data extraction, standard Business Suite extractors in finance, controlling, materials management and sales, table-level change capture through SLT for data with no extractor, and legacy BW datasources as a last fallback. Consumers subscribe to those providers and the framework routes the changes.

The operational delta queue is the buffer. Change events wait there until a consumer reads them, and several consumers can read the same queue independently. Persistence, sequencing and resilience to a consumer being down all live in the queue, not in the flow.

That is why a replication flow can complete and deliver nothing. It read the queue; the queue was empty; it reported success. Flow status says the flow executed. Only the queue says whether the source published.

SAP's learning material adds a rule that catches people out: for ABAP sources the flow accepts CDS views enabled for extraction and tables with a primary key, and CDS views or providers without a primary key can only be loaded with the initial-only load type, because delta is not available for them. A requirement for a delta feed on a keyless object is therefore a source-modelling task before it is a Datasphere task.

The one shortcut never to take

Do not read source tables directly over a database connection to avoid the extraction framework. It is the most persistent anti-pattern on this platform and it fails in the most expensive way: the initial load is correct, so the flow is trusted and promoted, and only the deltas are silently missing. A direct table read bypasses the change-tracking framework entirely.

The discovery point is month-end close, when actuals do not match. By then the model has fed reports for weeks, and the reconciliation is no longer technical but a question of credibility with finance.

The price of doing it properly is an operational surface to watch: queue depth, error rate and subscriptions. It is the honest trade, and it is the only route that gives resilient, multi-consumer, delta-safe replication. There is a second reason to stay on the annotated-view route. A CDS view with the extraction annotation is at once a source for Datasphere and a candidate for the SAP knowledge graph that grounds Joule, whereas a table read or a table-level capture produces a working pipe with no semantic surface. A legacy datasource or a direct read keeps the data moving but leaves nothing for an agent to reason about, and it will have to be rebuilt later.

A Saturday night failure, read in two screens

A flow feeding a finance data product fails at 23:47 on a Saturday. The on-call consultant has two artefacts to read before touching anything: the queue monitor view for the affected subscription, and the extractor's exception log in the source system.

In this worked case, the queue depth climbed from around 200 events to just over 14,000 across six hours, with zero events consumed. That signature means a downstream consumer stopped pulling; it does not mean extraction failed. The exception log, read second, was clean: the source had emitted every change correctly. The fault sat entirely on the consuming side, so restarting the source jobs would have done nothing and might have hidden the evidence.

The method generalises into a short triage order. Is the queue growing? If yes, the consumer is down or throttled. Is the queue flat and empty while business users report missing data? Then the source did not publish, and the exception log is where to look. Is the queue healthy and the numbers still wrong? Then the problem is in modelling or in the target, and the flow is innocent. Three questions, two screens, and no guessing at which team to wake up.

What to alert on, and what the thresholds really are

Queue depth. Our working band is healthy below roughly a thousand pending events; growth past 10,000 events without a consumer pulling indicates an upstream backlog or a consumer that is down, and it is visible days before anyone sees a wrong number. The 10,000 figure is from the platform's extraction card; the thousand-event floor is our own field guidance.

Extraction error rate. Under about 0.01 percent is our working band. Above it, something fails intermittently, and intermittent failures are the ones that surface at reconciliation rather than in an alert.

Subscription hygiene. A consumer that was deliberately de-registered and one that was simply forgotten look identical from the queue's side, and only one of them is fine. A quarterly audit stops the queue bloating on behalf of a project that ended two years ago.

Thread settings. Replication parallelism is a capacity decision, not a detail. The platform's sizing card warns that the default of eight threads per flow, multiplied across dozens of concurrently scheduled flows, can saturate the input and output allotment even when raw compute has headroom. The symptom looks like a capacity shortfall and is really a concurrency setting. SAP documents that thread limits and delta frequency can be changed without redeploying the flow, and that initial loads can parallelise up to a maximum of 16 jobs, so tune both deliberately rather than leaving the defaults to collide.

Treat all of these as starting points. None is an SAP-published service level, and your own baseline after four weeks of normal traffic is a better threshold than ours.

Capacity: what a replication flow costs you

Replication flows draw a sustained and largely predictable load, unlike live analytics queries, which spike at peak business hours. The two compete for the same pool in a tenant, and a tenant sized on average utilisation will queue month-end close queries behind background flows. Queuing is the worse failure mode for users, because it looks like the platform is simply slow rather than failing with a clear error.

SAP's service description guide v10-2026 prints no tenant floor of that kind, so entry sizing goes through SAP and its capacity estimator. Replication flows are one of six workload types that consume the same pool, so a sizing proposal should show the replication line separately from SAC concurrency and Joule volume.

Writing to non-SAP targets has its own commercial line. SAP's community guidance on replication flows describes a premium outbound integration allocation that an administrator must assign before a non-SAP target can be used. SAP does not publish a single public price for it, so do not let a proposal quote a number you have not seen in the customer's own order form. Capacity-unit choices are budget choices; loop in finance before you size, not six months after go-live.

When not to replicate at all

Replication is a choice, not a default. Datasphere federates as a first-class capability: a live query passes through with nothing copied. The case for moving data is therefore made object by object, on latency, source-system load or source availability.

Our working break-even is around 100 gigabytes of facts, or more than fifty concurrent business-intelligence users. Below that, federation usually wins on total cost and freshness; above it, repeatedly paying the source's live-query cost adds up in money and in patience. These are our own planning figures, not an SAP threshold.

One cost hides on the federation side: querying an external platform in place consumes compute on that platform's bill. A federation-heavy design can look cheaper in SAP capacity units and be more expensive overall, and nothing in the capacity-unit conversation surfaces it.

Decision grid, one line per case. Governed finance data from S/4HANA, deltas required — Choice: replication flow on an annotated CDS view · Watch: queue depth and error rate. Table with no extractor, change tracking needed — Choice: replication flow on table-level capture · Watch: it stays outside the knowledge graph until a semantic layer exists. Data in a partner's Delta table or an object-store lake — Choice: federate or share in place · Watch: compute lands on the other platform's bill. Large facts, many concurrent users, stable schema — Choice: replicate · Watch: capacity-unit load and thread limits. Keyless object, delta wanted — Choice: fix the source model first · Watch: initial-only is the only load type available.

In a hybrid estate, with ECC and S/4HANA side by side for the three to five years a transition typically takes, two-way replication is one of three ways to harmonise master data, next to federated read and a master-data governance hub. Replication gives both systems local, low-latency data but is operationally heavy and fragile when schemas diverge. Write down which of the three you chose and why.

What we cannot assert

The queue-depth floor (about a thousand events), the error-rate band, the 100 GB and fifty-user break-even are our own field guidance, not SAP service levels. The default delta check interval, per-flow object limits and the price of premium outbound integration could not be verified on SAP's JavaScript-rendered help pages and are therefore not stated.

Frequently asked

Why did my SAP Datasphere replication flow succeed but load nothing?

It read an empty delta queue. The flow depends on the source side publishing change events; if an extractor failed or a source job did not run, the flow pulls zero rows and correctly reports success. Check the queue depth and the extractor exception log before changing anything in Datasphere.

What load types does a replication flow support?

SAP documents three: initial only, initial and delta, and delta only. What you can actually use depends on the source and target connection, and some sources are initial-load-only. Objects without a primary key, such as certain CDS views and providers, can only run as initial only.

Can I replicate a table that has no standard extractor?

Yes. Table-level change capture through SLT covers that case and plugs into the same change-tracking framework. What you should not do is read the table directly over a database connection, which loads the initial set and silently misses every delta.

What should we monitor on a replication flow?

The source side more than the flow: queue depth (past 10,000 events with no consumer pulling means a backlog or a stopped consumer), extraction error rate (our working band is below roughly 0.01 percent), and a quarterly audit of which consumers are still subscribed. Add thread settings if many flows run at once.

Should we replicate or federate?

Federate by default and justify each replication in writing. Past roughly 100 GB of facts or fifty concurrent business-intelligence users, replication is usually cheaper overall in our planning model. Remember that a federated query consumes compute on the source platform's own bill.

When is a SAP Datasphere replication flow the wrong tool?

When federation answers the question, when the object has no key and you need deltas, or when the source cannot act as a delta source. Each of those is cheaper to discover in design than at month-end close.

Engage the editor

This guide is written by . 27 years of SAP — Datasphere, Business Data Cloud, SAC Planning, Joule. Freelance or permanent · Hybrid / remote · Anywhere in EMEA.

Profile and contact

What this page is built on

External sources

Read next

Alert me to new SAP Datasphere missions

No account needed. We first ask you to confirm by e-mail; nothing is sent without that confirmation, and every alert has a one-click unsubscribe link. Privacy

Open missions

120 live roles and contracts in SAP Datasphere right now — see the SAP Datasphere missions · freelance SAP contracts.