An integration runtime is the compute environment that executes integration flows: it runs the connectors, mappings, transformations and routing logic that a designer defines, and it supplies the network position, credentials and processing capacity those flows need. In automation and system integration, the runtime is the part that actually does the work, while the design tool, the flow definition and the management console only describe and supervise it.
The distinction matters because the same integration logic can behave very differently depending on where the runtime sits. A runtime hosted by a cloud provider reaches public SaaS APIs easily but cannot see a database behind a corporate firewall, whereas a self-hosted runtime installed inside the private network can. Choosing, sizing and securing the runtime is therefore an architectural decision, not a deployment detail.
How an integration runtime executes a flow
An integration runtime separates the control plane from the data plane. The control plane, usually a cloud management service, stores flow definitions, schedules, credentials references and monitoring data. The data plane is the runtime itself: it receives a deployed flow, pulls or receives the trigger, opens connections, processes records and reports status back.
A typical execution cycle follows a fixed sequence. The runtime registers with the control plane through an outbound, encrypted channel, so no inbound firewall port is usually required. When a trigger fires, whether a schedule, a webhook, a polling interval or a message on a queue, the runtime loads the flow version, resolves secrets from a vault, executes each step and writes execution logs and metrics that operators read in the console.
The following table summarizes the main components found in most runtime implementations and what each one contributes.
| Component | Role in execution | Typical failure symptom |
|---|---|---|
| Execution engine | Interprets the flow and runs steps in order | Flows stall or exceed time limits |
| Connector library | Speaks each target protocol or API | Authentication or schema errors |
| Agent or gateway | Maintains the channel to the control plane | Runtime shown as offline |
| Secret resolver | Fetches credentials at run time | Expired or missing credentials |
| Buffer and retry store | Holds messages during outages | Duplicates or lost records |
| Telemetry module | Emits logs, metrics and traces | Blind spots during incidents |
Types of integration runtime by hosting model
Integration platforms usually offer three hosting models, and many organizations run more than one at the same time. The choice depends on where the data lives, which compliance rules apply and who is expected to operate the infrastructure.
| Hosting model | Where it runs | Best suited for | Main trade-off |
|---|---|---|---|
| Cloud-managed runtime | Vendor infrastructure | SaaS-to-SaaS flows and public APIs | No direct reach into private networks |
| Self-hosted runtime | Customer servers or virtual machines | On-premises databases and file shares | Customer patches, scales and monitors it |
| Hybrid runtime | Cloud control plane with local workers | Mixed estates with sensitive data | Two operational models to manage |
| Containerized runtime | Kubernetes or a container service | Elastic workloads and DevOps teams | Requires container platform skills |
Some platforms add specialized runtimes for particular workloads, for example a runtime dedicated to executing legacy SQL Server Integration Services packages, or one optimized for large data movement with parallel copy. These variants share the same principle: the runtime provides compute and network position, and the flow definition remains portable across them as far as the connectors allow.
Integration runtime compared with related concepts
The integration runtime is often confused with the platform, the connector or the middleware pattern it executes. The runtime is infrastructure, not architecture: an enterprise service bus, an iPaaS product or a point-to-point integration each need some runtime to execute, but none of them is identical to it.
| Concept | What it is | Relationship to the runtime |
|---|---|---|
| iPaaS | A cloud integration platform delivered as a service | Provides managed runtimes plus design and monitoring tools |
| Integration connector | A packaged adapter for one system | Loaded and executed by the runtime |
| Enterprise service bus | A centralized mediation architecture | Runs on its own runtime nodes |
| Message queue | Durable storage for asynchronous messages | Feeds triggers to the runtime or buffers its output |
| API orchestration | Coordinating several API calls into one process | A workload the runtime executes |
A useful mental test is to ask what would break if the runtime were stopped. Flow definitions, mappings and connector configurations would survive in the platform repository, but no data would move. That asymmetry is why runtime availability is usually the first metric that integration operations teams watch.
Choosing and sizing an integration runtime
Runtime selection starts with data location and network reachability, followed by volume, latency and compliance requirements. A design that ignores network position tends to fail late, during testing, when a cloud runtime cannot resolve an internal host name.
| Situation | Recommended runtime | Reason |
|---|---|---|
| Both endpoints are public SaaS | Cloud-managed | No infrastructure to operate |
| Source database behind a firewall | Self-hosted near the source | Outbound-only connectivity, no open ports |
| Data must stay in one jurisdiction | Runtime pinned to that region | Keeps processing within residency rules |
| Large nightly batch transfers | Scaled-out or multi-node runtime | Parallel copy shortens the batch window |
| Spiky event-driven traffic | Containerized with autoscaling | Capacity follows demand |
Sizing depends on concurrency, payload size and transformation complexity rather than on record counts alone. A flow that parses large XML documents or performs lookups for every record consumes far more memory and CPU than a simple file copy of the same volume. Practitioners load-test with realistic payloads and plan for at least two nodes when the integration supports a business-critical process, so that patching or a node failure does not stop the flow.
Limitations, risks and common mistakes
Most runtime incidents come from operational neglect rather than design flaws. Self-hosted runtimes in particular are easy to install and easy to forget, which leaves outdated versions, expiring certificates and single points of failure in production.
- Single-node deployment for a critical flow, so one reboot interrupts the business process.
- Version drift between runtime and control plane, which disables newer connector features or blocks deployment.
- Over-privileged service accounts that give the runtime broader database or file access than the flows need.
- Unmonitored queues and retry stores, where failed messages accumulate silently until a backlog surfaces downstream.
- Shared runtimes without isolation, where one heavy batch flow starves latency-sensitive real-time flows.
Mitigations are well established: run high-availability node groups, apply least-privilege credentials stored in a vault, separate batch and real-time workloads onto different runtimes, and alert on runtime heartbeat, queue depth and error rate. Treating the runtime as production infrastructure, with change management and capacity reviews, prevents most surprises.
In simple terms
An integration runtime is the engine that does the actual moving of data between systems. Someone draws up the instructions, such as take new orders from the shop and put them into the accounting system, but the instructions alone do nothing. The runtime is the worker that carries them out, every time they are needed.
Where that worker stands matters. A worker based in the cloud can reach other cloud services easily, but cannot walk into a locked office building. If the data is inside the company network, the worker has to be placed inside too, which is what a self-hosted runtime does.
A practical example
A regional distributor keeps its inventory in an on-premises SQL database and sells through a cloud commerce platform. The integration team designs a flow that sends stock levels to the storefront every 15 minutes. A cloud-managed runtime cannot reach the internal database, so the team installs a self-hosted runtime on two virtual machines inside the data center, both registered to the same cloud control plane through outbound HTTPS.
The two nodes share the load and cover for each other during monthly patching. A separate cloud-managed runtime handles order notifications between the storefront and the shipping provider, because both are public services. After go-live, stock discrepancies on the storefront fell from several per day to almost none, and no inbound firewall port had to be opened.