# Autoscaling {% hint style="warning" %} Autoscaling features are only available for realms using the **Consumption-Based model**. If your realm uses this model, it is highly recommended to activate Autoscaling to benefit from automatic elasticity, which can reduce consumption costs. {% endhint %} ## What is autoscaling Autoscaling is the process of adjusting computational resources based on workload. Pipelines are horizontally scalable, meaning that replicas can be **dynamically created or removed** depending on the number of messages (invocations) in the queue. The Digibee Integration Platform continuously monitors the queue and adjusts the number of replicas to ensure **optimal resource usage** and **efficient processing**, regardless of the trigger type that generated the messages. ## How autoscaling works
Every pipeline execution starts with a **trigger**. Different trigger types, such as [Scheduler](https://app.gitbook.com/s/EKM2LD3uNAckQgy1OUyZ/triggers/scheduling/scheduler), [Event ](https://app.gitbook.com/s/EKM2LD3uNAckQgy1OUyZ/triggers/messaging-and-events/event)or [HTTP](https://app.gitbook.com/s/EKM2LD3uNAckQgy1OUyZ/triggers/web-protocols/http), have unique behaviors and activation criteria. These triggers determine whether a message is sent to the execution queue for further processing or whether it should be skipped or rejected altogether. ### Autoscaling algorithm The autoscaling algorithm scales pipelines **horizontally**. Every **10 seconds**, the Platform monitors the system and can execute the following commands: #### Activation (from 0 to 1) * If the **minimum number of replicas is set to zero**, the pipeline can receive messages but has no replicas running. * The first replica is deployed when the **first message arrives**, which can **add latency**. * The minimum startup time is **20 seconds**. * Once initialized, the replica processes the first and subsequent messages. {% hint style="info" %} The number and type of connectors in the pipeline can impact startup time. {% endhint %} #### Scale up (from 1 to N) When the number of messages in the queue exceeds **70% of total concurrent executions**, the Platform deploys **additional replicas**. #### Scale down (from N to 1) The Platform **reduces replicas** when fewer resources are needed. The system calculates the required number of replicas using the formula below: {% code fullWidth="true" %} ``` Replicas = [(Message rate per second * Response time in seconds) / Total concurrent executions)] ``` {% endcode %} #### Scale-to-Zero (Cooldown) If the **minimum replica is zero** and a single replica is idle for **more than 60 seconds**, the pipeline scales **down to zero**. New replicas are deployed only when new messages arrive. * **Custom Cooldown:** You can configure a custom cooldown time between 60 and 300 seconds in the pipeline’s Advanced Settings. This defines how long the system waits before scaling back to zero after idle. ## Autoscaling on the Digibee Integration Platform The autoscaling process on the Digibee Integration Platform follows these steps: ### Phase 1: Activation of the first replica 1. The system executes an **Activation command** to create and start the **first replica**. 2. This command triggers the processing of the **first message in the queue**. 3. Components for pipeline processing are initialized only when the replica is deployed, that is, when messages arrive. {% hint style="info" %} Since the last step can add latency to the first message’s total response time, it’s recommended to set **min replicas ≥ 1** for latency-sensitive pipelines. {% endhint %} ### Phase 2: Processing of messages 1. After the first replica is deployed and initialized, the Platform processes the first set of messages according to the concurrency settings. 2. When **new messages arrive:** * They are processed by available consumers if capacity allows. * Otherwise, a [Scale Up command](#scale-up-from-1-to-n) starts a **new replica**. {% hint style="info" %} Scale Up impact is smaller on total response time because it triggers only when the queue reaches 70% of concurrent execution capacity. {% endhint %} **Total processing capacity formula:** ``` Total capacity = Replicas × Simultaneous executions per replica ``` ### Phase 3: Scaling down * If the queue has only a few messages, the Platform reduces the number of running replicas until only one remains. * If the single replica remains idle for **more than 60 seconds**, a [Cooldown command](#scale-to-zero-cooldown) scales the pipeline down to zero replicas. * New replicas are deployed only when the next set of messages arrives. ## Autoscaling strategies There are two types of autoscaling deployment strategies on the Digibee Integration Platform: **Scale-to-Zero** and **Always On**. ### Scale-to-Zero The pipeline is deployed with autoscaling enabled and a **minimum of zero replicas**. Pipelines can receive new requests but operate with “zero replicas”, incurring no cost until the first message is received. #### Recommended usage scenario This strategy is ideal for workloads that **don’t require immediate response** for the first message. For example, a pipeline that runs once a day can remain inactive most of the time. See more scenarios where Scale-to-Zero is applicable: * Scheduled pipelines using the [Scheduler trigger](https://app.gitbook.com/s/EKM2LD3uNAckQgy1OUyZ/triggers/scheduling/scheduler). * Low-criticality workloads where the first message can tolerate extra latency. * Workloads aiming to maximize cost savings during idle periods. ### Always On The pipeline is deployed with autoscaling enabled and **a minimum of 1 or more replicas**. At least one replica is always running, ready to process the first messages immediately. #### Recommended usage scenario This strategy ensures **low-latency response** even during low traffic periods. Always On is applicable on the following scenarios: * [API-triggered pipelines](https://app.gitbook.com/s/EKM2LD3uNAckQgy1OUyZ/triggers/web-protocols/api). * High-criticality workloads where extra response time for the first message is not acceptable. {% hint style="success" %} **Optimization tip:** Replace a Large pipeline with 1 replica and no autoscaling with a Small pipeline with autoscaling enabled, using 1 to 4 replicas. {% endhint %}