Baseten adds autoscaling schedules to cut cold-start delays
Baseten has introduced autoscaling schedules, a feature that raises the minimum number of replicas ahead of predictable traffic so that requests do not wait on cold starts. Schedules can recur, such as every weekday morning, or run once, for example before a product launch.
- Baseten's autoscaling schedules raise the minimum number of replicas before predictable traffic arrives.
- Schedules can recur (for example weekday mornings) or run once (for example ahead of a launch).
- Up to 10 schedules are allowed per environment, with one shared timezone.
- Windows cannot overlap and must last at least five minutes.
- Configuration is available through the dashboard, CLI, or Management API.
What cold starts are and why they matter
Autoscaling adjusts the number of replicas to match incoming traffic. When demand rises, new replicas need time before they can handle requests. Baseten calls this the cold-start time, and says it includes allocating compute resources, starting the serving software, and loading the model.
Requests that wait during a cold start have longer end-to-end response times, according to Baseten. Autoscaling schedules address this by starting additional replicas before the traffic arrives, rather than reacting after demand has already increased.
Scheduling around known demand
Baseten points out that many workloads have predictable traffic. Autoscaling schedules let users change their autoscaling settings at set times, so they do not need to maintain an external cron job to do it.
The company describes two main use cases:
- Recurring peaks: raise the minimum number of replicas during work hours and lower it overnight. Baseten gives the example of a meeting-transcription app that scales up before meetings begin and back down after the peak.
- Planned events: schedule a temporary capacity increase for product launches, load tests, or migrations to a new deployment.
How the schedules are configured
Each schedule sets a time window along with the minimum and maximum number of replicas for that window. The minimum is the lowest number of replicas the autoscaler keeps running during the window. The maximum limits the total number of replicas the autoscaler can run. When no schedule is active, the environment uses its default settings.
Baseten lists the following rules:
- Up to 10 schedules can be created per environment, all using one shared timezone.
- Windows can be daily, hourly, or one-time.
- Windows cannot overlap and must last at least five minutes.
- Schedules can be set up through the dashboard, the CLI, or the Management API.
What it means for users
For teams running workloads with known traffic patterns, the feature offers a built-in way to prepare capacity in advance instead of relying on reactive scaling alone. Baseten says that starting additional replicas before an expected surge can help keep end-to-end response times consistent as request volume increases. The company directs users to its setup guide, and offers to discuss specific workloads with its engineers.
