Scaling Apps - Cerebrium

Cerebrium’s scaling system automatically manages computing resources to match app demand, from single requests to multiple simultaneous requests, optimizing for both performance and cost.

How Autoscaling Works

The scaling system monitors a configurable metric - such as concurrency utilization, requests per second, CPU usage, or memory usage - and compares it against a target threshold. When the metric exceeds the target, new instances start within seconds. See Scaling Metrics for details on each option.

Scaling Configuration

The cerebrium.toml file controls scaling behavior through several key parameters:

[cerebrium.scaling]
min_replicas = 0           # Minimum running instances
max_replicas = 3           # Maximum concurrent instances
cooldown = 60              # Cooldown period in seconds
replica_concurrency = 1    # The maximum number of requests each replica of an app can accept

Minimum Instances

The min_replicas parameter defines how many instances remain active at all times. Setting this to 1 or higher maintains warm instances for immediate response, eliminating cold starts but increasing costs. Use this for apps requiring consistent response times or specific SLA guarantees.

Maximum Instances

The max_replicas parameter sets an upper limit on concurrent instances, controlling costs and protecting backend systems. When traffic increases, new instances start automatically up to this configured maximum.

Cooldown Period

The cooldown parameter specifies the time window (in seconds) that must pass at reduced concurrency before an instance scales down. This prevents premature scale-down during brief traffic dips that might be followed by more requests. A longer cooldown period helps handle bursty traffic patterns but increases instance running time and cost.

Replica Concurrency

The number of requests an app instance can handle concurrently is dictated by the replica_concurrency parameter. This is a hard limit, and an individual replica will not accept more than this limit at a time. By default, once this concurrency limit is reached on an instance and there are still requests to be processed in-flight, the system will scale out by the number of new instances required to fulfil the in-flight requests. For example, if replica_concurrency=1 and there are 3 requests in flight with no replicas currently available, Cerebrium will scale out 3 instances of the application to meet that demand.

Typically most GPU applications will require that replica_concurrency is set to 1. If the workload requires GPU but higher throughput is desired, replica_concurrency may be increased so long as access to GPU resources is controlled within the application through batching.

Processing Multiple Requests

Apps can process multiple requests simultaneously through batching and concurrency. Cerebrium supports frameworks with built-in batching and enables custom implementations through the custom runtime feature. See the Batching & Concurrency Guide for details.

Instance Management

Cerebrium automatically restarts failed instances, starts new instances to maintain capacity, and monitors instance health continuously. Apps requiring maximum reliability combine several scaling features:

[cerebrium.scaling]
min_replicas = 2              # Maintain redundant instances
cooldown = 600                # Extended warm period
max_replicas = 10             # Room for traffic spikes
response_grace_period = 1200  # Maximum request lifespan ensuring graceful exit

The response_grace_period parameter stipulates how long in seconds a request would need at most to finish, and provides time for instances to complete active requests during normal operation and shutdown. During normal replica operation, this acts as a request timeout value. During replica shutdown, the Cerebrium system sends a SIGTERM signal to the replica, waits for the specified grace period, issues a SIGKILL command if the instance has not stopped, and kills any active requests with a GatewayTimeout error.

When using the Cortex runtime (default), SIGTERM signals are automatically handled to allow graceful termination of requests. For custom runtimes, you’ll need to implement SIGTERM handling yourself to ensure requests complete gracefully before termination. See our Graceful Termination guide for detailed implementation examples, including FastAPI patterns for tracking and completing in-flight requests during shutdown.

Performance metrics available through the dashboard help monitor scaling behavior:

Request processing times
Active instance count
Cold start frequency
Resource usage patterns

Using Scaling Metrics

Cerebrium supports multiple scaling metrics beyond the default replica_concurrency. Four scaling metrics are available:

concurrency_utilization
requests_per_second
cpu_utilization
memory_utilization

Specify a metric and target in the cerebrium.scaling section:

[cerebrium.scaling]
min_replicas = 0
cooldown = 600
max_replicas = 10
response_grace_period = 120
replica_concurrency = 1
scaling_metric = "concurrency_utilization"
scaling_target = 100

Concurrency Utilization

concurrency_utilization is the default scaling metric, with a default target of 100%. This metric maintains a maximum percentage of replica_concurrency averaged across all instances. For example, with replica_concurrency=1 and scaling_target=70, Cerebrium maintains 0.7 requests per instance, ensuring 30% excess capacity. With replica_concurrency=200 and scaling_target=80, Cerebrium maintains 160 requests per instance and scales out once that target is exceeded.

Requests per Second

requests_per_second maintains a maximum application throughput measured in requests per second, averaged over all instances. This metric is more effective than concurrency_utilization when application throughput has been benchmarked. It does not enforce concurrency limits, so it is not recommended for most GPU applications. For example, scaling_target=5 maintains a 5 requests/s average across all instances.

CPU Utilization

cpu_utilization scales based on maximum CPU percentage utilization averaged over all instances, relative to the cerebrium.hardware.cpu value. For example, with cpu=2 and scaling_target=80, Cerebrium maintains 80% CPU utilization (1.6 CPUs) per instance. This metric requires min_replicas=1 since scaling relative to 0 CPU units is undefined.

Memory Utilization

memory_utilization scales based on maximum RAM percentage utilization averaged over all instances, relative to cerebrium.hardware.memory. This refers to RAM, not GPU VRAM. For example, with memory=10 and scaling_target=80, Cerebrium maintains 80% memory utilization (8GB) per instance. This metric requires min_replicas=1 since scaling relative to 0GB of memory is undefined.

Keeping a Scaling Buffer

For apps with long startup times or predictable traffic, a replica buffer maintains consistent excess capacity above what the scaling metric suggests. The scaling_buffer option adds a fixed number of extra replicas to the autoscaler’s recommendation. This is available with the following scaling metrics:

concurrency_utilization
requests_per_second

Add scaling_buffer to the cerebrium.scaling section:

[cerebrium.scaling]
min_replicas = 1
cooldown = 600
max_replicas = 10
response_grace_period = 120
replica_concurrency = 1
scaling_metric = "concurrency_utilization"
scaling_target = 100
scaling_buffer = 3

With the above config: when no traffic is received, the app runs 1 replica as a baseline — the buffer scales based on requests. With concurrency_utilization at 100 and replica_concurrency=1, receiving 1 request causes the autoscaler to suggest 1 replica. With scaling_buffer=3, the app scales to (1+3)=4 replicas. The buffer adds a static number of replicas on top of the autoscaler’s recommendation. After the request completes, the cooldown period applies and the replica count scales back to the 1 replica baseline.

Evaluation Interval

Requires CLI version 2.1.5 or higher.

The evaluation_interval parameter controls the time window (in seconds) over which the autoscaler evaluates metrics before making scaling decisions. The default is 30 seconds, with a valid range of 6-300 seconds.

[cerebrium.scaling]
evaluation_interval_seconds = 30  # Evaluate metrics over 30-second windows

A shorter interval makes the autoscaler more responsive to traffic spikes but may cause more frequent scaling events. A longer interval smooths out transient spikes but may delay scaling responses.

For bursty workloads, a shorter evaluation_interval (e.g., 10-15 seconds) helps the system respond quickly to demand. For steady workloads, a longer interval (e.g., 60 seconds) reduces unnecessary scaling churn.

Load Balancing

Requires CLI version 2.1.5 or higher.

The load_balancing_algorithm parameter controls how incoming requests are distributed across your replicas. When not specified, the system automatically selects the best algorithm based on your replica_concurrency setting.

[cerebrium.scaling]
load_balancing_algorithm = "min-connections"  # Explicitly set load balancing algorithm

Default behavior: When load_balancing_algorithm is not set, the system uses first-available for replica_concurrency <= 3 (typical for GPU workloads) and round-robin for higher concurrency.

Available Algorithms

round-robin

Cycles through replicas starting from the last successful target. Each replica’s concurrency limit is respected - if a replica is at capacity, the algorithm proceeds to the next one in rotation.

Characteristic	Value
Selection complexity	O(1) typical, O(N) worst case when scanning for available capacity
Latency profile	Consistent p50, good p90 under uniform load
Strategy	Stateful index rotation with mutex synchronization; skips full replicas

Best for: Workloads with predictable request times where you want even distribution across replicas over time.

first-available

Scans replicas from the start of the list and selects the first one with available capacity.

Characteristic	Value
Selection complexity	O(1) typical, O(N) worst case
Latency profile	Optimal p50 when load is light, may degrade p90 under high load
Strategy	Linear scan from list start; returns first replica that accepts via Reserve()

Best for: GPU workloads with low concurrency (replica_concurrency <= 3). Maximizes utilization of warm replicas before spreading load, reducing cold starts and keeping models in VRAM. Tradeoff: Earlier replicas in the list handle more traffic. This is desirable for GPU workloads but may cause uneven distribution for CPU workloads.

min-connections

Linear scan to find the replica with the fewest in-flight requests, then attempts to reserve it. If that replica cannot accept (at capacity), falls back to trying other replicas in iteration order.

Characteristic	Value
Selection complexity	Θ(N) - always scans all replicas to find minimum
Latency profile	Best p90/p99 tail latency
Strategy	Single pass to find minimum in-flight; fallback in iteration order

Best for: Workloads with variable request times (e.g., LLM inference where output length varies). Routes new requests to the least busy replica, preventing fast requests from queuing behind slow ones.

random-choice-2

Implements the “Power of Two Choices” algorithm: randomly samples two replicas and routes to the one with lower weight (based on active request tracking). Ties are broken randomly.

Characteristic	Value
Selection complexity	Θ(1) - constant time regardless of replica count
Latency profile	Good balance of p50 and p90
Strategy	Sample 2 random replicas, compare weights, pick lighter one

Best for: High-throughput scenarios with many replicas where selection overhead matters. Research shows this achieves exponentially better load distribution than pure random selection. Note: Uses weight-based tracking rather than reservation-based concurrency limiting, making it suitable for unlimited concurrency scenarios.

Choosing an Algorithm

Scenario	Recommended	Reason
GPU inference, `replica_concurrency=1`	`first-available` (default)	Maximizes GPU utilization, keeps models warm
LLMs with variable output lengths	`min-connections`	Prevents head-of-line blocking, best tail latency
High-throughput, many replicas	`random-choice-2`	Θ(1) selection with near-optimal distribution
Uniform request times, even distribution	`round-robin`	Predictable rotation, no hot spots over time
Latency-sensitive with variable load	`min-connections`	Minimizes p90/p99 by routing to least busy replica

Compute Tier

Requires CLI version 2.1.6 or higher.

The compute_tier parameter controls whether your replicas are scheduled on spot (interruptible) or on-demand (protected) instances. This directly affects cost and availability.

[cerebrium.scaling]
compute_tier = "protected"  # Use on-demand instances for higher availability

Available Tiers

Tier	Description
`interruptible`	(Default) Scheduled on preemptible instances. Lower cost, but may be interrupted by the cloud provider during periods of high demand.
`protected`	Scheduled on instances with higher availability with no interruptions, at increased cost.

Choosing a tier:

Use interruptible (default) for batch workloads, development, or cost-sensitive applications that can tolerate occasional interruptions.
Use protected for production services with strict availability requirements or long-running requests where interruption would be costly.

​How Autoscaling Works

​Scaling Configuration

​Minimum Instances

​Maximum Instances

​Cooldown Period

​Replica Concurrency

​Processing Multiple Requests

​Instance Management

​Using Scaling Metrics

​Concurrency Utilization

​Requests per Second

​CPU Utilization

​Memory Utilization

​Keeping a Scaling Buffer

​Evaluation Interval

​Load Balancing

​Available Algorithms

​round-robin

​first-available

​min-connections

​random-choice-2

​Choosing an Algorithm

​Compute Tier

​Available Tiers

How Autoscaling Works

Scaling Configuration

Minimum Instances

Maximum Instances

Cooldown Period

Replica Concurrency

Processing Multiple Requests

Instance Management

Using Scaling Metrics

Concurrency Utilization

Requests per Second

CPU Utilization

Memory Utilization

Keeping a Scaling Buffer

Evaluation Interval

Load Balancing

Available Algorithms

round-robin

first-available

min-connections

random-choice-2

Choosing an Algorithm

Compute Tier

Available Tiers