Drop default min resources for fleets

Runs and fleets have non-zero min resources by default, cpu: 2.., memory: 8GB.., disk: 100GB..:

dstack/src/dstack/_internal/core/models/resources.py

Lines 377 to 384 in f7dacf7

    
           class ResourcesSpec(generate_dual_core_model(ResourcesSpecConfig)): 
        
               # TODO: Remove Range[int] in 0.20. Range[int] for backward compatibility only. 
        
               cpu: Annotated[Union[CPUSpec, Range[int]], Field(description="The CPU requirements")] = ( 
        
                   CPUSpec() 
        
               ) 
        
               memory: Annotated[Range[Memory], Field(description="The RAM size (e.g., `8GB`)")] = ( 
        
                   DEFAULT_MEMORY_SIZE 
        
               )

This makes sense for runs to avoid provisioning runs on tiny instances by default and failing, e.g. due to OOM. Having min resources on fleets gets in the way when the instance is provisioned for a run. Consider a case:

Create a default fleet with no resources.
Create a k8s cluster with nodes having 100GB disk or 8GB RAM (or less *) and configure a k8s backend.
You'll get no offers even if specifying lower resources for runs because the fleet min resource filter out the offers.
You have to recreate the fleet or create a new fleet.

100GB disk and 8GB RAM will translate to less available resources on k8s, so the exact lower bound will filter out such nodes.

We may still apply the same min resource to fleets if they are not specified for provisioning instances without run (non-elastic fleets, maintaining nodes.min, etc).

This issue is stale because it has been open for 30 days with no activity.

Analysis: Are min resources needed for fleet offer generation?

Short answer: No. Backends work fine without min resource constraints. Here's why:

Offline providers (AWS, GCP, Azure, Lambda, Nebius, OCI, RunPod) — No issue

dstack backends extend ComputeWithAllOffersCached and call gpuhunt.default_catalog().query(). For offline
providers, this just filters a pre-fetched catalog.

Kubernetes — No issue

K8s backend works fine with no constraints but it won't be used for fleet provisioning anyway.

Online providers (Cudo, VastAI, etc.) — Minor concern only

Cudo has a nested loop generating CPU × GPU × memory combinations for API queries. With low minimums and high
machine capacity, this could create many combinations. But Cudo has its own internal MIN_CPU=2 and MIN_MEMORY=8 fallback defaults when filter values are None.

VastAI uses filter parameters as API hints. With low minimums, it just fetches a broader set and post-filters locally.

Disk

Disk is safe to set to None in fleet resources because there are multiple fallback layers. choose_disk_size_mib() falls back to DEFAULT_DISK.size (100GB) when requirements_disk_size is None.

Conclusion

We can safely drop default min resources for fleets entirely — no need to keep them even for non-elastic fleets.

But will we require resources if user set nodes>0 in fleet config

@peterschmidt85

@peterschmidt85, no, there is no problems with provisioning nodes>0 fleets without resources – dstack will provision the cheapest/smallest offer available. See the example offers in #3776 (comment)

Navigation Menu

Repository navigation

Drop default min resources for fleets #3585

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	class ResourcesSpec(generate_dual_core_model(ResourcesSpecConfig)):
	# TODO: Remove Range[int] in 0.20. Range[int] for backward compatibility only.
	cpu: Annotated[Union[CPUSpec, Range[int]], Field(description="The CPU requirements")] = (
	CPUSpec()
	)
	memory: Annotated[Range[Memory], Field(description="The RAM size (e.g., `8GB`)")] = (
	DEFAULT_MEMORY_SIZE
	)

Drop default min resources for fleets #3585

Description

Activity

github-actions commented on Mar 21, 2026

r4victor commented on Apr 14, 2026

peterschmidt85 commented on Apr 14, 2026

r4victor commented on Apr 14, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions