Depot Container Builds

Depot Container Builds

Depot runs Docker image builds on remote high-performance builders (16 CPU, 32 GB RAM, NVMe SSD cache). depot build is a drop-in replacement for docker build / docker buildx build. depot bake replaces docker buildx bake.

Project Selection for Multi-Org Users

Container build commands target a specific project, not an organization. If expected projects aren't visible or a build unexpectedly prompts to select a project, the current default org may be wrong:

depot org show              # Current org ID
depot org list              # Orgs the user belongs to
depot org switch <org-id>   # Optional: set default org

Key Concepts

• Builds run remotely on ephemeral EC2 instances — images stay in remote cache by default
• Use --load to download to local Docker, --push to push to a registry, --save to store in Depot's ephemeral registry
• Cache is fully automatic on persistent NVMe SSDs — no manual cache config needed
• Multi-platform builds use native CPU builders (no QEMU emulation) for amd64 and arm64 simultaneously
• All team members on a project share the same layer cache

depot build — Essential Patterns

# Build remotely (image stays in remote cache)
depot build -t repo/image:tag .

# Build + download to local Docker daemon
depot build -t repo/image:tag . --load

# Build + push directly to registry (fast — doesn't route through local network)
depot build -t repo/image:tag . --push

# Multi-platform build (native CPUs, no emulation)
depot build --platform linux/amd64,linux/arm64 -t repo/image:tag . --push

# Save to Depot ephemeral registry (7-day default, configurable per repository)
depot build --save .
depot build --save --save-tag my-tag .

# Suppress provenance metadata (fixes "unknown/unknown" platform in registries)
depot build -t repo/image:tag --push --provenance=false .

# Lint Dockerfile before building
depot build -t repo/image:tag . --lint

# Build with secrets
depot build --secret id=mysecret,src=./secret.txt -t repo/image:tag .

# Build with SSH forwarding
depot build --ssh default -t repo/image:tag .

# Specify a Depot project explicitly
depot build --project <project-id> -t repo/image:tag .

Key Flags

depot bake — Multi-Image Builds

Drop-in replacement for docker buildx bake. Builds multiple images in parallel.

depot bake                                    # Default file lookup
depot bake -f docker-bake.hcl                 # Specific HCL file
depot bake -f docker-compose.yml --load       # Build compose services + load locally
depot bake --save --save-tag myrepo/app:v1    # Save to Depot Registry
depot bake --print                            # Print resolved config without building

Default file lookup order: compose.yaml → compose.yml → docker-compose.yml → docker-compose.yaml → docker-bake.json → docker-bake.hcl → docker-bake.override.json → docker-bake.override.hcl

HCL Bake File Example

variable "TAG" {
  default = "latest"
}

group "default" {
  targets = ["app", "worker"]
}

target "app" {
  dockerfile = "Dockerfile"
  platforms  = ["linux/amd64", "linux/arm64"]
  tags       = ["myrepo/app:${TAG}"]
  args = { NODE_VERSION = "20" }
}

target "worker" {
  dockerfile = "Dockerfile.worker"
  tags       = ["myrepo/worker:${TAG}"]
  contexts   = { app = "target:app" }  # Share base between targets
}

Override variables: TAG=v2.0 depot bake

Docker Compose with Per-Service Project IDs

services:
  api:
    build:
      dockerfile: ./Dockerfile.api
      x-depot:
        project-id: abc123
  web:
    build:
      dockerfile: ./Dockerfile.web
      x-depot:
        project-id: def456

Docker Compose Integration

# Preferred: build all services in parallel, then load
depot bake -f docker-compose.yml --load
docker compose up

# Alternative: zero code change (less efficient, each service = separate build)
depot configure-docker
docker compose build

Migration from Docker

# docker build → depot build (same flags, one-line swap)
depot build -t my-image .

# docker buildx bake → depot bake
depot bake -f docker-bake.hcl

# Zero code change via Docker plugin
depot configure-docker
docker build .  # Routes through Depot (look for [depot] prefix in logs)

When migrating, remove these flags — Depot handles caching automatically:

• --cache-from type=gha — causes "services aren't available" errors
• --cache-to type=gha — same issue
• Any manual BuildKit cache configuration

Common Mistakes

Builder Sizes

Billed per-second. Large and Extra Large builders consume your included build minutes faster (via the multiplier) in addition to costing more per minute.

Build Autoscaling

Autoscaling is disabled by default, so Depot routes a project's builds for each architecture to one builder. Enable horizontal autoscaling in the project settings when concurrent builds need separate builders. The settings form preselects 16 builds per instance; Depot starts another builder when concurrent builds reach the configured limit.

Each builder has an independent cache. Heavy autoscaling can reduce cache efficiency, so try a larger builder first when one build needs more CPU or memory. Autoscaling is available on all plans without an additional fee.

One depot bake invocation counts as one build for autoscaling, regardless of its target count. Separate project IDs create separate build requests.

Depot Registry

An OCI-compliant registry that works both as a lightweight ephemeral store for build output and as a full primary registry for all your images (including ones not built with Depot). Each org gets its own subdomain {orgId}.registry.depot.dev, backed by a global CDN, and stores images and OCI artifacts (Helm charts, AI models) up to 50 GB. Browse repositories, tags, and attestations in the Registry Explorer in the dashboard.

# Save a Depot build's output to the registry (project-scoped, no tagging step)
depot build --save -t myapp .

# Push any image to Depot Registry as a primary registry
docker push {orgId}.registry.depot.dev/your/app:v1

# Pull a saved image (by build ID or by tag)
depot pull --project <id> <build-id>
depot pull --project <id> my-image:tag
depot pull --project <id> --platform linux/arm64 <build-id>   # pull a specific platform

# Generate a short-lived (1-hour) read-only token to pull from the Depot Registry
depot pull-token --project <id>
depot pull-token --project <id> <build-id>                    # scoped to a specific build

# Push saved image to another registry
depot push --project <id> -t registry/image:tag <build-id>

# Docker auth for Depot Registry (each org has its own subdomain)
docker login {orgId}.registry.depot.dev -u x-token -p <depot-token>
# Registry URL: {orgId}.registry.depot.dev/your/app:tag
# (the legacy registry.depot.dev/<project-id>:<build-id> form still works for existing usage)

Image retention defaults to 7 days but is configurable per repository from the Registry Explorer, either by age (1, 7, 14, or 30 days) or by tag count (5, 10, 25, 50, or 100 tags).

Pull-through cache: Depot Registry can proxy an external registry and cache its layers on the CDN, so repeat pulls skip the upstream hop (faster CI pulls, global distribution, simpler auth for registries like Google Artifact Registry). Configure an upstream in Registry Settings; supported providers are Docker Hub, GitHub Container Registry, GitLab Container Registry, Quay.io, AWS ECR, Google Artifact Registry, Azure Container Registry, and custom registries. Pull-through must be set when the repository is created, it can't be added to an existing repository.

Special Output Formats

# estargz (lazy-pulling for faster container startup)
depot build --output "type=image,name=repo/image:tag,push=true,compression=estargz,oci-mediatypes=true,force-compression=true" .

# zstd compression (faster Fargate/K8s startup)
depot build --output type=image,name=repo/image:tag,oci-mediatypes=true,compression=zstd,compression-level=3,force-compression=true,push=true .

# SOCI v2 (lazy-loading for faster startup of large images, for example CUDA/PyTorch)
depot build -t repo/image:tag --push --output type=image,soci=true,compression=gzip,force-compression=true,oci-mediatypes=true .

SOCI only indexes gzip layers, so compression=gzip is required and force-compression=true ensures every layer (including base-image layers) is indexed. The index is written alongside unchanged layers in the registry, so there's no separate post-push step; layers under 10 MB are skipped by default (soci-min-layer-size to change). To get the startup benefit, the pulling platform must use the SOCI snapshotter: AWS Fargate does this automatically, and Amazon EKS/EC2 or your own hosts need the soci-snapshotter enabled in containerd.