Immich on GKE Autopilot

Immich is an open-source, AGPL-3.0-licensed self-hosted photo and video management platform — a Google Photos alternative with mobile auto-backup, a timeline, albums, sharing, CLIP-powered smart search, and face recognition. Every feature is free: the optional $99 product key is a supporter badge only, with zero feature gating. This module deploys Immich on GKE Autopilot on top of the App_GKE foundation, which provisions and manages the shared Google Cloud and Kubernetes infrastructure.

There is deliberately no CloudRun variant of this module. Immich's media library is a local filesystem — it has no S3/GCS storage backend — and photo/video uploads are routinely multi-GB, neither of which fits Cloud Run's request model.

This guide focuses on the cloud services Immich uses and how to explore and operate them from the Google Cloud Console and the command line. For the mechanics common to every GKE application — Workload Identity, ingress, autoscaling, CI/CD, Cloud Armor, IAP, Binary Authorization, VPC Service Controls, backups, and the deployment lifecycle — refer to the App_GKE foundation guide.

---

What Immich costs on RAD, and how that compares

Immich on RAD's GKE module costs about US$150 a month in your own project, plus a 300-credit (US$30) module fee once per deployment — with every feature unlocked, including face recognition and CLIP search (the optional $99 Immich product key is a supporter badge only). If you only need it occasionally, see "Pause it for free" below to stop paying without losing your photo library. Immich ships only as a GKE module on RAD; there is no Cloud Run variant.

What you pay on RAD

	In a Google Cloud project you own	In a project RAD manages for you
Module fee, once per deployment	300 credits (US$30.00)	270 credits (10% lower)
Build time	A few credits per build	The same
Google Cloud running cost	Billed by Google to your own billing account, about US$149.81 a month for the default configuration (table below)	Metered hourly in credits; RAD publishes 116 credits a day
Who owns what	You own the project and its billing; RAD deploys and updates it	RAD owns the project, with guardrails, quotas and budget alerts; you get console access, and a minimum purchased balance is held in reserve

10 credits cost US$1 on a one-off top-up, less on a plan. In a project RAD manages, the database, file server and network are shared by every application in the project, so a second application does not add a second database.

Default running cost in your own project (us-central1, Google list prices):

Resource (module default)	Per month
GKE Autopilot pod, 2x vCPU / 4 GiB	US$79.35
Cloud SQL for PostgreSQL, 1 vCPU / 3.75 GB, zonal, 10 GB SSD	US$51.02
NFS/cache file server (small VM)	US$13.43
Cloud NAT and networking	US$5.00
Cloud Storage (add-ons, backups)	US$1.00
GKE cluster management fee	shared across every GKE app in the project; $0 if this is your only cluster, else ~73
Total	about US$149.81

How it compares

• Immich itself runs no official hosted SaaS — its maintainers are explicit that self-hosting is the point — so the honest comparison is against the services it replaces. Google Photos' own paid storage (Google One) starts at $1.99/month for 100 GB and $2.99/month for 200 GB, but a shared family photo library with video quickly outgrows that, and you give up auto-tagging control and ownership of the originals in exchange.
• Immich's own footprint (2 vCPU, 4 GiB, plus a database and an NFS VM) is close to a Hetzner CPX22 or DigitalOcean 2 vCPU/4 GB droplet (about $24/month) or a GCP Compute Engine e2-standard-2 (about $49/month) — cheaper in cash, but then you run PostgreSQL, the NFS share, backups, security patches and the machine-learning container's upgrades yourself.
• What RAD adds over a bare VPS: managed Cloud SQL with automated backups and point-in-time recovery, Secret Manager for credentials, and monitoring — the things that turn a weekend self-host project into something you can leave running unattended.

Pause it for free: delete a RAD-managed project, restore it when you need it

If Immich runs in a project RAD manages for you and you expect to come back to it — a break between courses, a seasonal lull, or simply not knowing yet whether you'll need it next month — this is the better option, and it costs almost nothing.

Deleting a RAD-managed project unlinks its billing first, then asks Google to delete the project outright: Google's own 30-day recoverable soft delete. Unlike deleting Immich as a single module, this does not tear down the GKE pod, Cloud SQL database, NFS file server and networking one by one — the whole project simply stops, and nothing is charged while it waits, because billing is already unlinked.

Restoring, within 30 days and only by the project's owner, asks Google to undelete the project and reattaches its billing account, then asks you to run Update on each deployment to confirm everything came back. Because nothing was individually destroyed, that Update finds the same resources already there — it is a check, not a rebuild, and an Update never charges the module fee again. This costs only a handful of credits (under US$1) in build time for a typical chain of deployments.

What this needs: you must own the project (not one RAD only manages billing for), you must restore it yourself within 30 days — after that Google deletes it for good — and restoring is admitted like creating a new project, so your purchased credit balance must still clear the tier's floor (100 credits for the sandbox tier most use fits). Google says most services are fully working again within 36 hours of a restore.

One real gap: nightly backups are written to a bucket inside the project, and that bucket has Cloud Storage's soft-delete explicitly turned off, so it is very likely gone as soon as you delete the project, even though the project itself is recoverable for 30 days. Immich's own nightly export of the database and library metadata is kept in that bucket; the original photos and videos live in the GKE pod's attached storage and the NFS share, which are not individually destroyed by this path, so your library itself survives the 30 days intact — it is specifically the metadata/database backup copy that is not recoverable from a deleted project's bucket.

Pay only while you use it, the other way: delete and redeploy

In your own project, or once the 30-day window above has passed, the fallback is to delete Immich outright and redeploy it later.

• What a redeploy costs: the module fee again, plus the builds — roughly the module fee shown above plus 3–6 credits of build time.
• When it's worth it: deleting saves money only once Immich would otherwise sit unused for about 6 days or more in your own project (about 2 days or more in a RAD-managed one). Immich's database and NFS file server are the bulk of the running cost here, so if nothing else in the project shares them, deleting Immich removes the whole bill above.
• Keep data first. Nightly backups are written to a bucket inside the deployment, and that bucket is deleted with it — copy the latest backup out (to Google Drive, or a bucket you keep) before deleting if you want to keep it.

Lab sessions and Managed Environments

• Lab sessions, for training. A trainer runs a session for a class. Each participant gets the app in their own Google Cloud project for 15 minutes to 24 hours, within an allowance the trainer sets. Either the trainer funds every place, or each participant pays for their own. Everything is deleted when the session ends and unused credits go back to the trainer.
• Managed Environments, for consultancies. A partner runs the app for a client from a ring-fenced wallet it funds, and settles with the client directly. If the wallet runs low, billing pauses and the data is kept, so nobody receives an unexpected charge. At the end the partner hands the project over and the deployments become the client's own.
• Credits can be bought in more than 20 currencies, including XAF, XOF, NGN, GHS, KES and ZAR, by card, bank transfer or mobile money.

Sources (8 October 2026): Google Cloud list prices from the Cloud Billing Catalog API; RAD fees and the daily-credit estimate from radmodules.dev/pricing; Google Cloud: delete and restore projects for the pause/restore mechanism. Prices change; check each source before relying on a figure.

---

1. Overview

Immich runs as two services: the Immich server (API + in-process background workers, custom-built image) and a separate prebuilt machine-learning container.

Capability	Google Cloud service	Notes
Compute (server)	GKE Autopilot	Custom image over ghcr.io/immich-app/immich-server, port 2283, 2 vCPU / 4 GiB, exactly one replica
Compute (ML)	GKE Autopilot	Prebuilt ghcr.io/immich-app/immich-machine-learning, port 3003, internal-only, CPU inference (no GPU), 2 vCPU / 4 GiB
Database	Cloud SQL for PostgreSQL 15	Required — with the pgvector extension (DB_VECTOR_EXTENSION=pgvector; Cloud SQL has no VectorChord)
Media library	NFS (shared platform volume)	Mounted at /usr/src/app/upload (IMMICH_MEDIA_LOCATION) — Immich has no object-storage backend
Cache & queue	Redis	Required — Immich's job queue and pub/sub; the NFS-server co-hosted Redis is injected by default
Image build	Cloud Build + Artifact Registry	Thin custom build adding the cloud entrypoint
Ingress	Cloud Load Balancing	External LoadBalancer, optional custom domain + managed certificate

Sensible defaults worth knowing up front:

• enable_nfs = true is validated, not just defaulted. The entire photo/video library lives at IMMICH_MEDIA_LOCATION; without NFS that path is the pod's ephemeral disk and every restart wipes the library. The plan fails if you turn it off.
• enable_redis = true is validated. Immich refuses to start without Redis. With no explicit redis_host, the platform injects the NFS-server co-hosted Redis IP.
• max_instance_count = 1 is validated. One NFS media library, one writer; Immich's in-process job workers assume a single instance.
• The Deployment uses the Recreate strategy, not RollingUpdate — App_GKE switches automatically for NFS-backed apps, so version updates incur a short outage instead of deadlocking two pods on the same library.
• application_version = "latest" resolves to Immich's rolling release tag. Immich publishes release and vX.Y.Z tags but no latest; the shared layer maps latest → release via the app-specific IMMICH_VERSION build ARG, and the machine-learning image uses the same resolved tag (lock-step).
• pgvector is the vector backend. Cloud SQL has no VectorChord (Immich's preferred extension); Immich runs on its documented pgvector fallback. The db-init job pre-creates vector and earthdistance extensions.
• Probes hit GET /api/server/ping — Immich's unauthenticated liveness endpoint ({"res":"pong"}).
• No app-level secrets. Immich's JWT signing keys live in the database; DB_PASSWORD is injected by the foundation under the exact name Immich reads.
• First run is interactive. Visit the web UI and create the admin account on the sign-up screen; the Immich mobile apps (iOS/Android) then connect to the same server URL.

---

2. Google Cloud Services & How to Explore Them

All commands assume you have run gcloud container clusters get-credentials <cluster> --region <region> --project <project> and that PROJECT, REGION, and NAMESPACE are set. The namespace and other identifiers are reported in the deployment Outputs.

A. GKE Autopilot — the Immich server and ML workloads

The server Deployment runs a single replica; the machine-learning service runs as a second Deployment in the same namespace, reachable only inside the cluster.

• Console: Kubernetes Engine → Workloads → filter by the namespace to see both Deployments, pods, and events. Kubernetes Engine → Services & Ingress shows the external IP.
• CLI:

  kubectl get pods -A | grep immich                       # find the namespace fast
  kubectl get pods,svc,deploy -n "$NAMESPACE"
  kubectl logs -n "$NAMESPACE" deploy/"$(kubectl get deploy -n "$NAMESPACE" -o name | grep -v ml | head -1 | cut -d/ -f2)" --tail=100
  # The cloud entrypoint logs the resolved DB/Redis/media config on the first lines:
  kubectl logs -n "$NAMESPACE" <server-pod> | head -15
  # Confirm the ML URL wiring:
  kubectl exec -n "$NAMESPACE" deploy/<server-deploy> -- env | grep IMMICH_MACHINE_LEARNING_URL

B. Cloud SQL for PostgreSQL 15 + pgvector

Immich stores all metadata (users, albums, assets, EXIF, smart-search embeddings, face-recognition data) in Cloud SQL PostgreSQL 15, reached through the Cloud SQL Auth Proxy sidecar. The first-deploy db-init job creates the database and user and pre-creates the pgvector and earthdistance extensions; Immich runs with DB_VECTOR_EXTENSION = pgvector because Cloud SQL does not offer VectorChord.

• Console: SQL → select the instance.
• CLI:

  gcloud sql instances list --project "$PROJECT"
  gcloud sql connect <instance-name> --user=<db-user> --database=<db-name> --project "$PROJECT"
  # Inside psql — confirm the vector extension:
  #   SELECT extname, extversion FROM pg_extension WHERE extname IN ('vector','earthdistance');

The instance name, database name, user, and password secret are in the Outputs.

C. NFS — the media library

The shared platform NFS volume is mounted at /usr/src/app/upload (IMMICH_MEDIA_LOCATION). Every original, thumbnail, and encoded video lives there; the pod itself is disposable.

• CLI:

  # The NFS server is a Compute Engine VM managed by Services_GCP:
  gcloud compute instances list --project "$PROJECT" --filter="name~nfs"
  # Verify the mount and library content from inside the pod:
  kubectl exec -n "$NAMESPACE" deploy/<server-deploy> -- df -h /usr/src/app/upload
  kubectl exec -n "$NAMESPACE" deploy/<server-deploy> -- ls /usr/src/app/upload

D. Redis (job queue and pub/sub)

Immich queues every background job (thumbnail generation, metadata extraction, smart-search embedding, face detection) through Redis. By default the NFS-server VM co-hosts Redis and the platform injects its IP as REDIS_HOST; the cloud entrypoint maps it to Immich's REDIS_HOSTNAME and fails fast if it resolves empty.

• CLI:

  kubectl exec -n "$NAMESPACE" deploy/<server-deploy> -- env | grep -E 'REDIS_HOST'

E. Machine learning — CLIP smart search and face recognition

The prebuilt ghcr.io/immich-app/immich-machine-learning container serves CLIP embeddings, face recognition, and OCR over HTTP on port 3003 (internal-only). It is CPU inference — no GPU. Its env sets IMMICH_PORT = "3003" explicitly: the foundation propagates the server's env into additional services and the ML image reads the same IMMICH_PORT variable — without the override it inherits 2283, binds the wrong port, and its :3003 startup probe never passes. The server finds it via the injected IMMICH_MACHINE_LEARNING_URL, set via module_env_vars to the real Kubernetes Service DNS name (http://<service>-ml:3003 — the Service is named <service>-ml); the foundation's output_env_var_name mechanism composes an unresolvable bare-name URL (http://ml:3003), so its output is parked in the unused IMMICH_ML_URL_FOUNDATION_UNUSED env var. Model files download lazily on first use into /cache on the pod's ephemeral disk (re-downloaded after rescheduling — accepted).

• CLI:

  kubectl get pods -n "$NAMESPACE" | grep ml
  kubectl logs -n "$NAMESPACE" deploy/<ml-deploy> --tail=50    # model downloads + inference requests

F. Cloud Build & Artifact Registry

The server image is a thin custom build (FROM ghcr.io/immich-app/immich-server plus the cloud entrypoint) produced by Cloud Build and stored in Artifact Registry. The IMMICH_VERSION build ARG carries the resolved tag — app-specific on purpose, because the foundation injects APP_VERSION and wins the merge, and Immich has no latest tag.

• CLI:

  gcloud builds list --project "$PROJECT" --limit 5
  gcloud artifacts docker images list <region>-docker.pkg.dev/"$PROJECT"/<repo> 2>/dev/null | grep immich

G. Networking, Logging & Monitoring

External access is via a Cloud Load Balancing IP (service_type = LoadBalancer); a custom domain with a managed certificate and a reserved static IP are on by default (enable_custom_domain = true, reserve_static_ip = true).

• CLI:

  kubectl get svc -n "$NAMESPACE"
  gcloud compute addresses list --project "$PROJECT"
  gcloud logging read 'resource.type="k8s_container" AND resource.labels.namespace_name="'"$NAMESPACE"'"' \
    --project "$PROJECT" --limit 50

---

3. Immich Application Behaviour

• The cloud entrypoint maps env names at runtime. The custom image's cloud-entrypoint.sh maps the foundation-injected DB_HOST/DB_IP → DB_HOSTNAME (rewriting an Auth Proxy socket-directory path to 127.0.0.1 TCP), DB_USER → DB_USERNAME, DB_NAME → DB_DATABASE_NAME, and REDIS_HOST → REDIS_HOSTNAME/REDIS_PORT — the foundation may inject REDIS_HOST as host or host:port, so the entrypoint splits it into a bare hostname plus a port (default 6379 when none is present). Kubernetes $(VAR) references could not do this: K8s resolves $(VAR) only against env entries defined earlier in the alphabetically rendered list, and DB_DATABASE_NAME sorts before DB_NAME, so $(DB_NAME) would stay a literal string. DB_PASSWORD needs no mapping. The entrypoint also resolves the upstream start script across image layouts — Immich v3 moved the app to /usr/src/app/server (start script at server/bin/start.sh) while older images keep /usr/src/app/start.sh; both are probed and the working directory adjusted before exec. The entrypoint prints the resolved configuration at the top of the pod log — always the first place to look.
• First-deploy database setup. The db-init job (postgres:15-alpine) runs db-init.sh idempotently: user, database, grants, pgvector + earthdistance extensions, and a cloudsqlsuperuser grant so upstream migrations can manage extensions themselves. Immich applies its own schema migrations on every startup.
• Single writer, Recreate deploys. Because the media library is one NFS filesystem and job workers run in-process, only one replica is allowed (plan-time validated), and version updates replace the pod with Recreate (brief downtime) rather than a rolling surge.
• Smart search and face recognition are asynchronous. After an upload, the server queues jobs that call the ML service; the first smart-search query also triggers the CLIP model download, so it is noticeably slow once per ML pod lifetime. Job progress is visible in the web UI under Administration → Jobs.
• First-run admin. On first visit the web UI shows the sign-up screen; the first registered account becomes the admin. Immich's iOS/Android apps connect to the same server URL for mobile auto-backup.
• Telemetry is off (IMMICH_TELEMETRY_INCLUDE = "") and IMMICH_ENV = production; the API and background workers run in one container (upstream merged the microservices container in v1.106).

---

4. Configuration Variables

Variables are grouped exactly as they appear on the deployment platform. Only settings specific to or notable for Immich are listed; every other input is inherited from App_GKE with its standard behaviour and defaults.

Group 1 — Project & Identity

Variable	Default	Description
project_id	(required)	Target Google Cloud project.
region	us-central1	Region fallback when network discovery cannot determine it.

Group 2 — Deployment Environment

Variable	Default	Description
tenant_id	demo	Short suffix that makes resource names unique per environment.
support_users	[]	Emails granted project access and monitoring alerts.
resource_labels	{}	Labels applied to all resources.

Group 3 — Application Identity

Variable	Default	Description
application_name	immich	Base name for resources. Do not change after first deploy.
application_version	latest	Server + ML image tag, kept in lock-step. latest resolves to Immich's rolling release tag (no latest tag exists upstream); pin vX.Y.Z in production.

Group 4 — Runtime & Scaling

Variable	Default	Description
deploy_application	true	Set false to provision infrastructure only.
cpu_limit / memory_limit	2000m / 4Gi	Server container resources.
container_port	2283	Immich's native port (IMMICH_PORT).
min_instance_count	1	Keep at 1 so Immich is always available.
max_instance_count	1	Must be 1 — plan-time validated (single NFS writer, in-process workers).
ml_cpu_limit / ml_memory_limit	2000m / 4Gi	Machine-learning container resources. CLIP + face models need ~2–3Gi resident; 4Gi is the safe default.
enable_cloudsql_volume	true	Cloud SQL Auth Proxy sidecar.
enable_image_mirroring	true	Mirror images into Artifact Registry.

Group 5 — Environment Variables & Secrets

Variable	Default	Description
environment_variables	{}	Extra settings merged over the Immich defaults (IMMICH_PORT, IMMICH_MEDIA_LOCATION, DB_VECTOR_EXTENSION, IMMICH_ENV, telemetry). Do not set DB_*/REDIS_* here — the entrypoint owns the mapping.
secret_environment_variables	{}	Map of env var → Secret Manager secret name.

Group 6 — GKE Backend & Cluster

Variable	Default	Description
service_type	LoadBalancer	External IP for the web UI and mobile apps.
session_affinity	ClientIP	Sticky routing (single replica, so mostly moot).
network_tags	["nfsserver"]	Required for NFS connectivity.
termination_grace_period_seconds	60	Grace before force-kill.
deployment_timeout	1800	Seconds Terraform waits for the rollout.

Group 10 — Observability & Health

Variable	Default	Description
startup_probe	HTTP /api/server/ping, 30s delay, 30×10s	Application startup probe (Group 14 on the platform).
liveness_probe	HTTP /api/server/ping, 30s delay	Application liveness probe.
uptime_check_config	disabled, path /api/server/ping	Optional Cloud Monitoring uptime check.

Group 11 — Jobs & Additional Services

Variable	Default	Description
initialization_jobs	[]	Leave empty for the built-in db-init job. Immich migrates its own schema on startup.
additional_services	[]	Appended after the built-in machine-learning service.

Group 13 — Filesystem (NFS)

Variable	Default	Description
enable_nfs	true	Must stay true — plan-time validated. The entire media library lives on NFS; Immich has no S3/GCS backend.
nfs_mount_path	/usr/src/app/upload	Mounted exactly at IMMICH_MEDIA_LOCATION.

Group 16 — Database Backend

Variable	Default	Description
db_name	immich_db	PostgreSQL database name. Immutable after first deploy.
db_user	immich_user	Application database user. Immutable after first deploy.
database_password_length	32	Generated password length (16–64).

Group 17 — Backup & Maintenance

Variable	Default	Description
backup_schedule	0 2 * * *	Automated database and NFS backup cron (UTC).
backup_retention_days	7	Raise for production — the NFS library is the only copy of your photos.

Group 19 — Custom Domain, Static IP & Networking

Variable	Default	Description
enable_custom_domain	true	Gateway API + managed certificate.
application_domains	[]	Hostnames to serve.
reserve_static_ip	true	Stable external IP across redeploys.

Group 20 — Identity-Aware Proxy (IAP)

Warning: the Immich mobile apps authenticate against the Immich server, not Google — IAP in front of the API breaks mobile auto-backup unless every device can complete the Google sign-in flow.

Variable	Default	Description
enable_iap	false	Requires custom domain and both OAuth credentials (validated).

Group 21 — Redis & Cloud Armor

Variable	Default	Description
enable_redis	true	Must stay true — plan-time validated; Immich refuses to start without Redis.
redis_host	""	Empty = the NFS-server co-hosted Redis IP is injected.
enable_cloud_armor	false	Attach a WAF policy to the Ingress backend.

Group 22 — VPC Service Controls & Audit Logging

Standard App_GKE inputs: enable_vpc_sc, vpc_cidr_ranges, vpc_sc_dry_run, organization_id, enable_audit_logging. See App_GKE.

---

5. Outputs

Output	Description
service_name	Kubernetes Service name.
namespace	Namespace the workloads run in.
service_cluster_ip / stage_service_cluster_ips	In-cluster ClusterIPs.
service_external_ip	External LoadBalancer IP (when a static IP is reserved).
web_url	URL for the Immich web UI — external LoadBalancer IP if available, otherwise internal cluster URL.
database_instance_name / database_name / database_user	Cloud SQL identifiers.
database_password_secret	Secret Manager secret holding the DB password.
database_host / database_port	DB endpoint (127.0.0.1 via the Auth Proxy) / port.
storage_buckets	Created Cloud Storage buckets.
network_name / network_exists / regions	VPC network details.
container_image / container_registry	Deployed image and Artifact Registry repo.
monitoring_enabled / monitoring_notification_channels	Monitoring status and channels.
initialization_jobs / db_import_job	Setup and (optional) import job names.
deployment_id / tenant_id / resource_prefix	Naming identifiers.
project_id / project_number	Project identifiers.
cicd_enabled / cicd_configuration and GitHub/trigger outputs	CI/CD status and details.
kubernetes_ready	Whether all Kubernetes resources were deployed (false on the first apply of a new inline cluster).
vpc_sc_enabled / vpc_sc_perimeter_name / vpc_sc_dry_run_mode	VPC-SC status.
audit_logging_enabled / artifact_registry_cmek_enabled	Audit logging and CMEK status.

---

6. Configuration Pitfalls & Sensible Defaults

Risk: Critical (data loss / outage / security) — High (service degraded) — Medium (cost or partial degradation) — Low (minor).

Plan-time validation. Immich_GKE carries its own validation guards on top of the App_GKE foundation checks: enable_nfs = false, enable_redis = false, max_instance_count > 1, Redis with no host source, and IAP without OAuth credentials all fail the plan with a named error before any resource is created.

Setting	Sensible value	Risk	Consequence if wrong
enable_nfs	true (validated)	Critical	With NFS off, the entire media library sits on the pod's ephemeral disk — every pod rescheduling wipes all photos and videos. Immich has no S3/GCS backend to fall back to. Blocked at plan time.
max_instance_count	1 (validated)	Critical	More than one replica means multiple writers on one NFS library and duplicated in-process job workers — library corruption and racing jobs. Blocked at plan time.
db_name / db_user	Set once	Critical	Immutable after first deploy; renaming recreates the DB/user and orphans all asset metadata, albums, and users.
backup_retention_days	Raise for production	High	The NFS library is the only copy of the media; 7 days of backups is thin for a photo archive.
enable_redis / redis_host	true / "" (validated)	High	Without Redis the server exits at startup (the entrypoint fails fast with a clear error). With Redis on but no NFS and no explicit host, REDIS_HOST is empty — also blocked at plan time.
ML IMMICH_PORT override	Keep the built-in IMMICH_PORT = "3003" on the ML service	High	The foundation propagates the server's env into additional services, and the ML image reads the same IMMICH_PORT variable as the server — without the override the ML container inherits 2283, binds the wrong port, and its :3003 startup probe never passes (the ML Deployment never becomes Ready).
IMMICH_MACHINE_LEARNING_URL	Leave the injected real-DNS URL alone	High	The foundation's output_env_var_name mechanism composes the URL from the bare additional-service name (http://ml:3003), but the Service it creates is named <service>-ml — the bare name does not resolve and smart search/face recognition fail. The module injects the real Service DNS URL via module_env_vars and parks the foundation-composed value in the unused IMMICH_ML_URL_FOUNDATION_UNUSED.
ml_memory_limit	4Gi (default; treat as the floor)	Medium	Below the ~2–3Gi the CLIP + face models need resident, model load OOM-kills the ML pod — smart search and face recognition silently fail while the main app looks perfectly healthy (uploads and browsing still work). Watch for OOMKilled restarts on the ML pod.
enable_iap	false unless mobile access is managed	Medium	IAP intercepts the API the mobile apps call; auto-backup breaks for devices that cannot complete Google sign-in.
application_version	latest (→ release) or a pinned vX.Y.Z	Medium	Setting a tag that doesn't exist upstream fails the build/pull; the server and ML tags are kept in lock-step automatically — do not point them at different versions manually.
Version updates	Expect brief downtime	Low	NFS-backed apps deploy with the Recreate strategy — the old pod stops before the new one starts. This is intentional (rolling updates deadlock on the shared library).
pgvector vs VectorChord	Accept pgvector on Cloud SQL	Low	Smart-search index builds and queries are slower than Immich's preferred VectorChord — expected on Cloud SQL, which does not offer VectorChord. Functionally complete, just slower on large libraries.
First ML query latency	Expect a slow first search	Low	CLIP/face models download on first use (cache on ephemeral disk, re-downloaded after ML pod rescheduling) — the first smart-search query after a deploy is slow.
Startup probe path	/api/server/ping	Low	Any authenticated endpoint returns 401/403 to the unauthenticated kubelet probe and wedges the rollout; keep the default.

---

For the foundation behaviour referenced throughout — IAM and Workload Identity, autoscaling, ingress and certificates, CI/CD, Cloud Armor, IAP, Binary Authorization, VPC-SC, backups, and image mirroring — see App_GKE. The Immich-specific shared application layer (image, entrypoint, database bootstrap, probes) is described in Immich_Common.

Related guides

• Hands-on lab: Immich on GKE Autopilot — deploy it step by step, with the console screens and commands at each stage.
• Immich Common — Shared Application Configuration — the configuration shared by both deployment targets.
• Deployed alongside Nextcloud on Google Cloud Run, Filebrowser on GKE Autopilot, Vaultwarden on Google Cloud Run, Kopia on GKE Autopilot in the Personal Cloud solution.