Add OpenStack Terraform support and TempCNN ml_tune_grid R example.

.github/workflows/terraform-validate.yml

@@ -23,7 +23,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        cloud: [aws, azure, gcp]
+        cloud: [aws, azure, gcp, openstack]
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6

.gitignore

@@ -16,6 +16,7 @@ crash.log
 terraform/aws/terraform.tfvars
 terraform/azure/terraform.tfvars
 terraform/gcp/terraform.tfvars
+terraform/openstack/terraform.tfvars
 
 # inst/demo-lps-2025/00_ml_month.ipynb — openEO job downloads
 inst/demo-lps-2025/data/output-month/

DESCRIPTION

@@ -22,7 +22,6 @@ URL: https://github.com/Open-Earth-Monitor/openeocraft
 BugReports: https://github.com/Open-Earth-Monitor/openeocraft/issues
 Encoding: UTF-8
 Roxygen: list(markdown = TRUE)
-RoxygenNote: 7.3.3
 Depends:
     R (>= 4.5.0)
 Imports:
@@ -58,3 +57,4 @@ Collate:
     'pgraphs.R'
     'mock.R'
     'run.R'
+Config/roxygen2/version: 8.0.0

README.md

@@ -208,10 +208,10 @@ docker run -d --name openeocraft --restart unless-stopped \
 
 #### Terraform (recommended)
 
-For repeatable cloud provisioning, use the Terraform stacks in [`terraform/README.md`](terraform/README.md). Defaults target **GPU with at least 32 GB RAM** on AWS (`g4dn.2xlarge`), Azure (`Standard_NC8as_T4_v3`), and GCP (`g2-standard-8`), with NVIDIA drivers and `docker run --gpus all`. See each cloud README for quota, auth, and CPU fallbacks.
+For repeatable cloud provisioning, use the Terraform stacks in [`terraform/README.md`](terraform/README.md). Defaults target **GPU with at least 32 GB RAM** on AWS, Azure, and GCP; OpenStack uses a configurable `flavor_name` with a **32 GB RAM minimum**. See each cloud README for quota, auth, and CPU fallbacks.
 
 ```bash
-cd terraform/aws    # or azure/ or gcp/
+cd terraform/aws    # or azure/, gcp/, or openstack/
 cp terraform.tfvars.example terraform.tfvars
 # Set credentials, key pair / SSH key, ssh_cidr_blocks, api_cidr_blocks
 terraform init

inst/examples/06_ml_tempcnn_tune_grid.R

@@ -0,0 +1,102 @@
+# TempCNN fine-tuning (grid search) + inference via the openEO R client.
+#
+# Prerequisites:
+#   - OpenEOcraft backend running locally (e.g. source docker/plumber.R in R/RStudio,
+#     or docker compose up)
+#   - openeo R package installed
+#
+# Run from repo root:
+#   Rscript inst/examples/06_ml_tempcnn_tune_grid.R
+#
+# Default connection below matches docker/plumber.R demo credentials (user brian).
+# With a local plumber process, job artefacts are written under the backend
+# workdir, e.g. <workdir>/workspace/<user>/jobs/<job_id>/ (tuning_results.json,
+# saved model, GeoTIFF). Set the workdir when starting the API if needed.
+#
+# Cube bands must match the training RDS timeline and features (P16D; add NDVI
+# when samples were built with an NDVI band). See inst/demo-lps-2025/00_ml_month.ipynb.
+# Grid search uses a small param_grid for demo wall time; see inst/demo-lps-2025/.
+
+library(openeo)
+
+con <- connect(
+    host = "http://127.0.0.1:8000",
+    user = "brian",
+    password = "123456"
+)
+
+p <- processes()
+
+deforestation_data <-
+    "https://github.com/e-sensing/sitsdata/raw/main/data/samples_deforestation_rondonia.rds"
+
+tempcnn_model_init <- p$mlm_class_tempcnn(
+    optimizer = "adam",
+    learning_rate = 0.0005,
+    seed = 42,
+    epochs = 20,
+    batch_size = 64,
+    verbose = TRUE
+)
+
+param_grid <- list(
+    learning_rate = c(0.0005, 0.0001),
+    epochs = c(20, 40)
+)
+
+tempcnn_tuned <- p$ml_tune_grid(
+    model = tempcnn_model_init,
+    training_data = deforestation_data,
+    target = "label",
+    parameters = param_grid,
+    scoring = "accuracy",
+    cv = 0,
+    seed = 42
+)
+
+tempcnn_model <- p$save_ml_model(
+    data = tempcnn_tuned,
+    name = "tempcnn_rondonia_tuned_v1",
+    return_model = TRUE
+)
+
+datacube <- p$load_collection(
+    id = "mpc-sentinel-2-l2a",
+    spatial_extent = list(
+        west = -63.50,
+        south = -8.92,
+        east = -63.35,
+        north = -8.78
+    ),
+    temporal_extent = c("2022-01-01", "2022-12-31"),
+    bands = list(
+        "b02", "b03", "b04", "b05", "b06", "b07", "b08",
+        "b11", "b12", "b8a"
+    )
+)
+
+datacube <- p$cube_regularize(
+    data = datacube,
+    period = "P16D",
+    resolution = 30
+)
+
+datacube <- p$ndvi(
+    data = datacube,
+    red = "B04",
+    nir = "B08",
+    target_band = "NDVI"
+)
+
+data <- p$ml_predict(data = datacube, model = tempcnn_model)
+
+ml_job <- p$save_result(data = data, format = "GTiff")
+
+job <- create_job(
+    graph = ml_job,
+    title = "TempCNN fine-tuning + inference",
+    description = "Grid-search TempCNN on Rondonia samples; predict GeoTIFF"
+)
+job <- start_job(job)
+
+describe_job(job)

terraform/README.md

@@ -1,6 +1,6 @@
 # OpenEOcraft Terraform (multi-cloud)
 
-Repeatable provisioning for OpenEOcraft on **AWS**, **Azure**, and **Google Cloud**. Each stack launches a single **x86_64** Ubuntu 22.04 VM, installs Docker via cloud-init, optionally installs the **NVIDIA driver + NVIDIA Container Toolkit**, pulls the OpenEOcraft image, and starts the API on port **8000**.
+Repeatable provisioning for OpenEOcraft on **AWS**, **Azure**, **Google Cloud**, and **OpenStack**. Each stack launches a single **x86_64** Ubuntu 22.04 VM, installs Docker via cloud-init, optionally installs the **NVIDIA driver + NVIDIA Container Toolkit**, pulls the OpenEOcraft image, and starts the API on port **8000**.
 
 All defaults target **at least 32 GB instance RAM** (30 GB container limit, leaving headroom for the OS).
 
@@ -9,6 +9,7 @@ All defaults target **at least 32 GB instance RAM** (30 GB container limit, leav
 | AWS | [`aws/`](aws/) | `g4dn.2xlarge` (32 GB, 1× T4) | `m5.2xlarge` (32 GB) |
 | Azure | [`azure/`](azure/) | `Standard_NC8as_T4_v3` (56 GB, 1× T4) | `Standard_D8s_v5` (32 GB) |
 | GCP | [`gcp/`](gcp/) | `g2-standard-8` (32 GB, 1× L4) | `n2-standard-8` (32 GB) |
+| OpenStack | [`openstack/`](openstack/) | cloud-specific GPU flavor | `flavor_name` ≥ 32 GB RAM (default `m1.8xlarge`) |
 
 Bootstrap scripts live in [`shared/cloud_init.tftpl`](shared/cloud_init.tftpl) and are shared across clouds. Workspace disks are mounted only via stable cloud-specific device paths (never heuristic disk guessing); if the data disk is not ready, bootstrap falls back to a directory on the root filesystem.
 
@@ -17,7 +18,7 @@ Bootstrap scripts live in [`shared/cloud_init.tftpl`](shared/cloud_init.tftpl) a
 Pick a cloud, copy the example variables, edit credentials and CIDR blocks, then apply:
 
 ```bash
-cd terraform/aws    # or azure/ or gcp/
+cd terraform/aws    # or azure/, gcp/, or openstack/
 cp terraform.tfvars.example terraform.tfvars
 # Edit terraform.tfvars (key pair / SSH key, your public IP /32)
 terraform init
@@ -34,14 +35,15 @@ Allow **15–20 minutes** on first boot when `enable_gpu = true`.
 | ----------- | ----- |
 | [Terraform](https://developer.hashicorp.com/terraform/install) >= 1.5 | |
 | Cloud account with billing enabled | GPU instances require quota in most accounts |
-| SSH access | AWS: EC2 key pair name; Azure/GCP: SSH public key |
+| SSH access | AWS/OpenStack: key pair name; Azure/GCP: SSH public key |
 | Your public IP as CIDR | `curl -s ifconfig.me` → `x.x.x.x/32` for SSH and API rules |
 
 ## Cloud-specific docs
 
 - [AWS EC2](aws/README.md)
 - [Azure VM](azure/README.md)
 - [Google Compute Engine](gcp/README.md)
+- [OpenStack Nova](openstack/README.md)
 
 ## What each stack creates
 

terraform/openstack/README.md

@@ -0,0 +1,104 @@
+# OpenEOcraft on OpenStack (Terraform)
+
+This stack provisions a single **x86_64** Ubuntu 22.04 Nova instance on a dedicated tenant network with a floating IP, installs Docker via cloud-init, optionally installs the **NVIDIA driver + NVIDIA Container Toolkit** (GPU mode), pulls the OpenEOcraft image, and starts the API on port **8000**.
+
+**Default target:** CPU flavor with **≥ 32 GB RAM** (`flavor_name`, default `m1.8xlarge` — verify on your cloud with `openstack flavor list`).
+
+## Prerequisites
+
+| Requirement | Notes |
+| ----------- | ----- |
+| [Terraform](https://developer.hashicorp.com/terraform/install) >= 1.5 | |
+| OpenStack project with compute + networking + block storage | |
+| Nova **key pair** | `openstack keypair create …` |
+| Flavor with **≥ 32 GB RAM** | `openstack flavor list` — names vary by cloud |
+| External network for floating IPs | Often `public` or `ext-net` |
+| Your public IP as CIDR | `curl -s ifconfig.me` → `x.x.x.x/32` |
+
+### OpenStack authentication
+
+Set standard OpenStack environment variables (from `openstack rc` or your cloud dashboard):
+
+```bash
+export OS_AUTH_URL=...
+export OS_USERNAME=...
+export OS_PASSWORD=...
+export OS_PROJECT_NAME=...
+export OS_USER_DOMAIN_NAME=Default
+export OS_PROJECT_DOMAIN_NAME=Default
+export OS_REGION_NAME=...
+openstack token issue
+```
+
+Or use `clouds.yaml` / application credentials as supported by the [OpenStack Terraform provider](https://registry.terraform.io/providers/terraform-provider-openstack/openstack/latest/docs).
+
+## Quick start
+
+```bash
+cd terraform/openstack
+cp terraform.tfvars.example terraform.tfvars
+```
+
+Edit `terraform.tfvars`:
+
+- `key_pair_name` — your Nova key pair
+- `flavor_name` — a flavor with **≥ 32768 MB RAM** on your cloud
+- `external_network_name` / `floating_ip_pool` — your provider external network
+- `ssh_cidr_blocks` / `api_cidr_blocks` — your public IP `/32`
+
+```bash
+terraform init
+terraform plan
+terraform apply
+```
+
+After apply:
+
+```bash
+terraform output api_url
+```
+
+**Bootstrap time:** allow **15–20 minutes** on first boot when `enable_gpu = true`.
+
+## What gets created
+
+| Resource | Purpose |
+| -------- | ------- |
+| Network, subnet, router | Tenant network with external gateway |
+| Security group | SSH (22) and OpenEOcraft API (8000) from your CIDR lists |
+| Nova instance | Ubuntu 22.04, Docker, optional NVIDIA stack, OpenEOcraft container |
+| Cinder volume (optional) | Persistent `/var/openeo/workspace` on `/dev/vdb` (default 200 GB) |
+| Floating IP | Public access to the API |
+
+## Variables
+
+See [`variables.tf`](variables.tf). Key settings:
+
+| Variable | Default | Notes |
+| -------- | ------- | ----- |
+| `flavor_name` | `m1.8xlarge` | Must exist on your cloud with **≥ 32 GB RAM** |
+| `enable_gpu` | `false` | Set `true` only with a GPU flavor; names are cloud-specific |
+| `docker_cpus` / `docker_memory_gb` | `8` / `30` | Container limits |
+| `workspace_volume_size_gb` | `200` | Set to `0` to skip separate Cinder volume |
+| `external_network_name` | `public` | Provider external network for router + floating IP |
+
+### GPU (optional)
+
+GPU flavor names differ widely between OpenStack deployments. Copy [`terraform.gpu.tfvars.example`](terraform.gpu.tfvars.example), set `flavor_name` to a GPU flavor from your cloud, and set `enable_gpu = true`.
+
+## Verify after apply
+
+```bash
+ssh ubuntu@$(terraform output -raw public_ip)
+
+nvidia-smi   # when enable_gpu = true
+sudo docker ps
+sudo docker logs openeocraft --tail 50
+curl -s -o /dev/null -w "%{http_code}\n" http://localhost:8000/
+```
+
+## Teardown
+
+```bash
+terraform destroy
+```