feat(deploy): use native XFS for Aliyun nodes

AGENTS.md

@@ -394,6 +394,15 @@ quotas for runsc and Firecracker use this local-disk filestore. Without an
 explicit quota, runsc retains its configured memory-backed overlay while
 Firecracker creates its configured sparse ext4 default.
 
+Terraform-managed Alibaba Cloud node pools instead attach a dedicated 300 GiB
+ESSD by default, have ACK format it as XFS, and mount it at `/home/akernel`.
+The Aliyun sandboxd configuration leaves `filestore_dir_size` unset and uses
+`/home/akernel/filestore` directly, so writable layers and
+`/home/akernel/checkpoints` share the native reflink-capable filesystem. Do
+not set a bounded filestore size for this profile because that reintroduces a
+loop-backed filesystem and disables the high-performance Firecracker C/R
+layout.
+
 The bundled node enables YuanRong's local-only sandbox snapshot data plane and
 stores checkpoint state under the persistent `/home/akernel/checkpoints`
 mount. RRT receives

deploy/README.md

@@ -313,6 +313,12 @@ through its own LoadBalancer when `install_monitor=true`. Set
 `install_dragonfly=true` to install the pinned official Dragonfly chart and
 inject its seed-client proxy into the node runtime configuration.
 
+Terraform-managed Alibaba Cloud nodes also receive a dedicated 300 GiB XFS
+disk mounted at `/home/akernel` by default. sandboxd consumes that native
+filesystem directly for writable layers and local checkpoints, without a
+loop-backed filestore. See the Aliyun guide for capacity, opt-out, and node
+replacement details.
+
 Only the AKernel all-in-one image is pushed to the registry selected by
 `make config`. etcd, Traefik, Grafana, Prometheus, Loki, Tempo, and BusyBox use
 their pinned official public images by default. Set the per-component image

deploy/scripts/configure.sh

@@ -312,6 +312,11 @@ node_pool_system_disk_size     = 100
 node_pool_data_disk_enabled    = true
 node_pool_data_disk_category   = "cloud_essd"
 node_pool_data_disk_size       = 100
+node_pool_extra_data_disk_enabled    = true
+node_pool_extra_data_disk_category   = "cloud_essd"
+node_pool_extra_data_disk_size       = 300
+node_pool_extra_data_disk_mount_path = "/home/akernel"
+node_pool_extra_data_disk_fs_type    = "xfs"
 node_pool_key_name             = "${node_pool_key_name}"
 
 core_namespace = "akernel"

deploy/terraform/aliyun/README.md

@@ -112,6 +112,45 @@ terraform apply \
   -var 'monitor_storage_class=alicloud-disk-essd'
 ```
 
+## AKernel node storage
+
+New Terraform-managed ACK node pools attach two data disks by default. The
+first remains available to ACK for the container runtime. The second is a
+dedicated 300 GiB ESSD that ACK formats as XFS and mounts at
+`/home/akernel` before the node joins the cluster. The same storage layout is
+applied to user-defined `extra_node_pools`; dedicated Dragonfly pools retain
+their own storage configuration.
+
+AKernel stores both `/home/akernel/filestore` and
+`/home/akernel/checkpoints` on that native XFS filesystem. The generated
+sandboxd configuration intentionally leaves `filestore_dir_size` unset, so
+sandboxd uses the directory directly instead of creating a loop-backed ext4
+or XFS filesystem. Keeping the writable layer and checkpoint artifacts on
+the same reflink-capable filesystem enables the fast Firecracker checkpoint
+and restore path.
+
+Change the dedicated disk capacity or category with
+`node_pool_extra_data_disk_size` and
+`node_pool_extra_data_disk_category`. Set
+`node_pool_extra_data_disk_enabled=false` to opt out, for example when an
+existing cluster already provisions `/home/akernel` itself. The option has no
+effect when `create_cluster=false`; existing-cluster users must mount a
+suitable host filesystem before deploying the chart. An explicit ext4
+override remains supported for compatibility, but it cannot provide the XFS
+reflink checkpoint path.
+
+Changing these settings does not migrate live sandbox data on existing
+nodes. Before replacing an existing node pool, drain its sandboxes and remove
+or relocate lifecycle-bound checkpoints. Review the Terraform plan as the
+default dedicated disk adds one cloud disk per AKernel node.
+
+After provisioning a node, verify the effective layout with:
+
+```bash
+findmnt -no SOURCE,FSTYPE,TARGET /home/akernel
+xfs_info /home/akernel | grep 'reflink=1'
+```
+
 Install Dragonfly P2P distribution:
 
 ```bash

deploy/terraform/aliyun/main.tf

@@ -32,6 +32,10 @@ resource "null_resource" "input_validation" {
       condition     = !(length(var.node_pool_key_name) > 0 && length(var.node_pool_login_password) > 0)
       error_message = "node_pool_key_name conflicts with node_pool_login_password. Set only one."
     }
+    precondition {
+      condition     = !var.create_cluster || !(var.node_pool_extra_data_disk_enabled && var.node_home_use_csi_ephemeral)
+      error_message = "Dedicated node data disk and CSI ephemeral home storage are mutually exclusive. Disable one storage mode."
+    }
     precondition {
       condition     = length(var.master_image_tag) > 0 && length(var.node_image_tag) > 0
       error_message = "master_image_tag and node_image_tag must be set to a published AKernel image tag."
@@ -87,78 +91,7 @@ locals {
     sandboxd_nat_backend = var.sandboxd_nat_backend
     max_user_namespaces  = var.node_pool_max_user_namespaces
   })
-  node_pool_bootstrap_user_data = var.node_pool_extra_data_disk_enabled ? join("\n", [
-    local.node_pool_bootstrap_base,
-    <<-EOT
-      # Mount extra data disk
-      set -euo pipefail
-      MOUNT_PATH="${var.node_pool_extra_data_disk_mount_path}"
-      FS_TYPE="${var.node_pool_extra_data_disk_fs_type}"
-      MARKER_PATH="${var.node_pool_extra_data_disk_mount_path}/.storage-ready"
-
-      if mountpoint -q "$MOUNT_PATH" 2>/dev/null; then
-        echo "[data-disk] $MOUNT_PATH is already mounted, skipping."
-        exit 0
-      fi
-
-      ROOT_SOURCE=$(findmnt -n -o SOURCE /)
-      ROOT_DEV=""
-      if [ -n "$ROOT_SOURCE" ]; then
-        ROOT_DEV=$(lsblk -ndo PKNAME "$ROOT_SOURCE" 2>/dev/null || true)
-      fi
-
-      TARGET_DEV=""
-      for dev in /dev/vd[b-z] /dev/xvd[b-z] /dev/sd[b-z] /dev/nvme[0-9]n1 /dev/nvme[1-9][0-9]n1; do
-        [ -b "$dev" ] || continue
-        dev_name=$(basename "$dev")
-        if [ -n "$ROOT_DEV" ] && [ "$dev_name" = "$ROOT_DEV" ]; then
-          continue
-        fi
-        if lsblk -ln -o MOUNTPOINT "$dev" 2>/dev/null | grep -q '/'; then
-          continue
-        fi
-        if pvs "$dev" &>/dev/null 2>&1; then
-          continue
-        fi
-        if blkid "$dev" &>/dev/null; then
-          EXISTING_FS=$(blkid -s TYPE -o value "$dev" 2>/dev/null || true)
-          if [ "$EXISTING_FS" = "$FS_TYPE" ]; then
-            TARGET_DEV="$dev"
-            break
-          fi
-          continue
-        fi
-        TARGET_DEV="$dev"
-        break
-      done
-
-      if [ -z "$TARGET_DEV" ]; then
-        echo "[data-disk] No available data disk found, skipping."
-        exit 0
-      fi
-
-      echo "[data-disk] Using device: $TARGET_DEV"
-      mkdir -p "$MOUNT_PATH"
-
-      if ! blkid "$TARGET_DEV" &>/dev/null; then
-        echo "[data-disk] Formatting $TARGET_DEV with $FS_TYPE..."
-        if [ "$FS_TYPE" = "xfs" ]; then
-          mkfs.xfs -f "$TARGET_DEV"
-        else
-          mkfs.ext4 -F "$TARGET_DEV"
-        fi
-      fi
-
-      mount -o defaults,noatime "$TARGET_DEV" "$MOUNT_PATH"
-      UUID=$(blkid -s UUID -o value "$TARGET_DEV")
-      if ! grep -q "$UUID" /etc/fstab 2>/dev/null; then
-        echo "UUID=$UUID $MOUNT_PATH $FS_TYPE defaults,noatime 0 0" >> /etc/fstab
-      fi
-      echo "[data-disk] Mounted $TARGET_DEV at $MOUNT_PATH ($FS_TYPE) successfully."
-      touch "$MARKER_PATH"
-      echo "[data-disk] Created readiness marker: $MARKER_PATH"
-    EOT
-  ]) : local.node_pool_bootstrap_base
+  node_pool_bootstrap_user_data = local.node_pool_bootstrap_base
   dragonfly_seed_user_data = var.dragonfly_seed_node_pool.mount_local_nvme ? join("\n", [
     local.node_pool_bootstrap_base,
     templatefile("${path.module}/../shared/local-nvme-mount.sh.tftpl", {
@@ -202,34 +135,32 @@ locals {
   traefik_internal_stats_image = length(var.traefik_internal_stats_image) > 0 ? var.traefik_internal_stats_image : "${local.acr_registry}/busybox:1.37.0-musl"
 
   core_values = templatefile("${path.module}/values-akernel.yaml.tmpl", {
-    acr_registry                      = local.acr_registry
-    acr_secret_enabled                = local.acr_secret_enabled
-    acr_host                          = local.acr_host
-    acr_username                      = var.acr_username
-    acr_password                      = var.acr_password
-    etcd_image_repository             = local.etcd_image_repo
-    etcd_image_tag                    = var.etcd_image_tag
-    master_image_repository           = local.master_image_repo
-    master_image_tag                  = var.master_image_tag
-    node_image_repository             = local.node_image_repo
-    node_image_tag                    = var.node_image_tag
-    traefik_image_repository          = local.traefik_image_repo
-    traefik_image_tag                 = var.traefik_image_tag
-    iam_litebus_data_key              = var.iam_litebus_data_key
-    enable_kruise                     = var.install_prereqs
-    master_service_type               = (var.master_public_access_8888 && !var.traefik_enabled) ? var.master_service_type : "ClusterIP"
-    traefik_enabled                   = var.traefik_enabled
-    sandboxd_nat_backend              = var.sandboxd_nat_backend
-    enable_runc                       = var.enable_runc
-    node_secret_create                = var.node_secret_create
-    node_home_use_csi_ephemeral       = var.node_home_use_csi_ephemeral
-    node_home_csi_storage_class       = local.effective_node_home_csi_sc
-    node_home_csi_size                = var.node_home_csi_size
-    node_pool_extra_data_disk_enabled = var.node_pool_extra_data_disk_enabled
-    node_storage_init_image           = var.node_storage_init_image
-    node_storage_init_mount_path      = var.node_pool_extra_data_disk_mount_path
-    oss_auths                         = local.oss_auths
-    registry_auths                    = local.registry_auths
+    acr_registry                = local.acr_registry
+    acr_secret_enabled          = local.acr_secret_enabled
+    acr_host                    = local.acr_host
+    acr_username                = var.acr_username
+    acr_password                = var.acr_password
+    etcd_image_repository       = local.etcd_image_repo
+    etcd_image_tag              = var.etcd_image_tag
+    master_image_repository     = local.master_image_repo
+    master_image_tag            = var.master_image_tag
+    node_image_repository       = local.node_image_repo
+    node_image_tag              = var.node_image_tag
+    traefik_image_repository    = local.traefik_image_repo
+    traefik_image_tag           = var.traefik_image_tag
+    iam_litebus_data_key        = var.iam_litebus_data_key
+    enable_kruise               = var.install_prereqs
+    master_service_type         = (var.master_public_access_8888 && !var.traefik_enabled) ? var.master_service_type : "ClusterIP"
+    traefik_enabled             = var.traefik_enabled
+    sandboxd_nat_backend        = var.sandboxd_nat_backend
+    enable_runc                 = var.enable_runc
+    node_secret_create          = var.node_secret_create
+    node_home_use_csi_ephemeral = var.node_home_use_csi_ephemeral
+    node_home_csi_storage_class = local.effective_node_home_csi_sc
+    node_home_csi_size          = var.node_home_csi_size
+    node_host_disk_path         = var.create_cluster && var.node_pool_extra_data_disk_enabled ? var.node_pool_extra_data_disk_mount_path : "/home/akernel"
+    oss_auths                   = local.oss_auths
+    registry_auths              = local.registry_auths
 
     etcd_storage_class = local.effective_storage_class
     etcd_cpu           = var.etcd_resources.cpu
@@ -313,14 +244,19 @@ locals {
     tempo_pvc_size       = var.tempo_resources.pvc_size
   })
 
+  akernel_extra_node_pools = [for pool in var.extra_node_pools : merge(pool, {
+    use_akernel_data_disk = true
+  })]
+
   # Dragonfly server pool (manager + scheduler) — fixed size, goes through extra node pools
   dragonfly_server_pool = var.install_dragonfly && var.dragonfly_server_node_pool.enabled ? [{
-    name              = var.dragonfly_server_node_pool.name
-    size              = var.dragonfly_server_node_pool.size
-    instance_types    = var.dragonfly_server_node_pool.instance_types
-    system_disk_size  = var.dragonfly_server_node_pool.system_disk_size
-    data_disk_enabled = true
-    data_disk_size    = var.dragonfly_server_node_pool.data_disk_size
+    name                  = var.dragonfly_server_node_pool.name
+    size                  = var.dragonfly_server_node_pool.size
+    instance_types        = var.dragonfly_server_node_pool.instance_types
+    system_disk_size      = var.dragonfly_server_node_pool.system_disk_size
+    data_disk_enabled     = true
+    data_disk_size        = var.dragonfly_server_node_pool.data_disk_size
+    use_akernel_data_disk = false
     labels = {
       (var.dragonfly_server_node_pool.node_label_key) = var.dragonfly_server_node_pool.node_label_value
     }
@@ -332,7 +268,7 @@ locals {
   }] : []
 
   # Seed pool is created as a dedicated resource (with autoscaling), not via extra
-  all_extra_node_pools = concat(var.extra_node_pools, local.dragonfly_server_pool)
+  all_extra_node_pools = concat(local.akernel_extra_node_pools, local.dragonfly_server_pool)
 
   dragonfly_values = var.install_dragonfly ? templatefile("${path.module}/values-dragonfly.yaml.tmpl", {
     storage_class              = local.effective_storage_class
@@ -504,8 +440,11 @@ resource "alicloud_cs_kubernetes_node_pool" "default_with_key" {
   dynamic "data_disks" {
     for_each = var.node_pool_extra_data_disk_enabled ? [1] : []
     content {
-      category = var.node_pool_extra_data_disk_category
-      size     = var.node_pool_extra_data_disk_size
+      category     = var.node_pool_extra_data_disk_category
+      size         = var.node_pool_extra_data_disk_size
+      auto_format  = "true"
+      file_system  = var.node_pool_extra_data_disk_fs_type
+      mount_target = var.node_pool_extra_data_disk_mount_path
     }
   }
 
@@ -546,8 +485,11 @@ resource "alicloud_cs_kubernetes_node_pool" "default_with_password" {
   dynamic "data_disks" {
     for_each = var.node_pool_extra_data_disk_enabled ? [1] : []
     content {
-      category = var.node_pool_extra_data_disk_category
-      size     = var.node_pool_extra_data_disk_size
+      category     = var.node_pool_extra_data_disk_category
+      size         = var.node_pool_extra_data_disk_size
+      auto_format  = "true"
+      file_system  = var.node_pool_extra_data_disk_fs_type
+      mount_target = var.node_pool_extra_data_disk_mount_path
     }
   }
 
@@ -589,6 +531,16 @@ resource "alicloud_cs_kubernetes_node_pool" "extra" {
       size     = coalesce(each.value.data_disk_size, var.node_pool_data_disk_size)
     }
   }
+  dynamic "data_disks" {
+    for_each = each.value.use_akernel_data_disk && var.node_pool_extra_data_disk_enabled ? [1] : []
+    content {
+      category     = var.node_pool_extra_data_disk_category
+      size         = var.node_pool_extra_data_disk_size
+      auto_format  = "true"
+      file_system  = var.node_pool_extra_data_disk_fs_type
+      mount_target = var.node_pool_extra_data_disk_mount_path
+    }
+  }
 
   key_name = length(var.node_pool_key_name) > 0 ? var.node_pool_key_name : null
   password = length(var.node_pool_key_name) == 0 && length(var.node_pool_login_password) > 0 ? var.node_pool_login_password : null

deploy/terraform/aliyun/terraform.tfvars.example

@@ -56,23 +56,23 @@ node_pool_system_disk_size     = 100
 node_pool_data_disk_enabled    = true
 node_pool_data_disk_category   = "cloud_essd"
 node_pool_data_disk_size       = 100
+node_pool_extra_data_disk_enabled    = true
+node_pool_extra_data_disk_category   = "cloud_essd"
+node_pool_extra_data_disk_size       = 300
+node_pool_extra_data_disk_mount_path = "/home/akernel"
+node_pool_extra_data_disk_fs_type    = "xfs"
 
 # Set one of the following for node login.
 node_pool_key_name = "<your-ecs-key-name>"
 # node_pool_login_password = "<your-strong-password>"
 
-# --- Node /home/akernel storage mode (choose one) ---
-# Default: hostPath from system disk (/home/akernel). No extra config needed.
+# --- Alternative node /home/akernel storage modes ---
+# The default above is a dedicated XFS cloud disk. Set
+# node_pool_extra_data_disk_enabled=false to use the system-disk hostPath.
 #
-# Option A: Extra data disk — dedicated cloud disk per node, mounted to /home/akernel via user_data.
-#   No PVC scheduling delay, scales with nodes.
-# node_pool_extra_data_disk_enabled    = true
-# node_pool_extra_data_disk_category   = "cloud_essd"
-# node_pool_extra_data_disk_size       = 300
-# node_pool_extra_data_disk_mount_path = "/home/akernel"
-# node_pool_extra_data_disk_fs_type    = "ext4"   # ext4 or xfs
-#
-# Option B: CSI ephemeral PVC — dynamic cloud disk per pod.
+# Alternatively, use a CSI ephemeral PVC per node pod. Disable the dedicated
+# data disk when selecting this mode.
+# node_pool_extra_data_disk_enabled = false
 # node_home_use_csi_ephemeral = true
 # node_home_csi_storage_class = "alicloud-disk-essd"
 # node_home_csi_size          = "300Gi"