docs-rs: provision watcher ecs task

Author: marcoieni

terragrunt/modules/docs-rs/rds.tf

@@ -22,8 +22,8 @@ resource "aws_security_group" "db" {
     from_port       = 5432
     to_port         = 5432
     protocol        = "tcp"
-    description     = "Connections from web servers on ECS, bastion, and builder"
-    security_groups = [aws_security_group.web.id, var.bastion_security_group_id, aws_security_group.builder.id]
+    description     = "Connections from web servers on ECS, watcher, bastion, and builder"
+    security_groups = [aws_security_group.web.id, aws_security_group.watcher.id, var.bastion_security_group_id, aws_security_group.builder.id]
   }
 }
 

terragrunt/modules/docs-rs/watcher.tf

@@ -0,0 +1,146 @@
+# The watcher is a long-running background daemon (docs_rs_watcher binary) that:
+# - polls the crates.io registry index for new crate publishes and queues builds
+# - periodically updates GitHub/GitLab repository stats
+# - queues automatic rebuilds of old releases
+
+# Dedicated security group so we can grant the watcher DB access
+resource "aws_security_group" "watcher" {
+  vpc_id = var.cluster_config.vpc_id
+  name   = "docs-rs-watcher"
+}
+
+# The watcher doesn't serve HTTP, so we use ecs-task + raw aws_ecs_service directly
+# instead of the ecs-app module, which requires expose_http and always creates
+# a load balancer target group.
+module "watcher_task" {
+  source = "../ecs-task"
+
+  name           = "docs-rs-watcher"
+  container_name = "app"
+  cpu            = 256
+  memory         = 512
+
+  environment_variables = {
+    DOCSRS_PREFIX       = "/tmp"
+    REGISTRY_INDEX_PATH = "/tmp/crates.io-index"
+
+    DOCSRS_STORAGE_BACKEND = "s3"
+    S3_REGION              = "us-east-1"
+    DOCSRS_S3_BUCKET       = aws_s3_bucket.storage.id
+
+    DOCSRS_LOG     = "docs_rs=debug,rustwide=info"
+    RUST_BACKTRACE = "1"
+
+    # - How often to poll the registry for new crates (seconds)
+    DOCSRS_DELAY_BETWEEN_REGISTRY_FETCHES = "60"
+    # - Cap on how many old releases to queue for rebuild per hourly cycle
+    DOCSRS_MAX_QUEUED_REBUILDS = "10"
+  }
+
+  secrets = {
+    # Used by docs_rs_repository_stats to call the GitHub API when updating
+    # repository stars/forks/etc.
+    DOCSRS_GITHUB_ACCESSTOKEN = "/docs-rs/github-access-token"
+  }
+
+  computed_secrets = {
+    # The watcher needs direct DB access: it writes to the build queue,
+    # updates repository stats, and runs consistency checks.
+    DOCSRS_DATABASE_URL = aws_ssm_parameter.connection_url.arn
+  }
+
+  # No HTTP — the watcher is a background daemon with no exposed ports.
+  port_mappings = []
+  docker_labels = {}
+}
+
+resource "aws_ecs_service" "watcher" {
+  name            = "docs-rs-watcher"
+  cluster         = var.cluster_config.cluster_id
+  task_definition = module.watcher_task.task_arn
+  # Exactly 1 replica: the registry watcher must be a singleton, otherwise
+  # crates would be added to the build queue multiple times.
+  desired_count    = 1
+  launch_type      = "FARGATE"
+  platform_version = "1.4.0"
+
+  enable_ecs_managed_tags = true
+
+  # No load_balancer block — the watcher has no HTTP endpoint.
+  network_configuration {
+    subnets = var.cluster_config.subnet_ids
+    security_groups = [
+      var.cluster_config.service_security_group_id,
+      aws_security_group.watcher.id,
+    ]
+    # Public IP needed to reach external services (crates.io index, GitHub API,
+    # ECR, SSM) without NAT gateway — same pattern as the web service.
+    assign_public_ip = true
+  }
+}
+
+# S3 permissions: the watcher deletes documentation from S3 when removing
+# crates/versions (delete_crate, delete_version) and reads/writes via the
+# storage abstraction layer.
+resource "aws_iam_role_policy" "watcher_s3" {
+  role = module.watcher_task.task_execution_role_id
+  name = "inline"
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [
+      {
+        Effect = "Allow"
+        Action = [
+          "s3:PutObject",
+          "s3:GetObjectTagging",
+          "s3:PutObjectTagging",
+          "s3:GetObject",
+          "s3:PutObjectAcl",
+          "s3:ListBucket",
+          "s3:DeleteObject",
+        ]
+        Resource = [
+          aws_s3_bucket.storage.arn,
+          "${aws_s3_bucket.storage.arn}/*",
+        ]
+      }
+    ]
+  })
+}
+
+# CI: reuse the web module's OIDC role rather than creating a new one.
+# The gha-oidc-role module derives the IAM role name from org/repo/environment,
+# so a second module instance for the same rust-lang/docs.rs repo would
+# conflict. Instead we attach the watcher's ECR and ECS permissions to the
+# existing role.
+resource "aws_iam_role_policy" "watcher_oidc_update_service" {
+  name = "update-ecs-service-watcher"
+  role = module.web.oidc_role_id
+
+  policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [
+      {
+        Sid    = "AllowUpdate"
+        Effect = "Allow"
+        Action = [
+          "ecr:GetAuthorizationToken",
+          "ecs:UpdateService",
+          "ecs:DescribeServices",
+        ]
+        Resource = aws_ecs_service.watcher.id
+      }
+    ]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "watcher_oidc_ci_pull" {
+  role       = module.web.oidc_role_id
+  policy_arn = module.watcher_task.policy_pull_arn
+}
+
+resource "aws_iam_role_policy_attachment" "watcher_oidc_ci_push" {
+  role       = module.web.oidc_role_id
+  policy_arn = module.watcher_task.policy_push_arn
+}

terragrunt/modules/ecs-app/outputs.tf

@@ -1,3 +1,7 @@
 output "role_id" {
   value = module.ecs_task.task_execution_role_id
 }
+
+output "oidc_role_id" {
+  value = local.oidc_role_id
+}