AWS Terraform architecture

Platform layers, services, IRSA roles, networking, and module dependencies for LangSmith self-hosted on AWS EKS.

This page documents what the AWS Terraform modules provision and how the modules wire the resulting deployment together.

Platform layers

LangSmith on AWS deploys in two stages with one optional add-on. The infrastructure stage provisions the cloud foundation. The application stage installs the LangSmith Helm chart. The LangSmith Deployment add-on is opt-in and adds the host-backend, listener, and operator services for managing LangGraph applications from the UI.

LangSmith Deployment add-on  (enable_langsmith_deployments = true)
  host-backend, listener, operator
  Per deployed graph: api-server, queue, redis, postgres (operator-managed)
  Requires: KEDA (installed alongside infrastructure via k8s-bootstrap)

LangSmith application  (deploy_langsmith = true)
  backend, frontend, playground, queue, ace-backend, clickhouse
  Storage: RDS PostgreSQL (metadata) + S3 (trace blobs via VPC endpoint)
  Ingress: AWS ALB | NGINX | Envoy Gateway | Istio

AWS infrastructure
  VPC + private/public subnets + single NAT gateway
  EKS cluster + managed node group + cluster autoscaler
  RDS PostgreSQL (private subnets)
  ElastiCache Redis (private subnets)
  S3 bucket + VPC Gateway Endpoint (no public route)
  ALB controller + EBS CSI driver + metrics server
  k8s-bootstrap: KEDA, ESO, optional Envoy Gateway
  Optional: Network Firewall, WAF, CloudTrail, ALB access logs

Component to storage mapping

Component	Storage backend	Access method
backend	RDS PostgreSQL	Private subnet, security group
backend	S3 bucket	IRSA + VPC Gateway Endpoint
clickhouse	EBS volume (GP3, EKS PVC)	Local
redis	ElastiCache or in-cluster	Private subnet, security group
LGP operator	RDS PostgreSQL (shared)	Private subnet, security group

Application core services

These pods run on every deployment. All write logs and metrics; the busier components (backend, queue, ingest-queue) scale horizontally.

Service	Purpose	Port	HPA	IRSA	Depends on
langsmith-frontend	React UI	3000	1 to 10	No	backend, platform-backend
langsmith-backend	Main API (traces, runs, projects, API keys, feedback)	1984	3 to 10	Yes (S3)	Postgres, Redis, ClickHouse, S3
langsmith-platform-backend	Org and user management, auth, billing, settings	1986	1 to 10	Yes (S3)	Postgres, Redis, S3
langsmith-playground	LLM prompt playground UI	3001	1 to 10	No	backend
langsmith-queue	Trace ingestion worker (Redis to ClickHouse + S3)	—	3 to 10 + KEDA	Yes	Redis, ClickHouse, S3
langsmith-ingest-queue	Dedicated high-throughput ingestion worker	—	3 to 10 + KEDA	Yes	Redis, S3
langsmith-ace-backend	Async compute (dataset runs, evaluations, background jobs)	—	1 to 5	No	Postgres, Redis
langsmith-clickhouse	Columnar store (trace spans, run metadata, eval results)	—	StatefulSet, single replica	No	EBS GP3 PVC

In-cluster ClickHouse is dev/POC only (single pod, no replication, no backups). For production use LangChain Managed ClickHouse or a self-managed external cluster.

One-time jobs

The Helm chart runs three jobs at install and upgrade time:

Job	Purpose
langsmith-backend-migrations	PostgreSQL schema migrations
langsmith-backend-ch-migrations	ClickHouse schema migrations
langsmith-backend-auth-bootstrap	Creates the initial org and admin account from initial_org_admin_password in langsmith-config

LangSmith Deployment add-on

When enable_langsmith_deployments = true, three additional services are installed and a LangGraphPlatform CRD is registered. Each deployment the user creates in the LangSmith UI produces a Kubernetes Deployment in the langsmith namespace, managed by the operator.

Service	Purpose
langsmith-host-backend	LangGraph control plane API. Manages deployment lifecycle, serves deployment metadata. IRSA for S3 access.
langsmith-listener	Watches host-backend for deployment state changes, creates and updates LangGraphPlatform CRDs. IRSA for S3 access.
langsmith-operator	Kubernetes operator. Reconciles LangGraphPlatform CRDs, creates and deletes Deployments and Services for each agent.

AWS managed services

When postgres_source = "external" and redis_source = "external" (the recommended production setting), Terraform provisions the following AWS managed services:

RDS PostgreSQL

• Default size: db.t3.large, private subnets, port 5432.
• Holds orgs, users, projects, API keys, settings.
• Secret flow: SSM /langsmith/{base_name}/postgres-password → ESO → langsmith-config.

ElastiCache Redis

• Default size: cache.m5.large, private subnets, TLS port 6379.
• Trace ingestion queue, pub/sub, short-lived cache.
• Secret flow: SSM /langsmith/{base_name}/redis-auth-token → ESO → langsmith-config.

S3 bucket

• Trace payloads: large inputs and outputs, attachments.
• IRSA via langsmith_irsa_role (no static keys). VPC Gateway Endpoint, no public internet.
• Prefixes: ttl_s/ (short TTL) and ttl_l/ (long TTL).
• Always required. Disabling blob storage breaks the cluster on large payloads.

SSM Parameter Store

• Centralized secret store for all LangSmith secrets.
• Flow: source infra/scripts/setup-env.sh writes secrets to SSM. The ESO ClusterSecretStore reads them and projects a langsmith-config Kubernetes Secret that the Helm chart mounts via config.existingSecretName.
• Prefix: /langsmith/{name_prefix}-{environment}/.

Cluster infrastructure

The k8s-bootstrap Terraform module installs the cluster-level services that LangSmith depends on:

Service	Namespace	IRSA	Purpose
aws-load-balancer-controller	kube-system	Yes	Provisions the AWS ALB from Kubernetes Ingress objects. Deleting the Ingress deprovisions the ALB and assigns a new DNS name on recreate, which breaks DNS records and OIDC redirect URIs.
cluster-autoscaler	kube-system	Yes	Scales EC2 node groups based on pod scheduling pressure.
ebs-csi-driver	kube-system	Yes	Provisions EBS volumes for PersistentVolumeClaims (used by ClickHouse).
KEDA	keda	No	Kubernetes Event-driven Autoscaling. Scales queue and ingest-queue on Redis queue depth. Required for the LangSmith Deployment add-on.
cert-manager	cert-manager	Optional (Route 53 IRSA when letsencrypt)	Automates TLS certificate issuance. Installed always; active for Let’s Encrypt only.
External Secrets Operator	external-secrets	Yes	Syncs SSM parameters into the langsmith-config Kubernetes Secret.

IRSA roles

IRSA replaces static credentials. The EKS cluster’s OIDC issuer is the trust anchor; service accounts in langsmith and kube-system are annotated with role ARNs and pods receive temporary credentials via the EKS token webhook.

Role	Defined in	Used by	Permissions
langsmith_irsa_role	modules/eks	backend, platform-backend, queue, ingest-queue, host-backend, listener	s3:GetObject, s3:PutObject, s3:DeleteObject, s3:ListBucket on the LangSmith bucket
aws_iam_role.eso	aws/infra/main.tf	ESO controller	ssm:GetParameter, ssm:GetParameters on /langsmith/*

Network topology

Default — ALB ingress

Internet
  → AWS Application Load Balancer (port 80 or 443, TLS via ACM or Let's Encrypt)
    → EKS Cluster (private subnets)
      • kube-system: aws-load-balancer-controller, cluster-autoscaler, ebs-csi-driver, keda
      • langsmith:   backend, frontend, playground, queue, clickhouse
                     redis (in-cluster) OR ElastiCache (private subnet)
                     RDS PostgreSQL (private subnet)
                     S3 bucket (VPC Gateway Endpoint, no public route)

Envoy Gateway — opt-in

Internet
  → AWS Network Load Balancer (NLB, ACM TLS termination at 443)
    → envoy-gateway-system: Envoy proxy (GatewayClass: eg, Gateway: langsmith-gateway)
      → langsmith namespace:        backend, frontend, playground, queue, clickhouse, ...
      → langsmith-agents namespace (optional dataplane): langgraph-dataplane listener + operator + agent pods
         (HTTPRoute attaches to shared langsmith-gateway via allowedRoutes: All)

Egress path with Network Firewall

When create_firewall = true, all outbound internet traffic from private subnets is inspected before reaching the NAT gateway:

EKS pods / RDS / ElastiCache (private subnets)
  → AWS Network Firewall (TLS SNI + HTTP Host inspection)
     ALLOWLIST: firewall_allowed_fqdns (default: beacon.langchain.com)
     DROP: all other established connections
  → NAT Gateway (public subnet)
  → Internet

Pod-to-pod, pod-to-RDS, and pod-to-ElastiCache traffic uses the local VPC route and never touches the firewall.

Ingress options

Four mutually exclusive ingress options ship with the modules. The choice determines whether split dataplane (agent pods in a separate namespace) is supported.

Option	Variable	Split dataplane	Traffic path	When to use
ALB (AWS LBC)	default	No	ALB → frontend NodePort	Default. Single-namespace deployments, POC, simplest TLS via ACM.
NGINX Ingress	enable_nginx_ingress = true	No	ALB → TGB → NGINX controller → frontend ClusterIP	When NGINX is the standard ingress in your organization.
Envoy Gateway	enable_envoy_gateway = true	Yes	ALB → TGB → Envoy proxy:10080 → HTTPRoute → services	Cross-namespace HTTPRoute routing. Recommended for split dataplane on new AWS deployments.
Istio	enable_istio_gateway = true	Yes	ALB → TGB → istio-ingressgateway:80 → VirtualService → services	Clusters with Istio already installed, or when an mTLS mesh is required.

Why ALB cannot support split dataplane

Standard Kubernetes Ingress is namespace-scoped. The ALB controller routes only to services in the same namespace as the Ingress resource. Agent pods in langsmith-agents are invisible to an Ingress in langsmith. Envoy Gateway and Istio both support cross-namespace routing via the Kubernetes Gateway API.

ALB plus Envoy Gateway (chained)

When the existing ALB already provides SSO (Okta or Cognito OIDC), WAF, and TLS, Envoy Gateway slots in behind it instead of replacing it:

Internet
  → ALB (unchanged: WAF, SSO, TLS, DNS)
    → Envoy Gateway NLB (internal-scheme, auto-provisioned by k8s-bootstrap)
       → HTTPRoute → langsmith namespace        (control plane)
       → HTTPRoute → langsmith-agents namespace (split dataplane)

The only change from the default ALB path is retargeting the ALB target group to the Envoy NLB. See helm/values/examples/langsmith-values-ingress-envoy-gateway.yaml in the modules repo for the values overlay.

TLS and DNS

The tls_certificate_source variable controls the certificate strategy:

Mode	Behavior	Compatible gateways
none	HTTP only, no certificate	Any
acm	HTTPS:443 with HTTP→HTTPS redirect. ACM certificate, auto-provisioned or BYO.	ALB, NGINX
letsencrypt	HTTPS via cert-manager + Let’s Encrypt DNS-01 (Route 53 IRSA)	Istio, Envoy

Why ACM versus cert-manager

ACM certificates are non-exportable. AWS attaches them directly to the ALB, which makes ACM the right choice when TLS terminates at the ALB. ACM cannot be used when TLS terminates inside the cluster (Istio Gateway, Envoy Gateway) because those gateways require the certificate material as a Kubernetes Secret.

cert-manager handles in-cluster TLS for Istio and Envoy. The letsencrypt value is a reference implementation: it installs cert-manager and a Let’s Encrypt ACME ClusterIssuer. In production, swap the ClusterIssuer for any cert-manager-compatible issuer.

Issuer	When to use
Let’s Encrypt (default)	Public domain, internet access, free
ACM Private CA (aws-privateca-issuer)	AWS-native, air-gap friendly, private domains, paid
Venafi (cert-manager-venafi)	Enterprise PKI, regulated environments
HashiCorp Vault (cert-manager-vault)	Self-hosted PKI
DigiCert, Sectigo, others	ACME or custom issuer plugins

The Terraform module provisions the cert-manager IRSA role and Route 53 permissions. Only the ClusterIssuer manifest changes between issuers.

Auto-provisioned DNS

When langsmith_domain is set and acm_certificate_arn is empty, Terraform activates the dns module which creates:

• A Route 53 hosted zone for the domain.
• An ACM certificate with DNS validation records.
• A Route 53 alias record pointing the domain to the ALB.

Staged deploy pattern: Set langsmith_domain with tls_certificate_source = "none" first. Terraform creates the hosted zone and certificate without blocking on validation. Delegate the NS records at your registrar, then flip to tls_certificate_source = "acm" in a later apply. Terraform blocks until the certificate validates and wires it into the HTTPS listener.

Bring your own certificate

Set acm_certificate_arn directly to skip the dns module. For in-cluster gateways, create a Kubernetes TLS Secret manually and reference it in the Gateway or VirtualService.

Module dependency graph

vpc ─► firewall (optional, create_firewall = true)
│
├─► eks ─► k8s-bootstrap (KEDA, ESO, Envoy Gateway [opt-in])
│            └─► cert-manager (Let's Encrypt DNS-01 via Route 53 IRSA)
│
├─► postgres    (RDS, private subnets from VPC)
├─► redis       (ElastiCache, private subnets from VPC)
├─► storage     (S3 bucket + VPC Gateway Endpoint)
├─► alb         (pre-provisioned ALB, public subnets)
│     └─► alb_access_logs (S3 bucket for access logs, opt-in)
├─► dns         (Route 53 zone + ACM cert, optional)
├─► bastion     (jump host for private EKS access, optional)
├─► cloudtrail  (audit logging, optional)
├─► waf         (WAF ACL on ALB, optional)
└─► firewall    (Network Firewall egress filter, optional)
       all ─► langsmith (root module)

Opt-in security modules

Module	Variable	Default	Purpose
Network Firewall	create_firewall	false	FQDN-based egress filtering. Allows only domains in firewall_allowed_fqdns (TLS SNI + HTTP Host). Requires create_vpc = true. Cost ≈ $0.40/hr/endpoint + $0.065/GB processed.
ALB access logs	alb_access_logs_enabled	false	Traffic analysis and compliance
CloudTrail	create_cloudtrail	false	API call logging. Skip if an organization trail already exists.
WAF	create_waf	false	WAFv2 Web ACL — OWASP Top 10, IP reputation, known bad inputs

Default resource sizes

Resource	Default	vCPU	Memory
EKS node	m5.4xlarge	16	64 GB
RDS PostgreSQL	db.t3.large	2	8 GB
ElastiCache Redis	cache.m6g.xlarge	4	13.07 GB
RDS storage	10 GB	—	—

For production sizing recommendations, see the scaling guide and the AWS deployment guide.

Validated behaviors and known constraints

These constraints were validated during the April 2026 gateway permutation test run.

#	Area	Constraint or fix
1	ACM wildcard SANs	langchain.com has 0 issue "amazon.com" CAA but not 0 issuewild "amazon.com". Wildcard SANs fail with CAA_ERROR. The dns module requests only the apex domain.
2	In-cluster Redis	The LangSmith Helm chart deploys Redis without requirepass. The k8s_bootstrap module writes redis://langsmith-redis:6379. Do not add an auth token unless you also configure the Helm chart Redis values.
3	name_prefix length	Maximum 15 characters. Names like dz-nginx-tst (12 characters) are valid.
4	Istio port	Istio 1.23+ ingressgateway listens on port 80 via NET_BIND_SERVICE, not port 8080. ALB TGB health check and security group rules must target port 80.
5	NGINX TGB port	NGINX ingress-nginx controller pods listen on port 80. The TargetGroupBinding target type is ip.
6	Envoy proxy port	Envoy proxy pods listen on port 10080 (not 80) when running as non-root. The TGB servicePort must be 10080.
7	Destroy order	Always run terraform destroy first and let Terraform handle namespace and Helm release lifecycle. Pre-deleting namespaces causes the helm_release resource to time out because Helm cannot uninstall cleanly into a terminating namespace.
8	Stuck terminating namespaces	KEDA’s stale external.metrics.k8s.io/v1beta1 API group causes NamespaceDeletionDiscoveryFailure. Fix: kubectl delete apiservice v1beta1.external.metrics.k8s.io before re-running terraform destroy.

Verification commands

# EKS cluster status
aws eks describe-cluster --name <cluster-name> --query "cluster.status"

# Node health
kubectl get nodes -o wide

# ALB status
kubectl get ingress -n langsmith

# RDS status
aws rds describe-db-instances \
  --query "DBInstances[?DBInstanceIdentifier=='<db-id>'].DBInstanceStatus"

# ElastiCache status
aws elasticache describe-replication-groups \
  --query "ReplicationGroups[?ReplicationGroupId=='<group-id>'].Status"

# S3 access from a pod (via VPC endpoint)
kubectl run s3-test --rm -it --image=amazon/aws-cli -n langsmith -- \
  aws s3 ls s3://<bucket-name>

---

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.