LangSmith Self-Hosted — Cluster Sizing Calculator

Cluster resource requirements per selected component · Based on Helm chart v0.16+ defaults

Layout

Components & load profile

Components

TracesLangSmith observability — ingest, query, UI. The floor for everything in the langsmith chart.DeploymentsLangSmith-managed agent servers — host-backend, listener, and the operator that provisions each one.FleetAgent Builder — api-server, tool-server, trigger-server, queue.also enables DeploymentsInsightsClio conversation analysis — api-server plus its own queue and backing stores.PollyAutomated evaluation — api-server plus its own queue and backing stores.Standalone AgentsOne langgraph-cloud release per agent server. Separate chart — needs no LangSmith control plane.

Load Profile

Concurrent users (reads)10

Low 8Med 30High 50

Traces / second (writes)≈ 130M / month50

Low 15Med 200High 1k

Reads: MediumWrites: Medium

Databases, agents & options

PostgreSQL (LangSmith)

Internal StatefulSet

Not recommended for production

External managed

RDS / Cloud SQL — recommended

Traces Database (Storage)

In-cluster — single instance

OK for low/medium read load

In-cluster — 3-node replicated

Required for high read load

External / ClickHouse Cloud

No cluster resources consumed

SmithDB

Enable SmithDBBETA

S3-native trace database — fixed pod footprint, arm64 node pool. Requires ClickHouse (above) to stay active in dual-write mode.

Redis (LangSmith)

Internal StatefulSet

Default 8 GiB — fine for low writes

External managed

Required for high writes (≥26 GiB)

Trace Retention (TTL)

~60 GiB/day (med writes), ~600 GiB/day (high writes) without blob. With blob enabled, CH stores only references + search tokens (~20% of that).

Options

Blob storage (S3 / GCS / Blob)

Required at high write scale

HPA + KEDA autoscaling

Replica counts become ceilings; tiles also show the steady-state floor

Sandboxes

Enable sandboxesBETA

Code-execution workloads — adds 1 dedicated bare-metal node, not counted in vCPU/RAM totals

Cluster Resource Requirements

MediumMedium

Profile: Low→Med/Low→Med · 20% scheduling overhead included

Traces

Total vCPU needed

25 cores

services + in-cluster DBs

at rest 23 → ceiling 25 cores

Total RAM needed

56 GiB

services + in-cluster DBs

at rest 51 → ceiling 56 GiB

Est. worker nodes

4 nodes

8 vCPU / 32 GiB each

at rest 3 → ceiling 4 nodes

ClickHouse storage

84 GiB

14d TTL

Service Replicas

16 total pods

Service	Replicas	CPU req	RAM req
backend	4	4.0	8.0 Gi
platform-backend	3	3.0	6.0 Gi
ingest-queue	4	4.0	8.0 Gi
frontend	2	1.0	2.0 Gi
ace-backend	1	1.0	2.0 Gi
playground	1	0.5	1.0 Gi
queue	1	1.0	2.0 Gi
Services total	—	14.5	29.0 Gi

Storage

ClickHouse

84 GiB

14d TTL

PostgreSQL

50 GiB

+ enable autoexpansion

Redis

9 GiB

managed externally

ClickHouse requires SSD with 7,000+ IOPS and 1,000 MiB/s throughput.

Node Pool Guidance

Application node pool

4× — 8 vCPU / 32 GiB RAM

m5.2xlargen2-standard-8Standard_D8s_v38 vCPU / 32 GiB equivalent worker

ClickHouse node pool

1× — 6 vCPU / 17 GiB RAM / 84 GiB SSD

Dedicate to its own node pool or use node affinity to prevent resource contention.

Databases

In-cluster rows are counted in the vCPU/RAM totals above. External rows are what to provision outside the cluster — sized here, but not part of those totals. Open a row for instance types and requirements.

Instance types come from the Production terraform sizing profile: Real traffic, multi-replica with HPA. Roughly 20 concurrent users, 100 traces/sec.

Database	Runs	CPU req	RAM req	Storage
LangSmith PostgreSQLscales with the sizing profile	external	4	32 Gi	50 Gi	▸
ClickHouse	single	6	17 Gi	84 Gi
LangSmith Redisscales with write load	external	2	9 Gi	—	▸
In-cluster DB total	—	6.0	17.0 Gi	—

Azure Cache for Redis (Basic/Standard/Premium) is retiring — size new deployments against Azure Managed Redis.

Scaling self-hosted LangSmith · Terraform sizing profiles · External Postgres · External Redis

Architecture Diagram

values.yaml snippet

# ═══════════════════════════════════════════════════════════════════════
# REFERENCE GUIDANCE ONLY — not a production-ready values.yaml.
# This snippet shows which keys to configure and rough sizing based on your
# inputs. Review against the current chart defaults, fill in real secrets/
# connection strings, and validate in a staging environment before deploying.
# ═══════════════════════════════════════════════════════════════════════

# ── langsmith chart ────────────────────────
config:
  blobStorage:
    enabled: true

# ── Scaled services ─────────────────────────
frontend:
  deployment:
    replicas: 2

platformBackend:
  deployment:
    replicas: 3

ingestQueue:
  deployment:
    replicas: 4
  autoscaling:
    keda:
      enabled: true
      minReplicaCount: 3
      maxReplicaCount: 4

backend:
  deployment:
    replicas: 4

# ── PostgreSQL ──────────────────────────────
postgres:
  external:
    enabled: true
    connectionUrl: "<postgres-connection-url>"
# Required: 4 vCPU / 32 GiB / 50 GiB storage with autoexpansion (terraform Production profile)
# Floor from the scale guide: 2 vCPU / 8 GiB / 10 GiB, flat at every load
# PostgreSQL 14 or newer, admin user able to create: btree_gin, btree_gist, pgcrypto, citext, ltree, pg_trgm

# ── Redis ───────────────────────────────────
redis:
  external:
    enabled: true
    existingSecretName: langsmith-redis-secret
# Required: ≥9 GiB, 2 vCPU, Redis 5 or newer, or Valkey 8
# Dedicated to this installation — never shared with another LangSmith installation

# ── ClickHouse / SmithDB ────────────────────
clickhouse:
  statefulSet:
    persistence:
      size: 84Gi
    resources:
      requests:
        cpu: "6"
        memory: "17Gi"
      limits:
        cpu: "12"
        memory: "26Gi"