What LLM observability costs when you build, partner or buy SaaS
A team with LLM features in production needs to decide how traces, evaluations and cost records will be stored. The core of Langfuse is MIT licensed with no feature limits, so the software itself carries no license fee. What you pay for is the infrastructure that runs it, the instrumentation work to send traces from your applications, the operations to keep it running, and any Enterprise license for features such as project-level RBAC, data retention policies, audit logs and server-side data masking, which the Langfuse license key documentation lists. The self-hosted stack adds PostgreSQL, ClickHouse, Redis or Valkey and S3-compatible blob storage, and ClickHouse is usually the largest recurring line item because trace volume grows with every model and tool call.
SaaS tracing plans are priced differently. The Arize pricing page lists tiers that are metered by spans and retention, and the LangSmith pricing page meters base traces and seats. Those are subscription fees that recur with usage, and the higher tiers add retention and features on top. Self-hosting shifts the spend toward your own cloud or data center bill and your staff time. Before comparing, confirm the figures that drive both sides: your monthly trace or span volume, spans per agent request, how many days you must keep prompts and outputs, which regions the data may sit in, and which roles and masking controls you need. Ask any implementation partner for a written quote that separates one-time work (installation, instrumentation, upgrade testing) from recurring work (support, upgrades, backups, storage growth), with the Enterprise license cost shown as its own line.
Do not decide on license cost alone. Retention and data residency often set the answer. If prompts contain personal data or regulated client data, a SaaS plan may need a data residency review, and some controls you need may sit only in a paid tier on either side. Keep a short worksheet with volume, retention, residency, features and the team that will run the platform, then compare two or three options on the same assumptions. DeployOpen does not publish a price list, and any figure in a quote comes from the written scope of your own environment.
Which costs are one-time and which recur
Self-hosting Langfuse moves most of the cost into two buckets: project work that happens once, and infrastructure and staff time that repeat for as long as traces are kept. Each card below names one cost, who pays for it, and what to ask for in a quote, so the comparison with a SaaS tracing plan is made on the same terms.
One-time design and instrumentation
Scoping which applications, agents and evaluation flows are traced, designing the trace hierarchy, and instrumenting each service through the Langfuse SDK or OpenTelemetry. This work is usually a project fee paid to an implementation partner or carried by your own engineers. Ask the quote to list each application, the framework it uses, and whether instrumentation includes testing that traces arrive with correct cost and user attributes.
One-time masking and access decisions
Deciding which prompt and output fields are masked before storage, who may view traces that contain client data, and which roles exist in each project. Some controls, such as project-level RBAC, data retention policies and server-side masking, need a Langfuse Enterprise license key, so this decision also sets the license line in the budget. Ask the partner to document each decision in writing before deployment.
Recurring compute and storage on your cloud bill
The web, worker and PostgreSQL containers, Redis or Valkey, ClickHouse and S3-compatible blob storage run in your own accounts and are billed by your cloud provider or data center team. Langfuse lists minimum resources of 2 CPU and 8 GiB for ClickHouse and at least 16 GiB of ClickHouse memory for larger deployments. Ask for a sizing estimate based on your expected trace volume and retention period, not on the minimums.
Recurring backups and upgrades
Backups of PostgreSQL, ClickHouse and blob storage add storage and job time each month, and restore tests add staff time. Langfuse applies database migrations automatically at startup for minor updates, but major upgrades follow separate migration guides that require ClickHouse 25.12 or later and advise about 3x ClickHouse disk headroom when backfilling historic data. Ask the partner who runs upgrades, how often, and whether restore tests are included in the retainer.
Recurring support and staff time
Ongoing support can come from a retainer with a partner, a Langfuse Self-Hosted Enterprise plan that adds a named support engineer and a support SLA, or your own platform team. Each option is a recurring invoice or a recurring internal cost. Ask each vendor for the response terms in writing and count the internal hours for on-call, access reviews and evaluation scoring, since those recur for the life of the system.
SaaS fees that scale with usage
A SaaS tracing plan bills on a recurring subscription, often with seats, traces or spans and a retention limit as the drivers. Arize's published plans, for example, set monthly span allowances and 15-day or 30-day retention on its lower tiers, while LangSmith's paid plans are priced per seat with base trace limits. These fees sit on a vendor invoice and rise with traffic. Ask each SaaS vendor for the price at your projected volume, the retention you would need for audit evidence, and the cost of exporting data out.
Cost components: in-house, partner and SaaS
This table compares the cost components of a self-hosted Langfuse deployment, instrumentation work, the optional Enterprise license and SaaS tracing plans. Read each row for who pays and whether the cost is one-time or recurring, because the self-hosted runtime shifts most spend to infrastructure and people. The main takeaway is that the software license is often not the largest line item, so the quote should separate runtime, storage and effort.
| Cost component | Who pays and how | One-time or recurring, and main driver |
|---|---|---|
| Langfuse Web and Worker containers | Your platform team runs them in your cloud account or data center; the open source version has no license fee | Recurring compute. Driven by ingestion volume and worker scaling on CPU load |
| PostgreSQL (minimum 2 CPU, 4 GiB) | Your infrastructure or managed database budget; the scaling guide sets the minimum | Recurring. Driven by backup, high availability and retention of metadata |
| ClickHouse (minimum 2 CPU, 8 GiB; 16 GiB or more for larger deployments) | Your infrastructure; typically the largest storage and memory line for trace data | Recurring. Driven by span volume, retention period and disk headroom for backfills |
| Redis or Valkey (minimum 1 CPU, 1.5 GiB) | Your infrastructure; small compared with ClickHouse | Recurring, low. Mostly fixed once sized |
| S3-compatible blob storage | Your storage account or on-premises object store | Recurring. Driven by stored prompts, outputs and retention rules |
| Instrumentation work | Partner or internal engineers add tracing to each application and agent, usually billed as fixed-scope effort | One-time for each application, plus ongoing work when prompts or frameworks change |
| Enterprise license (optional) | Paid to Langfuse under a custom-priced Self-Hosted Enterprise plan; adds RBAC roles, data retention policies, audit logs, server-side masking and SCIM | Recurring subscription. Ask for the quote to list which features and support levels are included |
| LangSmith SaaS plans | Paid to LangChain by seat and trace volume; Developer lists $0 for 1 seat and up to 5k base traces a month, Plus lists $39 per seat per month with up to 10k base traces | Recurring. Driven by seats and base trace counts; self-hosted options exist only in Enterprise |
| Arize SaaS plans | Paid to Arize by span volume and retention; AX Free lists $0 with 25k spans a month and 15-day retention, AX Pro lists $50 a month with 50k spans and 30-day retention | Recurring. Driven by span volume and retention; check whether your retention duty exceeds these limits |
How SaaS tracing plans are priced
LangSmith's published pricing combines a per-seat fee with a monthly allowance of base traces. The Developer plan is listed at $0 per seat per month for 1 seat with up to 5k base traces a month, and the Plus plan is listed at $39 per seat per month with up to 10k base traces a month. Both listed tiers are cloud-hosted only, so the seat count and trace allowance are the two levers that move the bill on those plans. Enterprise is priced on request and adds self-hosted and hybrid options, custom SSO, ABAC and RBAC.
- Seats scale the fee linearly: every engineer, reviewer or auditor who needs access on a paid tier adds a per-seat charge each month.
- Base traces set the volume ceiling: once the monthly allowance is exceeded, ask how overage is billed before you commit to a tier.
- Cloud-only on listed tiers: if trace data cannot leave your infrastructure, the listed plans are not an option and you are in Enterprise territory.
- Enterprise is custom: self-hosted or hybrid deployment is quoted on request, so the fee, support terms and deployment scope must be written into the order.
- Retention terms: confirm how long traces are kept on each tier and whether that meets your record-keeping duties.
Before signing, ask the vendor to model your expected trace count per month, your seat count in 12 months and the data residency of the cloud region.
Arize prices by spans rather than by traces, and retention is part of the tier. The AX Free plan is listed at $0 with 25k spans a month and 15-day retention. AX Pro is listed at $50 a month with 50k spans and 30-day retention. AX Enterprise is custom-priced and offers SaaS or self-hosted deployment with enterprise SSO. Data residency options are listed as US, EU or CA, so residency is a tier and region decision rather than an afterthought.
- Spans multiply quickly: an agent that makes many model and tool calls produces many spans per request, so span volume can outgrow a trace-based estimate.
- Retention is tied to the tier: the 15-day and 30-day windows on the listed plans may be shorter than the retention your policy or regulation requires.
- Residency is selected: confirm the region the account is hosted in, and whether backups, support access and logs stay within that region.
- Self-hosting is a separate price: the Enterprise tier is custom, so the deployment model and its operating costs must be quoted separately.
- Overage and upgrades: ask what happens when a month exceeds the span allowance and whether retention can be extended without changing tier.
Ask the vendor to confirm in writing the region, retention period, span limit and who can access stored data, then compare those answers with your own requirements.
Across both vendors, the bill moves with three variables: who uses the tool, how much telemetry each request generates, and how long the data is kept. Seat-based pricing rewards small expert teams and penalizes broad access for reviewers. Span- or trace-based pricing rewards short, simple requests and penalizes multi-step agents. Retention can be the hidden multiplier, because a longer window keeps more history under the same volume, and on some tiers it requires a higher plan.
- Illustrative example: a team that doubles its daily requests will roughly double its span count on a span-priced plan, while a seat-priced plan is unchanged until a new user is added.
- Retention versus volume: if retention rises from 15 to 30 days on a span-priced tier, storage and tier requirements both change.
- Prompt length matters: long prompts and tool outputs can increase the size of each record even when the span count is stable.
- Evaluation runs: scoring traces can create additional billable records or consume allowance, so include evaluation volume in your model.
- Growth horizon: model the next 12 to 24 months of usage, not only the pilot.
Ask for a worked estimate using your real request volumes, span counts per request and required retention, and verify that the estimate covers evaluation and support.
Data residency is the factor most likely to decide between a SaaS tracing plan and a self-hosted deployment. Arize lists US, EU and CA data residency options, so an EU deployment requires selecting that region at the tier you buy. LangSmith's listed Developer and Plus plans are cloud-hosted only, and its self-hosted and hybrid options sit under Enterprise. Traces often contain prompts, outputs and personal data, so the hosting region, sub-processors and support access all matter.
- Region is a contract term: confirm the hosting region for primary storage, backups and any replicas, not only the default region shown on the pricing page.
- Support and access: ask where vendor staff can access data and under what controls.
- Sub-processors: request the list of sub-processors that handle trace data and their locations.
- Hybrid versus full SaaS: determine whether trace payloads or only metadata leave your environment.
- Self-hosting as the alternative: a self-hosted Langfuse deployment keeps trace storage in your own accounts, though it carries infrastructure and operating costs of its own.
Before comparing prices, have your privacy and compliance leads confirm which regions and access patterns are acceptable for your trace data.
How to build your own cost estimate
This sequence turns a vague question about self-hosting versus SaaS into a sized estimate that you can compare line by line against a quote at the same trace volume and retention.
Inventory applications and agent steps to instrument
List every LLM application, chat assistant, RAG pipeline and agent that will send traces, and break each agent into its model calls, tool calls and retrieval steps. Name an owner for each application, since the owner decides which prompts and outputs are captured. Done when you have a table of applications with their instrumentation method, whether that is the Langfuse SDK or OpenTelemetry, and the team that will maintain it. This list also drives the instrumentation effort you will estimate in a later step.
Estimate trace and span volume per day
For each application, multiply expected requests per day by traces per request and spans per request, then add a margin for retries and growth. Agent workloads often produce many spans per request, so measure a few representative sessions in a test environment rather than guessing. Done when you have a daily trace count and a daily span count per application, plus a peak-day figure. Compare these numbers directly with the span allowances and monthly trace limits on SaaS pricing pages, because those plans meter on exactly these units.
Set retention and the regulatory minimum
Decide how long traces, scores and datasets must remain available, and who owns that decision (usually the compliance lead or model risk owner together with the application owner). Where the EU AI Act applies to a high-risk system, Article 19(1) requires providers to keep automatically generated logs for at least six months unless other law provides otherwise, so use six months as the floor and confirm the longer periods that sector rules or internal policy require. Done when you have a written retention period per data class. Retention is the single biggest driver of storage in a self-hosted build and of plan tier in a SaaS quote.
Size ClickHouse and blob storage
Translate daily span volume and retention into stored bytes, then size the ClickHouse nodes and object storage. Start from the Langfuse self-hosting minimums: 2 CPU and 8 GiB for ClickHouse, with at least 16 GiB of ClickHouse memory for larger deployments, and 2 CPU and 4 GiB each for Langfuse Web, Langfuse Worker and PostgreSQL, plus 1 CPU and 1.5 GiB for Redis or Valkey. Add disk headroom for backfills, since the Langfuse v3 to v4 guide advises about 3x ClickHouse disk headroom when backfilling historic data. Done when you have a storage and compute sizing sheet that your cloud or platform team has reviewed, and a note on whether the chosen ClickHouse version meets the 25.12 minimum for v4.
Price compute, storage and backups in your cloud
Price the compute, block and object storage, network egress and backup retention in the cloud accounts or data centers where the stack will run, using your own negotiated rates. Separate production from test, and include a second environment for upgrades, since Langfuse major upgrades require dedicated migration steps. Recurring costs here include nodes, storage growth over the retention window, and backup copies. One-time costs include initial provisioning and the migration of any backfilled history. Done when each line has a unit price from your own rate card and a monthly total, and you have asked your infrastructure team to confirm the backup restore time they can support.
Add instrumentation and operations effort
Estimate the people time for instrumenting each application, building masking into the SDK or OpenTelemetry Collector, configuring projects and roles, writing evaluations, and running upgrades and on-call support. These are the costs that SaaS vendors often exclude from their headline price. Ask your implementation partner for a written scope that lists each application, the instrumentation method, the features that need an Enterprise license key, and the support model, since the Self-Hosted Enterprise plan adds a named support engineer and a support SLA. Done when every effort line has an owner, a rate and an assumption that you can defend in review.
Add license and support costs
Confirm whether the core stack is enough for your needs. Langfuse states that all core features and APIs are available in its MIT-licensed open source version without limits, while features such as data retention policies, audit logs, server-side data masking, project-level RBAC roles and SCIM need an Enterprise license key. Treat any Enterprise license as a separate line with its own quote, and list the controls that depend on it. Done when you know which controls are included at no license cost and which require a paid quote, and you have a written support decision for each environment.
Compare with SaaS at the same volume and retention
Price the SaaS option using the same trace or span volume, the same retention period, the same number of seats and the same data residency requirement from step one. Pricing pages change, so pull current figures from each vendor and record the date. Check whether the plan meets your retention floor, where data is stored, whether SSO and roles are included, and whether the vendor allows the data to leave your region. Done when you have two totals for the same workload, a list of any requirement a SaaS plan does not meet, and a clear statement of what you would have to confirm with each vendor before signing.
How retention and logging duties change the bill
Retention is the largest lever on both sides of the comparison, and it works differently in each model. SaaS tracing plans cap how long traces are kept: the Arize plans listed on its pricing page keep spans for 15 days on AX Free and 30 days on AX Pro, so a longer obligation pushes you toward a higher tier or an enterprise contract. Self-hosting removes that cap but turns retention into a storage sizing decision. Every extra month of prompts and outputs adds ClickHouse disk, object storage and backup copies that your team or partner must provision, monitor and pay for, and the cost grows with trace volume rather than with seats.
Two articles of the EU AI Act are a useful input when you size retention for a high-risk system. Article 12(1) requires that such systems technically allow the automatic recording of events (logs) over their lifetime, and Article 19(1) requires providers to keep those logs for at least six months unless other law provides otherwise. Only your legal and compliance team can confirm which duties apply to your system, which role you hold, and whether sector rules set longer periods. Ask them for a written retention period per log type before you request a quote, because that number, not the tool's default, sets the storage budget.
Retention and audit controls also affect the license decision. In self-hosted Langfuse, retention policies and audit logs are Enterprise license features, listed in the license key documentation, so a retention requirement may move the deployment from the free open source tier to a custom-priced Self-Hosted Enterprise plan. Ask any partner to show which controls they will configure in the open source version, which need a license key, and how the license cost is quoted separately from implementation and infrastructure. Confirm the exact feature list with Langfuse before you finalize the budget.
What moves the cost up or down
Most of the cost of self-hosted Langfuse comes from the volume of data it stores, how long it keeps that data, and how much operational resilience the organization needs. The cards below name each driver and the direction it pushes the total, so a quote can be compared against the same assumptions.
Trace and span volume
Every prompt, model call, tool call and retrieval step adds rows to ClickHouse and objects to blob storage, so ingestion volume is the main driver of infrastructure size. Volume rises with traffic, with the number of spans per request, and with the length of stored prompts and outputs. Ask the partner to estimate spans per request from a real trace of your heaviest workflow, then multiply by expected daily requests. Volume up means storage, memory and worker capacity up.
Retention period
How long traces are kept sets how much storage accumulates and therefore how much disk is paid for and backed up. Retention policies are an Enterprise license feature in self-hosted Langfuse, so a longer or per-project retention schedule may move the total from the open source plan to a paid license. Shorter retention lowers storage cost but must still satisfy any record-keeping duty that applies to your AI system. Confirm the retention period with legal and compliance before the storage size is set.
Applications and agent steps
Each instrumented application, environment and agent adds traces, projects and integration work. Agents that make many model and tool calls per request multiply span counts, so an agent workload can cost several times more to store than a single prompt-and-answer feature. Instrumentation effort also grows with each framework that must be connected. Ask for a per-application inventory with estimated steps per request before agreeing a fixed scope.
High availability and production topology
Docker Compose on a single VM is documented for testing only and has no high availability, so it suits a proof of concept rather than a system that holds production records. Production runs on Kubernetes with Helm or Terraform, and it needs replicated web and worker containers, managed or clustered PostgreSQL, ClickHouse and Redis or Valkey, and tested backups. Moving from a single VM to a production topology raises both infrastructure and operations effort, and it is usually the largest one-time step in the budget.
Data residency and masking requirements
Keeping traces in your own storage, in a named region, with prompts masked before they are written, adds design and testing work. Client-side masking in the SDK or OpenTelemetry Collector can be built in-house, while server-side masking is an Enterprise license feature in self-hosted Langfuse. Stricter residency and masking rules increase engineering and review effort but reduce exposure if traces leak. Ask the partner which controls are built in the Collector, which depend on the license, and how each is tested.
Upgrade frequency and the v3 to v4 migration
Updates within a major version apply database migrations at startup, but the move from v3 to v4 changes the ClickHouse data model to observations-first and requires ClickHouse 25.12 or later, with 26.4 recommended, PostgreSQL 15 or later and Redis 7.0 or later. The guide advises about 3x ClickHouse disk headroom when backfilling historic data, so a migration can need temporary storage well beyond steady-state size. Plan upgrades as recurring work and ask the partner to price the migration separately from routine patching.
Questions buyers ask about LLM observability cost
How much does it cost to self-host Langfuse?
Langfuse itself has no license fee for the core, so the cost is the infrastructure and the people who run it. A self-hosted deployment runs Langfuse Web and Langfuse Worker alongside PostgreSQL, ClickHouse, Redis or Valkey, and S3-compatible blob storage. The Langfuse scaling guide lists minimum resources of 2 CPU and 4 GiB for web, worker and PostgreSQL, 1 CPU and 1.5 GiB for Redis or Valkey, and 2 CPU and 8 GiB for ClickHouse, with at least 16 GiB of ClickHouse memory for larger deployments. Your recurring cost is driven by trace volume, retention length and the cloud or data center capacity that hosts these services. Ask your team or a partner to model these against your expected ingestion before you fix a budget.
Is the core Langfuse platform free to self-host?
Yes, the core features and APIs are available in the MIT-licensed open source version without limits, and self-hosted deployments use the same codebase as Langfuse Cloud. The exceptions are the directories under ee/, web/src/ee/ and worker/src/ee/, which fall under a separate enterprise license. Features that need an Enterprise license key include project-level RBAC roles, data retention policies, audit logs, server-side data masking, and the SCIM and instance management APIs. If your requirements include those controls, price the Self-Hosted Enterprise plan from the Langfuse self-hosted pricing page as a separate line item.
How is a partner's fixed-scope work priced?
Most DeployOpen engagements are fixed-scope, so scope, deliverables and timelines are agreed in writing before work starts, and the price follows from that written scope. The work usually covers the deployment, instrumentation of your applications, storage sizing and backup, masking and access setup, and evidence for compliance reviews. Ongoing support, upgrades and continuous testing are available as retainers. DeployOpen does not publish a price list, so ask for a written scope, a list of assumptions, and a quote that separates one-time setup from recurring support.
How long does a self-hosted Langfuse deployment take?
The timeline depends on how many applications need instrumenting, how production-ready the environment must be, and what approvals your organization requires. Langfuse documents Docker Compose on a single VM for testing with no high availability, and Kubernetes with Helm or Terraform for production on AWS, Azure and GCP. A pilot on a single VM is a different project from a highly available production cluster with separate ingestion and user interface deployments. Ask any partner for a milestone plan tied to these deployment modes rather than a single number.
What does the work need from our internal team?
Your team provides the cloud or data center accounts, network and identity integration, storage and backup standards, and decisions about who may view traces that contain prompts or client data. Engineers on each instrumented application need time to add the SDK or OpenTelemetry instrumentation and review what is recorded. Security and compliance staff must approve masking rules and retention periods. Ask the partner to list these inputs by name and role in the written scope, so the dependencies are visible before the schedule is set.
Can our team deploy and run Langfuse without a partner?
Yes, if you have platform engineers who can run PostgreSQL, ClickHouse, Redis and object storage, and who can follow the Langfuse upgrade guides. Major upgrades such as v3 to v4 require ClickHouse 25.12 or later, PostgreSQL 15 or later and Redis 7.0 or later, and the v3 to v4 guide advises about 3x ClickHouse disk headroom when backfilling historic data. The harder work is usually ongoing: sizing storage, testing backups, keeping releases current, and setting evaluation and access controls. Estimate the operating hours for those tasks before you decide to build internally.
When is a SaaS tracing plan cheaper than self-hosting?
SaaS can be cheaper when trace volume is small, your data may leave your environment, and you do not need the enterprise controls. LangSmith lists a Developer plan at $0 per seat per month for 1 seat with up to 5k base traces a month, and Plus at $39 per seat per month with up to 10k base traces a month, both cloud-hosted only. Arize lists AX Free at $0 with 25k spans a month and 15-day retention, and AX Pro at $50 a month with 50k spans and 30-day retention. SaaS costs rise with seats, traces or spans and longer retention, while self-hosting costs rise with storage and operations. Compare both at your expected volume, and check data residency before you choose SaaS for prompts that contain personal data.
What to ask for in a written quote
A written quote for a self-hosted Langfuse deployment should state the components it covers and the assumptions behind them. Name each service: the Langfuse Web and Worker containers, PostgreSQL, ClickHouse, Redis or Valkey, and the S3-compatible blob storage. For each, the quote should give the CPU and memory sizing used and the reasoning, measured against the minimums in Langfuse's scaling guide. Ask whether the topology is a single VM for testing or a Kubernetes deployment for production, because that choice changes the cost of every line. Also ask for the trace volume and retention period the sizing assumes, since ClickHouse storage grows with both. Under the Langfuse self-hosting docs, the scope should say which cloud account or data center hosts each component.
The quote should also cover instrumentation scope, listed application by application. Ask for each LLM application, agent or pipeline to be named, with the SDK or OpenTelemetry path used, the prompts and outputs to be masked or not, and the evaluation work included. The quote should state which Enterprise license features are in scope, such as project-level RBAC, data retention policies, audit logs, server-side data masking and SCIM, because these are not in the open source version. Ask the partner to confirm in writing who buys the license key and who holds the support relationship. Then ask for backup and upgrade responsibilities: who runs backups and restore tests, who applies the migrations for each Langfuse major version, and who keeps the ClickHouse headroom the v3 to v4 guide calls for.
Finally, ask the quote to separate one-time work (design, deployment, instrumentation, hardening, initial evaluation setup) from recurring costs (infrastructure, license fees, support tier, upgrades and retained operations). DeployOpen does not publish prices. Quotes follow from a written, fixed scope agreed before work starts, so an unclear scope produces an unclear quote. Ask for the support tier explicitly and check whether it matches the Langfuse self-hosted plans described on the Langfuse self-hosted pricing page. Compare the result against SaaS tracing plans, using the seat, trace or span counts and retention limits each vendor lists, and confirm those figures on the vendor's current pricing page before you decide.
How DeployOpen works on this
DeployOpen plans and implements self-hosted Langfuse inside your own cloud accounts, data centers or air-gapped networks, so cost questions get answered from a written scope rather than a generic estimate. Engagements are fixed-scope, with scope, deliverables and timelines agreed in writing before work starts.
How we scope
- We list the applications, frameworks and model calls that need tracing, and the trace volume each one produces.
- We define which Langfuse features need an enterprise license key and which open source features cover the agreed needs.
- We document the storage components to size: PostgreSQL, ClickHouse, Redis or Valkey, and S3-compatible object storage.
- We agree in writing what the quote covers, including instrumentation work, deployment, testing and handover.
How we build
- We deploy Langfuse Web, Langfuse Worker and their dependencies on the infrastructure you choose, using Docker Compose for testing or Kubernetes for production.
- We instrument applications with the Langfuse SDKs or OpenTelemetry and set up masking in the SDK or Collector before data is stored.
- We test ingestion volume, cost attribution per project and retention settings against the scope before go-live.
- We configure SSO, project roles and backups for the storage layer as the written scope requires.
How we hand over
- We document the architecture, sizing assumptions, configuration keys and the upgrade path between Langfuse major versions.
- We provide a cost model that separates one-time implementation work from recurring infrastructure and support costs.
- We deliver runbooks for backup restore, storage growth, retention changes and access reviews.
- We confirm with you which evidence the deployment produces for model risk, audit or AI Act logging reviews.
How we support
- Ongoing support, upgrades and continuous testing are available as retainers after handover.
- Retainers can cover ClickHouse capacity reviews, Langfuse version upgrades and new application instrumentation.
- We recommend confirming with the vendor and your own environment which support tier, such as Langfuse's enterprise plan, you need.
