ClickHouse underneath. Langfuse on top. A runbook at the end.

Langfuse Activation

• We are on Langfuse Cloud and happy with it. Why would we need you?

  Stay on it. Langfuse Cloud is the right answer for most teams and we do not compete with it. Signing up is easy; what stalls is everything after tracing: nobody can say whether last week’s prompt change made answers better or worse, which evaluator to build first, or what a single agent run costs, and the prompts are still in the repo so every wording change is a deploy. Every Langfuse Activation scope runs on Langfuse Cloud.

• Could we just do this ourselves?

  Yes. Langfuse’s documentation is good, their workshop is public, and nothing we do is secret. What you are buying is the week. Someone who has already made the taxonomy mistakes, already calibrated a judge against human labels, and already reconciled cost figures to a provider invoice does in five days what a first-timer does in six weeks of nights and weekends.

• How does Langfuse billing work?

  A billable unit is one trace plus every observation plus every score, counted on ingestion. Two consequences people miss. Units Langfuse generates itself count, including LLM-as-a-judge output, annotation labels and experiment runs. And because units are counted on ingestion, deleting data never refunds a unit: retention reduces storage cost, not unit cost. We project unit volume before evaluators switch on across all traffic.

• What does the handover actually include?

  A recorded 90-minute session, a written runbook, and every artifact left in your project and your repo: the span taxonomy, the dashboards, the evaluator rubrics, the judge sampling rate and what it implies for your bill, and the alert configuration. Nothing sits in a Bonneville account. The test we hold ourselves to is whether it survives the engineer who leaves.

• Is SSO part of the Langfuse Enterprise licence?

  No. On self-hosted Langfuse, SSO and SSO enforcement are free in the open source build. This is the most common misconception in the stack and it costs teams real money. The Enterprise licence buys governance: project-level RBAC, audit logs, the admin and instance management APIs, data retention management, server-side masking, and the SOC 2 and ISO 27001 reports procurement will ask for. On Langfuse Cloud it works differently: SSO requires the Teams add-on on Pro, or the Enterprise plan.

ClickHouse

• Can we port our schema to ClickHouse table for table?

  You can, and it will fail quietly. Everything loads, the counts match, and the queries are slower than the source. The method that works is to mine the query history rather than the ER diagram, derive the schema from what people actually run, go wide on one event table per business process, and push dimension attributes onto the event at write time. The sort key is the expensive part to get wrong: ORDER BY and PARTITION BY are one-way doors, and changing them later means rebuilding the table.

• Why does our self-hosted ClickHouse bill keep climbing?

  Almost always one of three things, and a Health Check tells you which. The cluster was sized from trace count rather than observation count, so storage lands at ten or a hundred times the model. Retention never reclaims: lightweight deletes mark rows rather than removing them, and monthly partitions above the merge threshold are never rewritten, so deleted data keeps occupying disk; one publicly documented deployment carried 679 GiB of unreclaimed deleted rows in a single monthly partition. Or there is no TTL baseline and no storage tiering, so everything stays on hot storage forever.

• Do you quote a migration up front?

  No. We quote an assessment, and the migration is quoted from its findings. Most partners sell you a migration; this engagement tells you whether to do one. If it concludes that two of your five workloads should stay where they are, that is in the report. Assessment and code conversion are the hard parts of a migration and moving the data is the easy part, so skipping assessment is how teams discover the schema was wrong three months in.

• Will ClickHouse be cheaper than Snowflake, Splunk or Elasticsearch?

  Not always, and we will not promise it before seeing 90 days of your account usage. ClickHouse wins decisively on storage cost against Elasticsearch and Splunk, where you pay for data you rarely query, and on query cost under high concurrency. It can be more expensive than Snowflake on an idle-heavy workload that runs twice a day, because auto-suspend is genuinely hard to beat there. Scale-to-zero on idle services is the closest analogue, and whether it closes the gap depends on your tolerance for wake latency.

• Do we have to cut over to find out if it works?

  No. We run a shadow. Your existing system stays the system of record, the same data lands in ClickHouse in parallel, and the queries run against both on a schedule with automatic diffing. You get results-match-or-not, p95 on both sides, monthly cost on both sides, and an engineering-hours number. Nothing customer-facing changes, so during the shadow there is nothing to roll back. Cutover is a separate decision made with all of that in hand.