HubSpot to Snowflake and BigQuery.
The CRM stops being the warehouse. HubSpot reporting is good until the question spans HubSpot, the product database and the billing platform. At that point the answer belongs in a warehouse, and the CRM should receive the aggregate rather than try to compute it.
Two dashboards. One question. Different answers.
Marketing computes pipeline one way in the CRM, finance computes it another way in the warehouse, and both are internally consistent. The problem is not the tooling. It is that nobody owns the definition, so every system is free to invent one.
Nobody owns the definition.
Warehouse — the definition gapA warehouse layer somebody owns.
Extraction that survives schema change
Fivetran, Stitch, Airbyte or a direct build, chosen on cost and on how much control you need. Someone adds a HubSpot property on Tuesday and the pipeline should not break.
Modelled tables with tests
Raw extracts are not a reporting layer. Modelled tables with documented definitions and tests on those definitions are what stop two dashboards disagreeing.
Slowly changing dimensions
Deal stage history, owner history and lifecycle transitions have to be captured as changes over time, or every historical report silently reflects today rather than the day in question.
Write-back of the aggregates
Health scores, product usage summaries and account-level metrics computed in the warehouse and written back to HubSpot, so the people in the CRM can act on them.
Cost control
Warehouse bills grow quietly. Incremental models, sensible scheduling and clustering keep the thing affordable at the volume you will reach in two years.
Governance and access
Personal data in a warehouse carries obligations. Row-level access, masking and a documented retention position are part of the build.
Built to a contract, not to a connector.
Draw the boundary
Which system owns which field, in writing, before a line of code. Most broken integrations are two systems both believing they are the source of truth.
Model the join
The key that matches a record on one side to a record on the other, and the rule for what happens when it does not match — because it will not match.
Build and instrument
Sync built with retries, idempotency and a dead letter path, plus logging you can read without opening a console.
Reconcile
Counts compared on both sides on a schedule, with a report that names the drift rather than a green tick that hides it.
Hand over
A runbook covering the three failures that actually happen, so the next person does not have to reverse-engineer the design from the logs.
The warehouse becomes a second silo.
Data lands, dbt models get written, and the insight stays in a BI tool nobody in sales opens. A warehouse project only pays for itself when the numbers come back to where decisions happen — on the record, in the workflow, in front of the rep.
What people ask before they commit.
Do we need a warehouse at all?
Not if your questions live inside HubSpot. You need one when reporting spans several systems, when you need history the CRM does not retain, or when the volume makes CRM reporting slow. We will tell you honestly which side of that line you are on.
Snowflake, BigQuery or Databricks?
Mostly it follows your cloud and your team. BigQuery if you are on Google, Snowflake for straightforward warehousing with good governance, Databricks if there is real data engineering or machine learning alongside.
Fivetran or build it ourselves?
Fivetran for speed and for not maintaining connectors. Build it when the volume makes managed pricing painful or you need transformation in flight. The maintenance burden is the deciding factor, not the licence.
Can warehouse data come back into HubSpot?
Yes, and it should. Scores and aggregates written back to properties are what turn analysis into action. Reverse ETL tools do this, and so does a small scheduled job if the volume is modest.
How long does it take?
Four to twelve weeks for a first warehouse layer with modelled tables and write-back. Extraction alone is much faster and much less useful.
Where people go from here.
One definition, owned by someone.
Tell us what you are running
What the system does today, where it breaks, and when it has to work. An engineer reads it — you get an answer inside one business day, not a sequence.