top of page

Why Data Vault Splits Hubs, Links and Satellites Into Separate Tables

Writer: Rhys Hanscombe
Rhys Hanscombe
3 days ago
3 min read

Data Vault training draws the same picture. Business keys go in hubs. Relationships go in links. The details and their history go in satellites. Then someone asks the obvious question: why three tables, when one would do?


That question started one of the longest threads on the Data Community forum in 2025. It ran to 31 posts, with strong views on both sides.


The problem: the split looks like extra work

The member who asked had read the main Data Vault books and held the training badges. He still wanted a reason for the split beyond the split itself.


Links you model early can turn out wrong once new data arrives. Fixing them adds work, and more tables mean more joins. He asked what a team would lose if business keys sat in the satellites and links went away.


What members say you gain


Change lands in new tables

The first reply set out the core case. Data Vault is built to be insert-only. A new source adds a new satellite. A new business idea adds a new hub. A new relationship adds a new link. A new rule or calculation goes into the business vault.


Each change adds a table and leaves the old ones alone. That means less code to retest and less risk to the people using the data.


Loads can run side by side

Another member pointed to how loads run. Hubs, links and satellites each have a simple load pattern, so many of them can run at once. That helps when you join up data from many source systems. It also suits short sprints, since you can add one piece at a time.


He also described hubs as an index of business keys, which keeps lookups fast.


Some data needs its own space

A long-time contributor listed other reasons for the split:

  • Audit. Each row carries its load date and the source it came from.

  • Personal data. You can keep sensitive fields in their own satellite and control or delete them there.

  • Real change only. Satellites store a new row only when the values change.

  • Solve hard problems once. Do the complex work at load time, so each query does less.


He also noted that business keys in a satellite are not banned.


Where the split costs you

The thread did not end in full agreement, and the doubts are useful too.

  • Links can need rework. A relationship modeled on early data may not hold as more data arrives.

  • Speed is unproven either way. One member asked for side-by-side tests and none were shared. Others pointed to key size. A hash key takes 16, 20 or 32 bytes. Business keys are often smaller, so joining on them can cost less.

  • Getting the latest row takes work. PIT tables solve this, but they add tables and loads of their own.


One member proposed a different pattern. It drops links and moves business keys into the satellites. He argued the end result is the same, with fewer table types to load.


A simple way to decide for your team

  1. Count your sources. The more systems feed one business idea, the more the split helps.

  2. Look at your rate of change. If sources and rules change often, adding tables beats rewriting them.

  3. Check for personal data. If you need to isolate or delete it, a separate satellite makes that easier.

  4. Test on your own platform. Run your real queries with your real key sizes before you take a side on speed.

  5. Start with the standard pattern. Write down each exception and why you made it.


Frequently asked questions


Why does Data Vault use separate hubs, links and satellites?

The split lets you add new sources, ideas and relationships as new tables, without changing the old ones. It also lets loads run in parallel and keeps a clear audit trail. Members also cited keeping personal data apart and storing only real changes.


Can I put business keys in a Data Vault satellite?

Yes. One experienced member noted it is not banned. Hubs still give you one row per business key, which helps you join data from many systems.


Is Data Vault slower because of the extra joins?

The forum thread did not settle this. Speed depends on your platform, your key sizes and how you query the data. Test your own workload, and use PIT and bridge tables where they cut repeated work.


Join the conversation

This post sums up a real thread on the Data Community forum. Read the full thread and add your own view.


Data Community has more than 4,000 members. It has run since 2018, with monthly meetups and a forum where members test ideas like these against real projects.


Join Data Community and join the forum to bring your own modeling questions.

bottom of page