Self-hosted ETL buyer's checklist: 12 questions, 12 verifiable answers
Every vendor says their self-hosting is better. Nobody publishes how to verify any of them. These twelve questions are written so that the answer is a fact, not a feeling.
Buyer's guide · 2026-11-03 · Sumeta engineering
A self-hosting claim is only as good as the questions used to test it. Twelve questions in four groups — deployment, data, operations and commercial — convert the claim into evidence you can check in a POC.
How to score it
Each question is scored 0, 1 or 2: 2 when there is a verifiable answer, 1 when the answer is a claim without evidence, 0 when the answer is no or unknown. A shortlist should be compared on the total, not on the demo.
This checklist applies to any vendor. That is deliberate — a method is more useful, and more quotable, than a pitch.
Deployment group
| # | Question | A good answer sounds like | How to verify in a POC |
|---|---|---|---|
| 1 | Do extraction, transformation and load all run in our VPC? | Yes, including the control plane | Run a job with egress capture |
| 2 | Can it run with no outbound connection except our targets? | Yes; activation is local | Block egress except targets and re-run |
| 3 | Does upgrading require external connectivity? | No; offline upgrade path exists | Upgrade with the network isolated |
Data group
| # | Question | A good answer sounds like | How to verify |
|---|---|---|---|
| 4 | Which of the four data flows leave our environment? | A flow-by-flow answer | Compare against packet capture |
| 5 | Where are temporary files written? | A named local path or volume | Inspect the volume during a run |
| 6 | Where is schema metadata stored? | In the local control-plane store | Open the store and read a table definition |
Operations group
| # | Question | A good answer sounds like | How to verify |
|---|---|---|---|
| 7 | Can job logs and alerts be retained locally? | Yes, with configurable retention | Disable log shipping and confirm logs remain |
| 8 | What is the backup and restore procedure? | Documented, tested, with RPO/RTO | Restore a pipeline from backup |
| 9 | How is a failed batch rolled back? | Atomic rollback per batch | Force a mid-batch failure and inspect the target |
Commercial group
| # | Question | A good answer sounds like | How to verify |
|---|---|---|---|
| 10 | What exactly is the billing unit? | Capacity: nodes, vCPU, environments | Read the contract, not the slide |
| 11 | What happens above capacity? | A hard ceiling with notification | Ask for the clause in writing |
| 12 | How do we get our data out at exit? | Standard formats, no export fee | Ask for the exit clause |
Scoring your shortlist
| Score | Interpretation | Next step |
|---|---|---|
| 18–24 | Verifiable self-hosting | Proceed to a targeted POC |
| 10–17 | Hybrid in practice | Decide whether the control plane is acceptable |
| 0–9 | Managed SaaS with self-hosting language | Reframe the requirements or change vendor |
Run the checklist against us
Download the build and score all twelve questions in a single POC.