Every clean run returned the expected assertions.
First-party local experiment · 2 September 2026
We ran n8n.
Here is the boundary.
Five deterministic workload runs passed. One controlled failure was detected and recovered. A SQLite backup was restored. None of that makes this a cloud benchmark or a universal RAM recommendation.
Measured, then bounded
What happened on this machine.
The workflow generated 2,500 synthetic records, transformed them, grouped them into 20 buckets, and asserted exact totals. The full CLI process was measured, so startup and the task runner are included.
No customer, credential, or production workflow data.
End-to-end CLI time, not isolated node latency.
Working set on this machine, not a hosting recommendation.
Exact assertions · all five runs
The output did not merely “look right.”
- Item count
- 2,500
- Amount sum
- 126,250
- ID sum
- 3,126,250
- Unique buckets
- 20
Failure is part of operations
A happy-path demo was not enough.
The recovery exercise stayed isolated: no credentials, network calls, customer data, or external action nodes were present.
- 01
Induce a deterministic failure
A separate workflow returned a non-zero process exit. The runner detected the failure instead of counting it as success.
- 02
Correct the same workflow
The workflow was corrected and re-imported under the same workflow identity, then completed successfully.
- 03
Back up and restore SQLite
The isolated database was copied and restored. Its pre-read database hash and restored workflow material matched the recorded source.
Do not generalize past the evidence
Useful proof has an edge.
This experiment establishes actual n8n use and a repeatable local operating exercise. It does not settle the hosting decision for a different workflow.
Supported by this run
- n8n 2.37.7 was installed and executed locally
- The synthetic workload produced deterministic results
- A controlled failure was detected and corrected
- The isolated SQLite state could be restored
Still unmeasured
- n8n Cloud performance and availability
- Any managed or unmanaged hosting provider
- Your production payload, concurrency, and integrations
- A universally safe RAM, CPU, or latency target
Safe public artifact
Inspect the synthetic workload.
The download contains only the workflow fixture used for the bounded test. Runtime configuration, database files, encryption material, credentials, logs, and private account data are excluded.
SHA-2562B1BE51AC5C7F5E194F223BA764F6505B22DB0C2856E273A4304C492F25D32FE
Version pin: n8n 2.37.7 · Node.js 24.19.0 · isolated SQLite · five clean repetitions · Direct cost: ¥0 · Confirmed affiliate commission: ¥0.