14 years of NYC TLC yellow taxi data (2012–2025) — 168 monthly Parquet files, 19 columns, pulled specifically to cross 1 billion rows — scanned with a full multi-column aggregation query on a single GCP c3-standard-44-lssd instance. 19× faster than a pre-warmed Apache Spark cluster on the same query.
Every figure below was measured on a live Duck Data Master Guru instance running on GCP. Not simulated, not cherry-picked — actual API response times from a stress test across every feature.
2ms
SQL COUNT(*) 10 million rows
162ms
Full column profile SUMMARIZE · 10M rows
52ms
Fuzzy match Jaro-Winkler · 10M rows
10.8s
RandomForest train + score 2 features · 50k rows
ML-DSA-65
PQC signed exports NIST FIPS 204 · 50k rows
343 MB
CSV exported 10M rows · streamed
3–6s
NL → SQL → execute AI call + query end-to-end
90/100
Stability score 3rd benchmark run · all features
// TEST ENVIRONMENT
The instance that ran these tests.
This is an n2-standard-8 — the Standard compute tier. Every customer chooses their machine size. Bigger machines are available if your workload demands it.
Each row is a live API call against the running instance. Times are wall-clock from request to response including network round-trip from Phoenix, AZ to GCP us-central1.
Full session schema auto-injected · correct table list in reply
AI notebook generate
Notebook
4–7 s
AI generates full .ipynb · saved directly to /data/notebooks/ · instantly in JupyterLab
// RESPONSE TIME — VISUAL
SQL operations finish before you finish blinking.
The analytics engine processes data at memory speed. Only AI calls (Gemini) add latency — the query execution itself is near-instant at any scale.
SQL COUNT(*) 10M rows
2ms
SQL GROUP BY + SUM 10M rows
84ms
SUMMARIZE profile 10M rows
162ms
Fuzzy match 10M rows
52ms
PQC sign 50k rows
204ms
NL → SQL → result (AI)
3–6s
RandomForest train 50k rows
10.8s
CSV export 343 MB
~8s
* Bar widths are log-scaled for readability. AI call latency (NL, Chat, Suggest) is Gemini network time, not compute time — SQL execution within those calls is <10ms.
// VS MANAGED CLUSTERS
What you're replacing.
Duck Data Master is purpose-built for analytics workloads that finish in seconds on a single node. Here's the honest comparison.
⚡ Databricks / Snowflake
Monthly cost$3,000–$15,000+
InfrastructureShared multi-tenant cluster
Startup time30s–5min (cold cluster)
Data moves to their cloudYes
AI/NL queriesAdd-on / extra cost
ML scoringDatabricks ML / MLflow
PQC signed exportsNot available
Built-in notebookYes (shared compute)
🦆 Duck Data Master Guru
Monthly cost$99 + GCP at cost + 10%
InfrastructureDedicated GCP VM, your account
Startup timeAlways on — 0s
Data moves to their cloudNever — stays in your region
AI/NL queriesIncluded · 2,000/day
ML scoringBuilt-in · RandomForest / GBM / LR
PQC signed exportsML-DSA-65 (NIST FIPS 204)
Built-in notebookYes · JupyterLab + in-dashboard
// METHODOLOGY
How we measured.
Dataset: NYC Yellow Taxi trip records, January–December 2024 — 12 Parquet files, ~10 million rows total, 19 columns including timestamps, fare amounts, GPS coordinates, and categorical fields. All files loaded into the analytics engine's persistent session before testing.
Timing method: Wall-clock time measured from HTTP request to full JSON response received. Tests run sequentially from Phoenix, AZ (home network) over HTTPS. Times include network round-trip latency (~20ms baseline to GCP us-central1) and exclude browser rendering.
AI call latency: NL query, Python NL, Chat, and Notebook Suggest all route through Gemini (Vertex AI). The 3–6 second range reflects typical Gemini response time. SQL execution within those calls is measured separately at <10ms.
ML training: scikit-learn RandomForestClassifier, 100 estimators, 80/20 train/test split, random_state=42. Training auto-samples to 100,000 rows maximum — sufficient for production model quality. A memory guard (4 GB free RAM required) prevents OOM on large feature sets.
PQC signatures: ML-DSA-65 (CRYSTALS-Dilithium3, NIST FIPS 204) implemented via dilithium-py. Signs the SHA-256 digest of the exported CSV. Public key is 1,952 bytes (3,904 hex chars). Signature files (.sig) are written alongside the export.
90/100 score: Measured across 13 features in 3 benchmark runs. The 10 points reflect real-world constraints — not system failures: GCS ingest requires a configured bucket, ML has no streaming progress bar yet (visual only, results are correct), and DB attach is environment-dependent. Core analytics (SQL, NL, Profile, Fuzzy, Join, Export, PQC, Chat, Notebook, Extract) all scored 88–95.
Honest note: These numbers represent a lightly loaded single-user instance. Performance scales with your GCP machine tier. The n2-standard-8 is the Standard tier — larger instances (up to 176 vCPU / 704 GB RAM) deliver proportionally faster results.
// VM TIERS — PICK YOUR COMPUTE
More machine when you need it.
All benchmarks above were run on the Standard tier (n2-standard-8). Scale up to your workload — you pay GCP's exact rate + 10%. Turn it off when you're done.
Tier
vCPU
RAM
Rate
Best for
Starter
4
16 GB
$0.19/hr
Up to ~25M rows, light analytics
Standard(these benchmarks)
8
32 GB
$0.39/hr
Up to ~100M rows, typical analytics
Pro
22
88 GB
$1.07/hr
Up to ~500M rows, heavy ETL
Power
44
176 GB
$2.14/hr
Billion-row datasets, ML at scale
Ultra
88
352 GB
$4.29/hr
Full data warehouse replacement
Guru
176
704 GB
$8.57/hr
Enterprise-scale, real-time at any volume
Stopped instance: ~$8–12/mo disk-only cost. Zero compute when idle. We tell you to stop it — our revenue doesn't depend on you forgetting.
// GET STARTED
Run these benchmarks on your data.
3-day free trial. Full Guru access. Your GCP instance provisions automatically. No credit card required.