In one line
Five minutes of rough arithmetic tells you whether a design needs one machine or a hundred, gigabytes or petabytes, before anyone writes code.
The idea
You don’t need precise numbers; you need the right order of magnitude. Estimation answers:
- How many requests/events per second (average and peak)?
- How much storage per day, per year, in total?
- How much bandwidth in and out?
- How much compute (CPU, GPU-hours)?
- Roughly what will it cost?
How it works
Handy numbers
| Fact | Value |
|---|---|
| Seconds in a day | 86,400 ≈ ~10⁵ |
| Seconds in a month | ≈ 2.6 million |
| 1 million per day | ≈ 12 per second |
| 1 KB / 1 MB / 1 GB / 1 TB / 1 PB | 10³ / 10⁶ / 10⁹ / 10¹² / 10¹⁵ bytes (roughly) |
| Peak vs average | assume 2–10×, depending on burstiness |
Rough latencies (orders of magnitude)
| Operation | ~Time |
|---|---|
| Read from memory | nanoseconds to ~100 ns |
| Read 1 MB sequentially from SSD | ~1 ms or less |
| Round trip within an Azure region | ~1 ms or less |
| Round trip Europe ↔ India | ~100–150 ms |
| Read from Blob Storage (small object, hot tier) | ~tens of ms |
These matter for design: a page that makes 50 sequential cross-region calls will never be fast.
The method
- Write down assumptions (and label them as assumptions).
- Compute average rates, then apply a peak factor.
- Compute storage growth including replicas and metadata.
- Round aggressively. 86,400 → 100,000.
- Sanity-check: does the answer fit one machine, one database, one storage account?
On the platform: worked example
Assumptions (placeholders, replace with real ones):
- 20 sites × 10 robots × 1 image every 10 s while operating
- 16 operating hours/day
- 2D image average 3 MB; 10% of captures also produce a 50 MB 3D point cloud
- Metadata ~ 2 KB per image in MongoDB
Rate
- Images/s per robot = 0.1 → 200 robots → 20 images/s average
- Peak (commissioning, bursts) ×5 → ~100 images/s
- Per day: 20 × 16 h × 3,600 ≈ 1.15 M images/day
Storage per day
- 2D: 1.15 M × 3 MB ≈ 3.5 TB/day
- 3D: 115 k × 50 MB ≈ 5.8 TB/day
- Total raw ≈ ~9 TB/day → ~3.4 PB/year 😮
Metadata
- 1.15 M × 2 KB ≈ 2.3 GB/day → ~0.85 TB/year: comfortable for MongoDB with good indexes and maybe partitioning later.
Bandwidth in
- 9 TB / 57,600 s ≈ ~160 MB/s average (~1.3 Gbit/s), with peaks several times higher.
What the numbers tell us
- Raw storage cost dominates, so storage tiering and retention policy matter far more than API server size. See 10 Blob Storage Patterns.
- Do we really need to keep every frame? Sampling or deduplication at the edge could cut cost by 10×. That’s a product question worth asking.
- 100 images/s of validation is easy to scale horizontally with Functions or workers. Compute isn’t the bottleneck.
The best output of estimation is a question
“Do we need all 3.4 PB?” is worth more than any server sizing.
Common mistakes
- Forgetting peak load, replicas (×3 for redundancy), backups and indexes.
- Mixing bits and bytes (network in Gbit/s, storage in GB).
- False precision: “3,412.7 TB”. Say “~3.4 PB”.
- Not writing down assumptions, so nobody can challenge them.
Check yourself
Roughly how many requests per second is 10 million requests per day?
10⁷ / 86,400 ≈ ~115/s average (or 10⁷ / 10⁵ = 100/s as a quick round). Plan for several hundred per second at peak.
500,000 images per day at 4 MB each. Storage per year?
500 k × 4 MB = 2 TB/day → ~730 TB/year, before replication.
Why estimate peak separately from average?
Systems fail at peak. Capacity, autoscaling limits and queue sizing must handle bursts, not the daily mean.
Apply it
- Get real numbers from the team (images/day, average size, retention) and redo this estimate for the actual platform.
- Compare the storage estimate with the actual Azure bill for storage. If they differ a lot, find out why.
Go deeper
Related: 01 Requirements · 03 Vertical vs Horizontal Scaling · 10 Blob Storage Patterns