Back-of-the-Envelope Estimation Techniques

Back-of-the-envelope estimations are not intended to yield exact accounting precision down to decimal places. Rather, their purpose is to determine the order of magnitude (O(10^x)) of a system's physical constraints:

• Are we designing for 100 QPS (single node) or 100,000 QPS (distributed cluster)?
• Is 5-year storage 50 GB (single PostgreSQL database) or 50 PB (sharded S3 data lake)?
• Is egress bandwidth 10 Mbps (commodity network card) or 100 Gbps (dedicated CDN edge routing)?

By rapidly bounding the problem space, you immediately discard non-viable architectural designs (e.g., attempting in-memory joins on 50 TB of data) and justify distributed components (caching, sharding, message queues) before drawing boxes.

The 100,000 Seconds per Day Rule

A day has 24 × 60 × 60 = 86,400 seconds. In mental calculations, round this up to 100,000 (10^5) seconds:

Average QPS = \frac{Total Daily Requests}{100,000}

Peak QPS = Average QPS × 2 (Standard Enterprise) \quad or \quad Average QPS × 3 to 5 (Flash Sales / Consumer)

Instant QPS Benchmarks:

• 1 Million requests/day = \frac{10^6}{10^5} = 10 QPS (Peak ≈ 20-30 QPS)
• 100 Million requests/day = \frac{10^8}{10^5} = 1,000 QPS (Peak ≈ 2,000-3,000 QPS)
• 1 Billion requests/day = \frac{10^9}{10^5} = 10,000 QPS (Peak ≈ 20,000-30,000 QPS)

---

Powers of 2 vs Powers of 10 Approximation

Understanding the physics of latency across the computer storage and network hierarchy dictates why caching, indexing, and locality exist:

Architectural Insights from Latency Physics:

• Memory vs SSD: DRAM is 200-500× faster than NVMe SSDs for random reads.
• Disk Seeks: Random disk seeks on spinning magnetic platters are 100,000× slower than RAM, explaining why LSM-Trees use sequential log writes.
• Speed of Light: Network roundtrips across continents (100+ ms) dwarf internal server processing times (< 5 ms), mandating regional edge PoPs and CDNs.

When estimating hardware infrastructure from raw data math, real-world systems incur structural overheads that must be factored in:

1. Replication Factor Multiplier (3×): Enterprise distributed stores (HDFS, Kafka, Cassandra, Ceph) maintain 3 replicas across different availability zones or failure domains. Always multiply persistent raw storage by 3×.
2. Database Index & Metadata Overhead (+30\% - 50\%): B-Tree indexes, primary key lookups, and table metadata (PostgreSQL MVCC vacuum headers) add 30\% - 50\% storage on top of raw payload bytes.
3. Network Bit/Byte Conversion Trap: Network capacity is measured in Bits per second (bps, Gbps), while storage/memory is measured in Bytes per second (B/s, GB/s). Always multiply Bytes by 8 to obtain network line rates:

Bandwidth (Gbps) = Throughput (GB/s) × 8

1. Headroom Buffer: Production systems should run at ≤ 65-70\% CPU and disk capacity to absorb sudden traffic surges, garbage collection pauses, and failover re-balancing.