Section 3: A Million Tablets on a Napkin - Scale Estimation & Workload Analysis
Executive Summary: Webtable (400 TB) and Profiles (100 TB) make ~500 TB compressed, 1.5 PB on disk after GFS's three replicas. At a 0.5 GB average tablet the cell holds ~1 million tablets, ~1,000 per server, and the whole METADATA table is just 1 GB. Peak load is ~2 M ops/s. At 5,000 ops/s per server, an evenly spread peak needs only 400 servers; the other 600 are 2.5× headroom for range-partition skew and a lost rack. Bytes live in GFS, so skew, CPU and memory size this fleet, not data volume.
The interviewer taps Section 2's requirements. "Put numbers on it. And when you get to a server count, tell me why it's that number."
The Cell We Agreed On
| Ref | Assumption |
|---|---|
| Cell | One datacenter: 1,000 tablet servers in 25 racks × 40, sharing machines with the GFS chunkservers; 1 master; a 5-replica Chubby cell (12 s session lease, 45 s grace) |
| Server | 32 cores, 128 GB RAM, 10 Gbps NIC |
| Latency | In-datacenter RPC ≈ 0.5 ms; a random GFS HDD read ≈ 10 ms (one seek) |
You: "Remember that last line: one seek spends Section 2's whole 10 ms p99 read budget. Section 9 is built around that fact."
Sign in to continue reading
The rest of this lesson is available with a free account. Signing in with Google or Microsoft is free.
Sign in to read the full lesson