📋 Technical Program Manager (TPM) & System Design Learning Path¶

🚀 Hyperscale Architecture & Technical Leadership: Master 5-Stage Clos fabric scaling math, eBGP vs. iBGP architectural trade-offs, blast radius containment, SLA / convergence budget calculations, and multi-vendor RFP execution.

📊 Learning Path Overview¶

Metric	Target Specification
Estimated Completion Time	20 – 25 Hours (Architectural analysis, system design drills, and case studies)
Milestone Stages	5 Progressive Stages (Clos Scaling Math → Routing Architecture → SLA & Convergence Budgets → Overlay Virtualization → Vendor RFPs & Telemetry)
Target Roles	Technical Program Manager (TPM) - Infrastructure, Network Solutions Architect, Infrastructure Program Lead, Engineering Director
Target Employers	Google, Meta, Apple, AWS, Microsoft, ByteDance, NVIDIA, Global Financial Institutions, and Cloud Infrastructure Consultancies

🧠 Core TPM Engineering & Leadership Pillars¶

Program Domain	Key Architectural Decisions	Leadership & Program Function
Fabric Scaling Math	5-Stage Clos vs. 2-Tier Spine-Leaf	ASIC port-density calculations, oversubscription ratios, and modular pod expansion planning
Routing Protocol Selection	eBGP (RFC 7938) vs. iBGP + RRs	Blast-radius containment, routing loop elimination, and global ASN allocation governance
Convergence SLAs	Sub-50ms Ti-LFA vs. IGP Timers	Establishing production availability budgets (99.999%), link failover SLAs, and error budgets
Overlay Virtualization	EVPN-VXLAN ESI vs. Proprietary MLAG	Decoupling physical underlay from tenant overlays and eliminating vendor hardware lock-in
Telemetry Standardization	gNMI Streaming vs. SNMP Polling	Standardizing multi-vendor OpenConfig schemas across Arista, Cisco, and whitebox switches

🗺️ 5-Stage Progressive Milestone Roadmap¶

01

Stage 1 · 5-Stage Clos Fabric Sizing & Capacity Math

Fabric Math

Calculate non-blocking Clos port radix formulas and oversubscription ratios across Tier-1 Leaf, Tier-2 Spine, and Tier-3 Super-Spine switches.

5-Stage Clos Port Radix Formulas Oversubscription Ratios

02

Stage 2 · Routing Architecture & Blast Radius Design

Governance

Compare RFC 7938 eBGP leaf-spine fabrics against IGP+iBGP to eliminate routing loops and contain failure blast radiuses.

RFC 7938 eBGP ASN Allocation Strategy Failure Domain Isolation

03

Stage 3 · High-Availability & Convergence SLA Budgets

99.999% SLAs

Calculate production availability budgets and evaluate sub-50ms Ti-LFA fast reroute vs. traditional IGP convergence timers.

Ti-LFA Sub-50ms Error Budgeting BFD Hardware Offload

04

Stage 4 · Multi-Tenant Overlays & Open Standards

Open Standards

Decouple physical fabrics from tenant services using EVPN-VXLAN with ESI multihoming, avoiding proprietary MLAG/vPC lock-in.

EVPN-VXLAN ESI Multihoming Multi-Vendor Interop

05

Stage 5 · Telemetry Governance & Vendor RFP Delivery

Vendor RFPs

Write standardized multi-vendor OpenConfig telemetry specs and mandate automated PyATS acceptance tests for hardware procurement.

OpenConfig YANG gNMI Governance Automated RFP Acceptance

🚀 Interactive Lesson Directory (Click Any Lesson to Start)¶

Milestone Stage	Architecture & Program Focus	Clickable Lessons & System Designs	Technical Focus	Action
Stage 1 `Clos Sizing Math`	ASIC Port Radix Math, Oversubscription Ratios, Super-Spine Pods	• EVPN-VXLAN Clos Fabric Design • Hyperscale System Design Masterclass	Pod Math	Start Stage 1 →
Stage 2 `BGP Governance`	RFC 7938 eBGP Clos Design, Blast Radius, ASN Allocation Policies	• Phase 1 · BGP Fundamentals & Architecture • Phase 1 · Lab 03: Route Reflector Hierarchy	ASN Model	Start Stage 2 →
Stage 3 `SLA & FRR Budgets`	Sub-50ms Ti-LFA Fast Reroute, BFD Hardware Offload, Error Budgets	• Segment Routing (SR-MPLS) Ti-LFA Architecture • WAN Edge BFD Sub-Second Failover	99.999% SLAs	Start Stage 3 →
Stage 4 `Multi-Tenant Overlays`	RFC 8365/7432 EVPN-VXLAN, ESI All-Active Multihoming vs MLAG Lock-in	• EVPN Lab 03: ESI All-Active Multihoming • EVPN Lab 05: Multi-Site DCI Architecture	Open Standards	Start Stage 4 →
Stage 5 `Vendor RFP Delivery`	OpenConfig YANG Governance, gNMI Telemetry SLAs, PyATS Pre-Checks	• Phase 7 · Streaming Telemetry & Observability • Phase 5 · Automated Network Testing with PyATS	RFP Criteria	Start Stage 5 →

🧪 Detailed Milestone Curricula¶

📍 Stage 1: 5-Stage Clos Fabric Sizing & Capacity Math¶

Core Focus: Translating business compute requirements into non-blocking or strictly-bounded oversubscribed network topologies.
Formulas & Trade-offs:
- Maximum leaves supported by $S$ spines: $N_{leaf} = S \times \text{uplinks per leaf}$ .
- Non-blocking server capacity given $k$ -port switches: $N_{servers} = \frac{k^2}{2}$ .
- Calculating bandwidth oversubscription ratios ( $1:1$ , $2:1$ , $3:1$ ) across Tier-1 (Leaf), Tier-2 (Spine), and Tier-3 (Super-Spine) layers.

📍 Stage 2: Routing Architecture & Blast Radius Design¶

Core Focus: Selecting control plane protocols that prevent cascade failures across multi-pod datacenters.
Key Comparisons:
- RFC 7938 eBGP: Distinct ASN per tier/rack; AS-PATH automatically eliminates loops; routing policy applied at AS boundaries.
- iBGP + Route Reflectors: Requires full IGP underlay; risk of route oscillations without careful cluster-id hierarchy.

📍 Stage 3: High-Availability & Convergence SLA Budgets¶

Core Focus: Calculating downtime risk and defining strict architectural failover parameters for contractual SLAs.
Key Concepts:
- Sub-50ms Topology-Independent Loop-Free Alternate (Ti-LFA).
- BFD timer aggregation vs. ASIC CPU load.
- Establishing 99.99% ("four nines") vs. 99.999% ("five nines") quarterly error budgets.

📍 Stage 4: Multi-Tenant Overlays & Open Standards¶

Core Focus: Avoiding vendor lock-in when evaluating multi-million dollar switch hardware procurement.
Key Concepts:
- Why proprietary multi-chassis link aggregation (Cisco vPC, Arista MLAG) creates fragile dual-control-plane bugs.
- How RFC 8365/7432 EVPN-VXLAN with ESI (Ethernet Segment Identifier) provides multi-vendor, all-active multihoming.

📍 Stage 5: Telemetry Governance & Vendor RFP Delivery¶

Core Focus: Writing technical specifications and automated validation gates for hardware RFPs and carrier acceptance.
Key Concepts:
- Mandating OpenConfig YANG schemas and gNMI streaming telemetry support in vendor contracts.
- Requiring automated PyATS / Batfish pre-deployment gate checks before accepting network hardware delivery.

📁 System Design Drills & Reference¶

Google & Hyperscale System Design Masterclass: Scenario-based interview drills and architectural trade-offs.
Curriculum Architecture Roadmap: Complete protocol dependency matrix across all 10 network phases.
EVPN-VXLAN Clos Fabric Course: The underlying leaf-spine fabric implementation.
Streaming Telemetry Course: The gNMI/OpenConfig telemetry pipeline.

🎯 System Design Interview Drills for TPMs¶

❓ Question 1: How do you design a non-blocking fabric for 16,384 GPU servers using 64-port 800G switches?¶

Answer: A 2-tier spine-leaf fabric using 64-port switches can connect at most $32 \times 32 = 1,024$ nodes non-blocking (each leaf has 32 downlinks to GPUs and 32 uplinks to spines). To scale to 16,384 GPUs, a 3-tier (5-stage Clos) architecture is required: 1. Tier-1 (Leaf / ToR): 512 leaf switches, each with 32 downlinks to GPUs (total 16,384 GPUs) and 32 uplinks to Tier-2 spines. 2. Tier-2 (Spine / Fabric): Grouped into Pods. Each pod contains leaves connected to 32 spines. 3. Tier-3 (Super-Spine / Core): Aggregates inter-pod traffic across the entire cluster, maintaining 1:1 non-blocking bisectional bandwidth.

❓ Question 2: Why do hyperscalers prefer eBGP (RFC 7938) over IGP+iBGP for datacenter fabric control planes?¶

Answer: 1. Loop Prevention: BGP uses AS-PATH as a built-in loop prevention mechanism. If an update circles back to an AS, it is instantly discarded without complex graph computations. 2. Blast Radius Containment: Route flapping in one rack or pod can be damped and filtered at AS boundaries, preventing full-fabric SPF recalculation storms that occur in link-state IGPs (OSPF/IS-IS). 3. Multi-Vendor Uniformity: BGP is universally supported with identical operational semantics across Arista, Cisco, Juniper, and whitebox SONiC platforms.