Web development

Facebook / Meta Architecture: একটি ছোট PHP Application থেকে বিশ্বের অন্যতম বৃহৎ Distributed System

Facebook Architecture হলো কীভাবে Facebook/Meta বিলিয়ন ব্যবহারকারী, বিপুল ডেটা, ট্রাফিক, social graph, distributed database, caching, networking এবং AI infrastructure পরিচালনা করে—তার সামগ্রিক প্রযুক্তিগত কাঠামো।

Walid15 min read17 views
Facebook / Meta Architecture: একটি ছোট PHP Application থেকে বিশ্বের অন্যতম বৃহৎ Distributed System

আমরা যখন Facebook বা Instagram ব্যবহার করি, তখন আমাদের কাছে পুরো ব্যাপারটা খুব সহজ মনে হয়।

অ্যাপ খুললাম → Login করলাম → News Feed দেখলাম → কোনো Post-এ Like দিলাম → Comment করলাম → Messenger-এ Message পাঠালাম।

কিন্তু এই প্রতিটি ছোট action-এর পেছনে রয়েছে অসংখ্য distributed system, database, cache, networking infrastructure, machine-learning model, storage system এবং data center।

আজকের Meta infrastructure এতটাই বড় যে এটাকে শুধু “একটা backend server + একটা database” হিসেবে চিন্তা করলে পুরো architecture-টা বোঝা সম্ভব নয়।

বরং এটাকে ভাবতে হবে একটি massive distributed ecosystem হিসেবে।

Meta-এর নিজের Engineering Blog-এর infrastructure evolution নিয়ে প্রকাশিত লেখাটি দেখলে সবচেয়ে ভালোভাবে বোঝা যায়, কীভাবে Facebook-এর architecture সময়ের সঙ্গে পরিবর্তিত হয়েছে।

Meta-এর infrastructure evolution নিয়ে বিস্তারিত:

https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/

---

১. Facebook আসলে কীভাবে শুরু হয়েছিল?

আজকের Facebook-এর architecture দেখে মনে হতে পারে শুরু থেকেই হয়তো বিশাল distributed infrastructure ছিল।

বাস্তবে তা নয়।

Facebook শুরু হয়েছিল অনেক বেশি সাধারণ architecture দিয়ে।

মূল technology stack ছিল:

LAMP

যার অর্থ:

  • L — Linux
  • A — Apache
  • M — MySQL
  • P — PHP

অর্থাৎ খুব সরলভাবে architecture-টা এমন ভাবতে পারি:

User
  │
  ▼
Apache Web Server
  │
  ▼
PHP Application
  │
  ▼
MySQL Database

এটা আজকের একটি সাধারণ web application-এর architecture-এর সঙ্গে অনেকটাই মিলে যায়।

ধরো তুমি একটি application বানালে:

React / Next.js
       │
       ▼
Node.js / NestJS
       │
       ▼
PostgreSQL / MongoDB

Facebook-এর শুরুর architecture-এর fundamental idea-ও প্রায় একই ছিল।

তাহলে প্রশ্ন আসে—

এত বিশাল architecture কেন তৈরি করতে হলো?

উত্তর হলো:

Technology-এর complexity প্রথমে আসেনি; scale-এর কারণে complexity এসেছে।

---

২. Scale বাড়ার সঙ্গে প্রথম সমস্যা: Database

ধরো তোমার application-এ মাত্র 100 জন user আছে।

তাহলে:

Backend
   │
   ▼
MySQL

দিয়ে খুব সহজেই কাজ চলে যাবে।

তারপর user হলো:

1,000
10,000
100,000
1 million
10 million
100 million
1 billion+

এখন একই database-এর ওপর enormous amount of traffic আসতে শুরু করবে।

User শুধু login করছে না।

প্রতিটি user:

  • profile দেখছে
  • friend তৈরি করছে
  • post করছে
  • comment করছে
  • like করছে
  • photo upload করছে
  • message পাঠাচ্ছে
  • notification পাচ্ছে
  • অন্য user-এর profile দেখছে
  • বিভিন্ন content interact করছে

অর্থাৎ database-এর ওপর simultaneously অসংখ্য read এবং write তৈরি হচ্ছে।

একটি single database দিয়ে এই scale maintain করা কঠিন হয়ে যায়।

এখান থেকেই distributed architecture-এর প্রয়োজন তৈরি হয়।

---

৩. Facebook-এর University-Based Database Architecture

Facebook-এর early stage-এ database organization-এর একটি interesting approach ছিল।

University অনুযায়ী database আলাদা করা হয়েছিল।

ধরো:

                Web Servers
               /     |      \
              /      |       \
             ▼       ▼        ▼
         Harvard    Yale    Columbia
            DB        DB        DB

এতে load আলাদা করা যায়।

কিন্তু খুব দ্রুত নতুন সমস্যা তৈরি হলো।

ধরো Harvard-এর একজন student Yale-এর একজন student-এর সঙ্গে friend হলো।

এখন relationship কোথায় থাকবে?

Harvard DB
     │
     │
     └──────────► Yale DB

এখানে Facebook বুঝতে শুরু করল যে তারা শুধু user data store করছে না।

তারা আসলে relationships store করছে।

আর এখান থেকেই আসে Facebook-এর সবচেয়ে গুরুত্বপূর্ণ architectural concept:

Social Graph

---

৪. Social Graph কী?

Facebook-এর মূল data model-কে খুব সহজভাবে একটি graph হিসেবে বোঝা যায়।

Graph-এর দুইটি প্রধান অংশ থাকে:

Node

এবং

Edge

Facebook-এর ক্ষেত্রে node হতে পারে:

User
Page
Post
Photo
Video
Group
Comment

আর edge বা relationship হতে পারে:

FRIEND
FOLLOWS
LIKES
COMMENTS
MEMBER_OF
SHARED

উদাহরণ:

             User A
             /    \
            /      \
        friend     likes
          /          \
      User B        Post X
                      │
                      │
                    belongs
                      │
                      ▼
                    Page Y

অর্থাৎ Facebook-এর কাছে প্রশ্ন শুধু:

“User 123 কে?”

এমন নয়।

প্রশ্ন হতে পারে:

“User 123-এর friends কারা?”

অথবা:

“User 123 কোন pages follow করে?”

অথবা:

“User 123 কোন posts-এ interaction করেছে?”

অথবা:

“এই post-এর সঙ্গে কোন users-এর relationship আছে?”

এগুলো graph queries।

এই enormous social graph efficiently manage করার জন্য Meta তৈরি করে TAO

---

৫. TAO — Facebook-এর Social Graph Infrastructure

TAO Facebook architecture-এর অন্যতম গুরুত্বপূর্ণ infrastructure system।

TAO-কে খুব simplified ভাবে একটি graph-oriented data access layer হিসেবে ভাবতে পারো।

Architecture:

Application
    │
    ▼
   TAO
   / \
  /   \
Cache Database

Application সরাসরি প্রতিবার database-এর সঙ্গে কথা না বলে TAO-এর মাধ্যমে graph data access করতে পারে।

উদাহরণ:

"User A-এর friends কারা?"

Request:

Application
     │
     ▼
    TAO
     │
     ▼
Graph Data
     │
     ▼
Friends

TAO-এর বড় architectural idea হলো:

Massive social graph-কে scalable এবং efficient ভাবে access করার জন্য একটি specialized infrastructure layer তৈরি করা।

TAO নিয়ে Meta-এর classic engineering article:

https://engineering.fb.com/2013/06/25/core-infra/tao-the-power-of-the-graph/

---

৬. Database-এর পর আসে Cache

ধরো Facebook-এর কোটি কোটি user একই ধরনের data বারবার request করছে।

যেমন:

User Profile
Friend List
Popular Post
Page Information
Comments

প্রতিটি request যদি database-এ যায়:

100 million requests
        │
        ▼
      MySQL
        │
        💥

database খুব দ্রুত bottleneck হয়ে যাবে।

তাই Facebook massive caching ব্যবহার করে।

একটি simplified architecture:

Request
   │
   ▼
Application
   │
   ▼
 Cache
  /  \
 /    \
HIT   MISS
│       │
▼       ▼
Data  Database
        │
        ▼
      Cache

যদি cache-এ data থাকে:

Request
   ↓
Cache
   ↓
Response

Database পর্যন্ত যেতে হয় না।

এটা অনেক দ্রুত।

---

৭. Memcache এবং Distributed Caching

Meta-এর infrastructure evolution-এর মধ্যে Memcache গুরুত্বপূর্ণ ভূমিকা পালন করেছে।

কিন্তু distributed caching-এর সঙ্গে নতুন সমস্যা আসে।

ধরো database-এ user-এর নতুন profile picture save হয়েছে:

Database:
new-photo.jpg

কিন্তু cache-এ এখনও আছে:

old-photo.jpg

তাহলে:

Database = NEW
Cache    = OLD

এখন user পুরনো data দেখতে পারে।

এখানে আসে:

  • cache invalidation
  • consistency
  • replication
  • synchronization
  • stale data

এগুলো distributed systems-এর সবচেয়ে কঠিন সমস্যাগুলোর মধ্যে পড়ে।

---

৮. Facebook একসময় Global হয়ে গেল

Facebook যখন শুধু একটি region-এর application ছিল, তখন architecture তুলনামূলকভাবে সহজ ছিল।

কিন্তু global users আসার পর architecture এমন হলো:

                   Internet
                      │
        ┌─────────────┼─────────────┐
        ▼             ▼             ▼
       USA          Europe         Asia
        │             │             │
       DC            DC            DC

এখন প্রশ্ন:

User কোন data center-এ যাবে?

আরেকটি প্রশ্ন:

এক data center-এর data অন্য data center-এ কীভাবে যাবে?

আরও একটি:

User যদি Asia-তে থাকে, তাহলে কেন তাকে সবসময় US data center-এ যেতে হবে?

এখান থেকে আসে:

  • Edge infrastructure
  • POP
  • Global networking
  • Data center networking
  • Backbone network
  • Traffic routing

---

৯. POP এবং Edge Infrastructure

POP বা Point of Presence হলো network-এর এমন একটি location যা users-এর কাছাকাছি traffic handle করতে সাহায্য করে।

Conceptually:

User
 │
 ▼
Nearest Edge / POP
 │
 ▼
Meta Backbone
 │
 ▼
Data Center
 │
 ▼
Application Services

এতে user-এর request সরাসরি অনেক দূরের data center-এ যাওয়ার প্রয়োজন কমে।

Latency কমানো এবং enormous traffic efficiently handle করার জন্য global edge infrastructure অত্যন্ত গুরুত্বপূর্ণ।

---

১০. Facebook-এর Data Center Architecture

একসময় Facebook-এর server ছিল।

তারপর server-এর সংখ্যা হলো হাজার।

তারপর লাখ।

এখন architecture শুধু application architecture নয়।

এখন physical infrastructure-ও architecture-এর অংশ।

Application
     ↓
Servers
     ↓
Racks
     ↓
Data Center
     ↓
Power
     ↓
Cooling
     ↓
Network

অর্থাৎ Meta-scale infrastructure-এ software engineer-এর architecture চিন্তা করতে গেলে physical infrastructure-ও বুঝতে হয়।

কারণ millions of machines চালাতে হলে প্রয়োজন:

  • Electricity
  • Cooling
  • Networking
  • Hardware
  • Storage
  • Compute
  • Monitoring
  • Failure recovery

---

১১. Sharding — Data কীভাবে ভাগ করা হয়?

ধরো Facebook-এর database-এ billions of users আছে।

সব data এক জায়গায় রাখা practical নয়।

তাই data partition করা হয়।

এটাকে বলা হয়:

Sharding

উদাহরণ:

                 User Data
                    │
                    ▼
               Sharding
                    │
       ┌────────────┼────────────┐
       ▼            ▼            ▼
    Shard A      Shard B      Shard C

ধরো:

Shard A
User ID 0 - 999M

Shard B
User ID 1B - 1.999B

Shard C
User ID 2B - 2.999B

বাস্তবে অবশ্যই partitioning strategy আরও sophisticated।

Meta-এর infrastructure article অনুযায়ী, তাদের Shard Manager tens of millions of shards এবং hundreds of thousands of servers-এর scale manage করে।

এখানে একটা গুরুত্বপূর্ণ lesson:

Facebook-এর scale-এ database নিজেই একটি distributed system।

---

১২. Distributed Storage

Database-এর বাইরেও Meta-কে enormous amount of data store করতে হয়।

যেমন:

Photos
Videos
Messages
Logs
AI datasets
User data
Machine learning data

তাই distributed storage systems প্রয়োজন হয়।

Meta-এর infrastructure evolution-এর মধ্যে Tectonic গুরুত্বপূর্ণ distributed storage technology।

Conceptually:

Application
     │
     ▼
Storage Layer
     │
     ▼
Distributed Storage
     │
 ┌───┼────┬────┐
 ▼   ▼    ▼    ▼
Disk Disk Disk Disk

একটি file বা object-এর data এক জায়গায় না থেকে distributed infrastructure-এ রাখা হতে পারে।

এতে:

  • scalability
  • durability
  • fault tolerance
  • availability

বাড়ানো যায়।

---

১৩. Failure is Normal

একটি traditional application-এ developer ভাবে:

“Server যেন না নষ্ট হয়।”

Meta-scale infrastructure-এ চিন্তাটা ভিন্ন:

“Server নষ্ট হবেই। System কীভাবে চলতে থাকবে?”

এটাই distributed system-এর অন্যতম গুরুত্বপূর্ণ philosophy।

ধরো:

Server 1 ✓
Server 2 ✓
Server 3 ❌
Server 4 ✓
Server 5 ✓

Server 3 মারা গেলেও পুরো service বন্ধ হওয়া উচিত নয়।

Architecture:

                Request
                   │
                   ▼
              Load Balancer
              /    |    \
             /     |     \
            ▼      ▼      ▼
        Server1 Server2 Server3
           ✓       ✓       ❌

Traffic অন্য server-এ চলে যাবে।

এটাই fault tolerance।

---

১৪. Service Routing

Meta-scale architecture-এ হাজার হাজার বা তারও বেশি services থাকতে পারে।

ধরো:

Feed Service
User Service
Comment Service
Media Service
Message Service
Notification Service
Ads Service
Ranking Service
Payment Service

একটি service অন্য service-এর সঙ্গে কথা বলবে।

তখন প্রশ্ন:

কোন server-এ request পাঠাব?
কোন instance healthy?
কোন region ব্যবহার করব?
কোন route fastest?

এখানে আসে service routing এবং service infrastructure।

Meta-এর বর্তমান infrastructure-এ Service Router-এর মতো systems global service traffic পরিচালনায় গুরুত্বপূর্ণ ভূমিকা রাখে।

Conceptually:

Service A
   │
   ▼
Service Router
   │
 ┌─┼───────┐
 ▼ ▼       ▼
S1 S2      S3

---

১৫. News Feed আসলে Database Query নয়

এখানে Facebook architecture আরও interesting হয়ে যায়।

তুমি Facebook খুললে News Feed দেখো।

কিন্তু Facebook-এর কাজ:

"সব posts দেখাও"

এটা নয়।

বরং:

“এই particular user-এর জন্য এই মুহূর্তে কোন content সবচেয়ে relevant?”

এটা একটি ranking problem

Simplified architecture:

                 User
                  │
                  ▼
          Candidate Content
                  │
                  ▼
              Ranking
                  │
                  ▼
          Ranked Content
                  │
                  ▼
              News Feed

---

১৬. Ranking কীভাবে কাজ করতে পারে?

ধরো user-এর কাছে 10,000 সম্ভাব্য content আছে।

Facebook সব 10,000 দেখাবে না।

প্রথমে candidate generate করতে হবে।

তারপর ranking করতে হবে।

10,000 candidates
       │
       ▼
Candidate Selection
       │
       ▼
1,000 candidates
       │
       ▼
Ranking Model
       │
       ▼
100 candidates
       │
       ▼
Final Ranking
       │
       ▼
User Feed

এখানে machine learning গুরুত্বপূর্ণ হয়ে যায়।

---

১৭. Facebook থেকে Recommendation Engine

পুরোনো social network model-এ content-এর বড় অংশ আসত:

My Friends
   ↓
Their Posts
   ↓
Ranking
   ↓
Feed

কিন্তু modern recommendation systems-এর ক্ষেত্রে content universe অনেক বড়।

ধরো billions of videos।

তখন:

                Billions of Content
                         │
                         ▼
                 Candidate Retrieval
                         │
                         ▼
                    Ranking
                         │
                         ▼
                       User

এখানে AI/ML ছাড়া এই scale-এ personalized recommendation করা খুব কঠিন।

---

১৮. User Behavior-ও Data

ধরো একজন user:

Video A → দেখল
Video B → Skip করল
Video C → আবার দেখল
Video D → Like করল
Video E → Share করল

এই interactions থেকে system user-এর preference সম্পর্কে ধারণা তৈরি করতে পারে।

Conceptually:

User Actions
    │
    ├── View
    ├── Like
    ├── Click
    ├── Share
    ├── Skip
    └── Purchase
          │
          ▼
      User Signals
          │
          ▼
     ML Representation

এখান থেকেই modern recommendation architecture আরও sophisticated হয়।

---

১৯. Embeddings এবং Recommendation

AI systems-এ আমরা embeddings সম্পর্কে শিখি।

একটি content-কে vector representation দেওয়া যায়:

Video
  │
  ▼
ML Model
  │
  ▼
Embedding
  │
  ▼
[0.12, -0.42, 0.81, ...]

একইভাবে user-এর preference-ও একটি representation হিসেবে model করতে পারি।

তারপর:

User Representation
        +
Content Representation
        +
Context
        │
        ▼
Recommendation Model
        │
        ▼
Score

যে content-এর score বেশি, সেটি user-এর কাছে বেশি relevant হতে পারে।

এটা অবশ্যই modern Meta recommendation system-এর সম্পূর্ণ implementation নয়; এটি বোঝার জন্য simplified mental model।

---

২০. Ads Ranking আরও বড় ML Problem

Meta-এর 5 August 2026-এর Ads Ranking article এই বিষয়টিকে আরও গভীরভাবে দেখায়।

Meta-এর recommendation systems-কে enormous amount of user interactions এবং temporal signals process করতে হয়।

অর্থাৎ user শুধু কী করেছে তা নয়—

কখন করেছে, কী sequence-এ করেছে, কীভাবে করেছে—এসব signal-ও গুরুত্বপূর্ণ হতে পারে।

Simplified architecture:

User Interactions
       │
       ├── Click
       ├── View
       ├── Like
       ├── Purchase
       ├── Skip
       └── Timing
             │
             ▼
       User Sequence
             │
             ▼
      ML Representation
             │
             ▼
      Candidate Ranking
             │
             ▼
        Final Ranking
             │
             ▼
          Ad/User

Meta-এর article:

https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking/

এখানে সবচেয়ে গুরুত্বপূর্ণ বিষয় হলো:

Production AI system = শুধু একটি model নয়।

এর চারপাশে থাকে:

Data
 ↓
Feature / Representation
 ↓
Candidate Generation
 ↓
Model
 ↓
Ranking
 ↓
Serving
 ↓
Monitoring
 ↓
Feedback
 ↓
New Data

অর্থাৎ AI একটি সম্পূর্ণ production pipeline।

---

২১. AI আসার পর Architecture আবার বদলে গেল

আগের Facebook architecture অনেকটা ছিল:

Users
  ↓
Web Servers
  ↓
Services
  ↓
Database

Modern Meta architecture-এ AI infrastructure একটি বিশাল additional layer তৈরি করেছে।

Users
  ↓
Applications
  ↓
Services
  ↓
Recommendation / AI
  ↓
GPU Clusters
  ↓
Distributed Storage
  ↓
High-Speed Network
  ↓
Data Centers

এখানে হাজার হাজার GPU একসঙ্গে কাজ করতে পারে।

---

২২. GPU Cluster কেন দরকার?

Machine learning model-এর অনেক computation parallel ভাবে করা যায়।

তাই GPU এই workload-এর জন্য অত্যন্ত কার্যকর।

Simplified:

                ML Training
                     │
                     ▼
              GPU Cluster
       ┌─────────────┼─────────────┐
       ▼             ▼             ▼
     GPU 1         GPU 2         GPU 3
       │             │             │
       └─────────────┼─────────────┘
                     ▼
                 Model

কিন্তু এখানে নতুন সমস্যা:

যদি হাজার হাজার GPU একসঙ্গে একটি training workload-এ অংশ নেয়, তাহলে network, storage এবং failure handling অত্যন্ত গুরুত্বপূর্ণ হয়ে যায়।

একটি GPU failure-ও training job-কে প্রভাবিত করতে পারে।

তাই AI infrastructure নিজেই একটি distributed systems problem।

---

২৩. AI Infrastructure-এর নতুন Problems

Traditional web application:

Request
  ↓
Server
  ↓
Response

AI training:

Training Job
     │
     ├── GPU 1
     ├── GPU 2
     ├── GPU 3
     ├── GPU 4
     ├── ...
     └── GPU 1000+

এখন দরকার:

  • High-speed networking
  • Distributed storage
  • GPU scheduling
  • Failure recovery
  • Checkpointing
  • Monitoring
  • Resource management
  • Power
  • Cooling

অর্থাৎ AI infrastructure শুধু software problem নয়।

এটা hardware + software + network + storage + data center—সবকিছুর problem।

---

২৪. Meta-এর Infrastructure আসলে একটি Stack

এখন পুরো architecture-টাকে একবার একসঙ্গে দেখি।

┌──────────────────────────────────────┐
│          Facebook / Instagram        │
│          WhatsApp / Threads          │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│          Application Services        │
│ Feed • Graph • Messaging • Ads       │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│       Service Infrastructure         │
│ Routing • Load Balancing • Discovery │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│          Data Infrastructure         │
│ TAO • Cache • KV • Sharding          │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│              Storage                 │
│ MySQL • Distributed Storage          │
│ Tectonic • Object Storage            │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│              Network                 │
│ Edge • POP • Backbone • Data Center  │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│              Compute                 │
│ CPU • GPU • AI Accelerators          │
└──────────────────┬───────────────────┘
                   │
                   ▼
┌──────────────────────────────────────┐
│        Physical Infrastructure       │
│ Servers • Racks • Power • Cooling    │
└──────────────────────────────────────┘

এবং AI/ML প্রায় প্রতিটি layer-এর সঙ্গে interact করে।

---

২৫. Meta Architecture-এর সবচেয়ে বড় Lesson

Facebook-এর architecture দেখে সবচেয়ে গুরুত্বপূর্ণ যে lesson পাওয়া যায় সেটা হলো:

Scale creates architecture.

প্রথমে ছিল:

LAMP

তারপর:

More Users
   ↓
Database Scaling

তারপর:

More Reads
   ↓
Caching

তারপর:

More Relationships
   ↓
Social Graph
   ↓
TAO

তারপর:

More Data
   ↓
Sharding
   ↓
Distributed Storage

তারপর:

More Users Globally
   ↓
Edge + POP + Backbone

তারপর:

More Machines
   ↓
Fleet Management

তারপর:

More Failures
   ↓
Fault Tolerance

তারপর:

More Content
   ↓
Recommendation Systems

তারপর:

More Personalization
   ↓
Machine Learning

তারপর:

More AI
   ↓
GPU Clusters
   ↓
AI Infrastructure

এটাই Meta infrastructure evolution-এর মূল গল্প।

---

২৬. তোমার নিজের Backend-এর সঙ্গে এটি কীভাবে relate করে?

এখানেই বিষয়টি একজন Full Stack Engineer-এর জন্য সবচেয়ে interesting।

তুমি যদি একটি ছোট application বানাও:

Next.js
   ↓
NestJS
   ↓
PostgreSQL
   ↓
Redis

এটি fundamentally ভুল architecture নয়।

বরং Facebook-ও একসময় অনেক simpler architecture দিয়ে শুরু করেছিল।

Difference হলো scale।

যখন তোমার application-এ:

100 users

তখন:

1 Backend
1 Database

হতে পারে যথেষ্ট।

যখন:

100,000 users

তখন:

Load Balancer
Multiple Backend
Redis
Database Optimization

লাগতে পারে।

যখন:

Millions of users

তখন:

Multiple Services
Sharding
Distributed Cache
Message Queue
Object Storage
CDN
Observability

প্রয়োজন হতে পারে।

আর Meta-scale-এ:

Global Networking
Massive Sharding
Distributed Storage
Service Routing
Fault Tolerance
AI Infrastructure
GPU Clusters
Data Centers

এসব প্রয়োজন হয়।

অর্থাৎ architecture আগে থেকে unnecessarily complex করার দরকার নেই।

Scale বাড়ার সঙ্গে architecture evolve করবে।

---

২৭. পুরো Meta Architecture এক লাইনে

যদি পুরো বিষয়টা একটি sentence-এ বলতে হয়:

Meta-এর architecture হলো এমন একটি globally distributed infrastructure যেখানে massive user traffic, social relationships, content, messaging, storage এবং machine-learning workloads-কে হাজার হাজার distributed services, caches, databases, storage systems, networks, compute clusters এবং data centers-এর মাধ্যমে reliably এবং efficiently operate করা হয়।

আর সবচেয়ে সুন্দর বিষয় হলো—

এটা একদিনে তৈরি হয়নি।

এটা তৈরি হয়েছে:

Problem
  ↓
Scale
  ↓
New Bottleneck
  ↓
New Architecture
  ↓
New Infrastructure
  ↓
More Scale
  ↓
New Bottleneck
  ↓
New Architecture

এই cycle-এর মাধ্যমে।

---

২৮. শেষ কথা

Facebook-এর architecture শেখার সময় সবচেয়ে বড় ভুল হবে শুধু technology-এর নাম মুখস্থ করা:

TAO কী?
Tectonic কী?
Memcache কী?
ZippyDB কী?
Twine কী?

এগুলো জানা অবশ্যই দরকার।

কিন্তু তার চেয়েও গুরুত্বপূর্ণ হলো কেন এগুলো তৈরি করতে হয়েছিল সেটা বোঝা।

একজন strong backend/system engineer চিন্তা করবে:

Traffic বাড়লে কী হবে?

Database bottleneck হলে?

Cache stale হলে?

Server মারা গেলে?

এক region down হলে?

Data অন্য region-এ কীভাবে যাবে?

Billions of records কীভাবে shard হবে?

Billions of content কীভাবে rank হবে?

Thousands of GPUs কীভাবে coordinate করবে?

এই প্রশ্নগুলোর উত্তর খুঁজতে খুঁজতেই তুমি আসলে Meta Architecture বুঝতে শুরু করবে।

এবং তখন Facebook আর শুধু একটি social media website মনে হবে না।

তুমি এটাকে দেখতে শুরু করবে একটি বিশাল:

Distributed System + Data Platform + Networking Infrastructure + AI Platform + Global Computing System

হিসেবে।

---

Recommended Reading

Meta Engineering — Infrastructure Evolution and the Advent of AI

https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/

TAO — The Power of the Graph

https://engineering.fb.com/2013/06/25/core-infra/tao-the-power-of-the-graph/

From User Sequences to Scaling Laws — Meta's Ads Ranking Architecture

https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking/

এই তিনটি resource দিয়ে শুরু করলে Meta-এর architecture বোঝার জন্য একটি শক্ত foundation তৈরি হবে।