Case Study: Design YouTube — Video Streaming Platform

“YouTube giống một đài truyền hình quốc gia — nhưng thay vì một kênh chiếu cho tất cả, mỗi người xem một kênh khác nhau, cùng lúc hàng triệu người. Và ‘đài truyền hình’ đó phải phát mượt mà trên mọi thiết bị, mọi tốc độ mạng, mọi quốc gia.”

Tags: system-design case-study video-streaming youtube cdn transcoding alex-xu Student: Hieu Prerequisite: Tuan-02-Back-of-the-envelope · Tuan-03-Networking-DNS-CDN · Tuan-08-Message-Queue Lien quan: Tuan-06-Cache-Strategy · Tuan-07-Database-Sharding-Replication · Tuan-13-Monitoring-Observability · Tuan-15-Data-Security-Encryption Reference: Alex Xu, System Design Interview Vol 1 — Chapter 14: Design YouTube

0. Context & Why — Tai sao Video Streaming kho?

0.1 Analogie: Dai truyen hinh quoc gia thoi dai so

Hieu, hay tuong tuong mot dai truyen hinh quoc gia. Truoc day, dai phat mot kenh duy nhat — tat ca moi nguoi cung xem mot chuong trinh. Don gian.

Bay gio tuong tuong moi nguoi xem mot kenh khac nhau, cung luc. Nguoi xem phim hanh dong 4K, nguoi xem clip hai 360p tren dien thoai 3G, nguoi tai video len, nguoi dang stream live. Tat ca dien ra dong thoi voi hang trieu nguoi.

Do chinh la YouTube. Va do la ly do no la mot trong nhung bai toan system design kho nhat.

0.2 Nhung thach thuc cot loi

Thach thuc	Giai thich	Vi sao kho?
Data volume	Video la data type nang nhat — 1 phut video 1080p ~ 150MB	Khong gi trong web lon bang video
Bandwidth	Streaming can bang thong lien tuc, khong phai burst nhu text/image	Moi user tieu thu 2-5 Mbps lien tuc
Diversity	Muon van thiet bi: TV 4K, dien thoai cu, tablet, laptop	Phai transcode ra nhieu resolution + codec
Global	User o moi noi tren the gioi	CDN phai phu khap, latency phai thap
Cost	CDN bandwidth la chi phi #1 — len toi hang ty USD/nam	Toi uu chi phi CDN la song con
Upload vs Watch	Upload it nhung nang (processing), Watch nhieu nhung nhe (serving)	Hai pipeline hoan toan khac nhau

0.3 YouTube theo so lieu thuc te (tham khao)

Metric	Con so	Y nghia
DAU	~2 ty	Gan 1/4 dan so the gioi
Video duoc upload moi phut	500 gio	Khong bao gio xem het
Tong video da upload	> 800 trieu	Petabyte-scale storage
Avg watch time/user/day	~30 phut	Engagement cuc cao
So resolution can transcode	5-8	144p, 240p, 360p, 480p, 720p, 1080p, 1440p, 4K
CDN cost/nam (uoc tinh)	$1-5 ty	Chi phi lon nhat cua YouTube

Aha Moment: Video streaming khong phai la “gui file cho user.” No la chia file thanh hang ngan manh nho (segments), chon dung resolution cho tung thoi diem, va gui tung manh mot — giong nhu dut com cho em be, tung muong mot, dung toc do em be nuot duoc.

1. Step 1 — Understand the Problem & Establish Design Scope

1.1 Clarifying Questions (Cau hoi lam ro)

Trong interview, luon hoi truoc khi thiet ke. Duoi day la cac cau hoi quan trong:

Cau hoi	Tra loi	Ghi chu
Nhung feature nao can thiet ke?	Upload video + Stream video	Core features
Clients nao?	Mobile app, web browser, smart TV	Multi-platform
DAU bao nhieu?	5 trieu	Quy mo trung binh-lon
Thoi gian xem trung binh/ngay?	30 phut	Engagement metric
Can ho tro nhieu resolution?	Co — tu 240p den 4K	Adaptive bitrate
Can ho tro nhieu ngon ngu?	Co — subtitle, audio tracks	i18n
Video dai toi da bao lau?	12 gio (nhu livestream rerun)	Anh huong storage
Can DRM (bao ve ban quyen)?	Co	Dac biet cho premium content
Upload file toi da bao lon?	256 GB	Cho video dai, 4K
Co phan biet user mien phi va tra tien?	Co — priority transcoding	Business logic

1.2 Functional Requirements (Yeu cau chuc nang)

FR1: Upload video — user upload video tu client, he thong xu ly va luu tru
FR2: Stream video — user xem video muot ma, khong giat lag
FR3: Multi-resolution — tu dong chuyen doi video thanh nhieu resolution (240p → 4K)
FR4: Adaptive bitrate — tu dong dieu chinh chat luong theo toc do mang cua user
FR5: Search — tim kiem video theo title, description, tags
FR6: Thumbnail generation — tu dong tao thumbnail cho moi video
FR7: Video metadata — title, description, view count, like/dislike, comments
FR8: Resumable upload — neu upload bi gian doan, tiep tuc tu cho dang do

1.3 Non-functional Requirements (Yeu cau phi chuc nang)

Requirement	Target	Giai thich
Availability	99.99%	Video platform la entertainment — downtime = mat user
Latency (playback start)	< 2 giay	User nhan play → video phai bat dau trong 2s
Buffering ratio	< 1%	Ty le thoi gian user phai doi buffer
Upload reliability	99.9%	Upload khong duoc mat file
Scalability	5M DAU, scale to 50M	Thiet ke cho tuong lai
Durability	99.999999999% (11 nines)	Video upload len roi → khong bao gio mat
Global reach	Serve tu CDN gan nhat	User o VN xem video phai nhanh nhu user o My

1.4 Pham vi thiet ke (Out of scope)

De tap trung, chung ta khong thiet ke:

Recommendation engine (ML-based) — chi thiet ke data pipeline cho no
Comment system — da co trong case study khac
Live streaming — chi focus vao Video-on-Demand (VOD)
Monetization / Ads system
Content moderation chi tiet (chi overview)

2. Step 2 — Propose High-Level Design

2.1 Hai luong chinh (Two Main Flows)

Hieu, toan bo YouTube co the chia thanh hai luong chinh hoan toan khac nhau:

Luong	Mo ta	Dac diem
Video Upload Pipeline	User upload video → he thong xu ly → luu tru → phan phoi	Write-heavy, CPU-intensive, async
Video Streaming	User nhan play → he thong tra ve video segments	Read-heavy, bandwidth-intensive, real-time

Key Insight: Hai luong nay co yeu cau hoan toan khac nhau. Upload can processing power (transcoding). Streaming can bandwidth (CDN). Thiet ke rieng biet cho tung luong.

2.2 High-Level Architecture Overview

graph TB
    subgraph Clients
        A[Web Browser]
        B[Mobile App]
        C[Smart TV]
    end

    subgraph "Upload Flow"
        D[API Gateway / Load Balancer]
        E[Upload Service]
        F[Original Video Storage<br/>Blob Store]
        G[Transcoding Service<br/>DAG Scheduler]
        H[Transcoded Video Storage<br/>Blob Store]
        I[Thumbnail Generator]
        J[Metadata Service]
        K[Message Queue<br/>Kafka/RabbitMQ]
    end

    subgraph "Streaming Flow"
        L[CDN Edge Server]
        M[CDN Regional Server]
        N[CDN Origin Server]
        O[Metadata Cache<br/>Redis]
    end

    subgraph "Data Stores"
        P[(Video Metadata DB<br/>MySQL/Vitess)]
        Q[(User Activity DB<br/>Cassandra)]
        R[Search Index<br/>Elasticsearch]
    end

    A & B & C -->|Upload| D
    D --> E
    E --> F
    E -->|Trigger| K
    K --> G
    G --> H
    G --> I
    H --> N
    E --> J
    J --> P
    J --> R

    A & B & C -->|Stream| L
    L -->|Cache miss| M
    M -->|Cache miss| N
    N -->|Fetch| H

    J --> O
    O --> L

2.3 Component Overview

Component	Chuc nang	Tech stack (vi du)
API Gateway	Authentication, rate limiting, routing	Kong, AWS API Gateway
Upload Service	Nhan file, chia chunk, luu vao blob store	Custom service + pre-signed URL
Original Storage	Luu video goc truoc khi transcode	AWS S3, Google Cloud Storage
Transcoding Service	Chuyen doi video thanh nhieu resolution/codec	FFmpeg workers, AWS Elastic Transcoder
Transcoded Storage	Luu video da transcode o moi resolution	S3, GCS (rieng bucket)
CDN	Phan phoi video den user	CloudFront, Akamai, Google CDN
Metadata Service	Quan ly thong tin video	REST API + MySQL/Vitess
Message Queue	Decouple upload va transcoding	Kafka, RabbitMQ → Tuan-08-Message-Queue
Cache	Cache metadata, CDN config	Redis → Tuan-06-Cache-Strategy

3. Step 3 — Design Deep Dive

3.1 Video Upload Pipeline — Chi tiet

3.1.1 Upload Flow chi tiet

Khi user upload video, day la toan bo quy trinh dien ra:

Buoc 1 — Pre-signed URL

Client goi API: “Toi muon upload video, file size = 2GB”
API Server kiem tra authentication, quota, file type
Neu hop le → tao pre-signed upload URL tro thang toi Blob Storage (S3)
Client upload truc tiep len S3, khong di qua API server → giam tai server

Tai sao pre-signed URL? Neu 1000 user upload cung luc, moi file 2GB, API server phai xu ly 2TB bandwidth. Voi pre-signed URL, traffic di thang vao S3 (designed for massive throughput). API server chi xu ly metadata (nhe).

Buoc 2 — Chunked Upload (Resumable)

File 2GB duoc chia thanh nhieu chunks (thuong 5-10MB/chunk)
Client upload tung chunk mot
Neu bi mat mang → chi can upload lai chunk bi loi, khong phai upload lai tu dau
Server theo doi progress: chunk 1/400 ✓, chunk 2/400 ✓, … chunk 235/400 ✗ (retry)

Buoc 3 — Original Storage

Khi tat ca chunks da upload xong → S3 ghep lai thanh file hoan chinh
File goc duoc luu vinh vien (hoac theo retention policy)
Mot event duoc push vao Message Queue → trigger transcoding pipeline

Buoc 4 — Transcoding (xem phan 3.1.2)

Buoc 5 — CDN Distribution

Video da transcode duoc push len CDN origin
CDN tu dong replicate ra edge servers khi co request

Buoc 6 — Metadata Update

Sau khi transcode xong, Metadata DB duoc update:
- Video status: “processing” → “ready”
- Available resolutions: [240p, 360p, 480p, 720p, 1080p]
- Duration, thumbnail URL, manifest file URL
User nhan notification: “Video da san sang!“

3.1.2 Video Transcoding — DAG Architecture

Day la phan phuc tap nhat cua toan bo he thong. Transcoding khong phai chi la “chuyen 1080p thanh 720p.” No la mot DAG (Directed Acyclic Graph) gom nhieu buoc:

graph TD
    A[Original Video File] --> B[Video Splitting<br/>Chia thanh segments]
    B --> C1[Segment 1]
    B --> C2[Segment 2]
    B --> C3[Segment 3]
    B --> C4[Segment N...]

    C1 --> D1[Encode H.264 1080p]
    C1 --> D2[Encode H.264 720p]
    C1 --> D3[Encode H.264 480p]
    C1 --> D4[Encode H.264 360p]
    C1 --> D5[Encode H.264 240p]
    C1 --> D6[Encode VP9 1080p]
    C1 --> D7[Encode VP9 720p]
    C1 --> D8[Encode AV1 1080p]

    C2 --> E1[Encode H.264 1080p]
    C2 --> E2[Encode H.264 720p]
    C2 --> E3[...]

    D1 & D2 & D3 & D4 & D5 & D6 & D7 & D8 --> F[Merge Segments<br/>Per resolution/codec]

    A --> G[Audio Extraction]
    G --> G1[AAC 128kbps]
    G --> G2[AAC 256kbps]
    G --> G3[Opus 128kbps]

    A --> H[Thumbnail Generation<br/>Extract key frames]
    H --> H1[Thumbnail 1]
    H --> H2[Thumbnail 2]
    H --> H3[Thumbnail 3]

    A --> I[Watermark Overlay<br/>Optional - for copyright]

    F --> J[Generate Manifest Files<br/>HLS .m3u8 / DASH .mpd]
    G1 & G2 & G3 --> J
    J --> K[Upload to Transcoded Storage]
    K --> L[Push to CDN Origin]
    L --> M[Update Metadata DB<br/>Status = Ready]

3.1.3 Tai sao dung DAG?

Ly do	Giai thich
Parallelism	Moi segment duoc encode doc lap → chay song song tren nhieu worker
Flexibility	De dang them/bot resolution, codec, buoc xu ly moi
Fault tolerance	Neu encode segment 5 loi → chi retry segment 5, khong phai lam lai tu dau
Priority	Encode 720p truoc (pho bien nhat) → user xem duoc som hon
Resource efficiency	Moi task co resource requirement khac nhau → schedule thong minh

3.1.4 Video Codecs — H.264 vs VP9 vs AV1

Codec	Uu diem	Nhuoc diem	Dung khi nao
H.264 (AVC)	Ho tro 99% thiet bi, decode nhanh	File lon hon	Default cho moi video
VP9	Nho hon H.264 ~30%, free (Google)	Encode cham hon 10x	Chrome, Android, YouTube web
AV1	Nho hon VP9 ~20%, free	Encode cuc cham (50-100x H.264)	Content pho bien, tiet kiem bandwidth
HEVC (H.265)	Nho hon H.264 ~40%	License fee dat, ho tro thiet bi han che	Apple ecosystem, Smart TV

Chien luoc cua YouTube thuc te: Encode H.264 cho tat ca video (compatibility). Voi video popular (nhieu view), encode them VP9 va AV1 de tiet kiem bandwidth CDN — vi chi phi encode 1 lan nhung tiet kiem bandwidth cho hang trieu luot xem.

3.1.5 Transcoding Queue va Priority

Khong phai moi video deu duoc transcode nhu nhau:

Priority	Doi tuong	Xu ly
P0 — Critical	Paid creators (YouTube Partner)	Encode ngay, dedicated workers
P1 — High	Video tu channel lon (>1M subs)	Queue rieng, timeout ngan
P2 — Normal	User thuong	Queue chung, best-effort
P3 — Low	Re-encode video cu sang codec moi	Chi chay luc off-peak (2AM-6AM)

Implement bang multiple message queues voi priority khac nhau:

graph LR
    A[Upload Event] --> B{Priority<br/>Router}
    B -->|P0| C[Critical Queue<br/>Dedicated Workers]
    B -->|P1| D[High Queue<br/>Auto-scaling Workers]
    B -->|P2| E[Normal Queue<br/>Shared Workers]
    B -->|P3| F[Low Queue<br/>Spot Instances Only]

    C --> G[Transcoding Worker Pool]
    D --> G
    E --> G
    F --> G

    G --> H[Transcoded Storage]

Cost optimization: P3 jobs (re-encode video cu) chi chay tren spot instances (re hon 70-90% so voi on-demand). Neu spot bi recall → job tu dong quay lai queue, doi spot instance moi.

3.2 Video Streaming — Adaptive Bitrate

3.2.1 Streaming Protocol: HLS va DASH

Khi user nhan “Play,” dieu gi xay ra?

Buoc 1: Client request manifest file (playlist)

HLS: file .m3u8
DASH: file .mpd

Buoc 2: Manifest file chua danh sach tat ca resolution co san va URL cua tung segment:

#EXT-X-STREAM-INF:BANDWIDTH=800000,RESOLUTION=640x360
360p/segment_001.ts
360p/segment_002.ts
...

#EXT-X-STREAM-INF:BANDWIDTH=2400000,RESOLUTION=1280x720
720p/segment_001.ts
720p/segment_002.ts
...

#EXT-X-STREAM-INF:BANDWIDTH=5000000,RESOLUTION=1920x1080
1080p/segment_001.ts
1080p/segment_002.ts
...

Buoc 3: Client do bandwidth hien tai → chon resolution phu hop

Buoc 4: Client tai tung segment (thuong 2-10 giay/segment)

Buoc 5: Neu bandwidth thay doi (vao thang may, chuyen tu WiFi sang 4G) → client tu dong chuyen resolution cho segment tiep theo

Day la Adaptive Bitrate Streaming (ABR) — Key cua trai nghiem xem video muot. Thay vi buffer va cho, player giam resolution de video tiep tuc chay.

3.2.2 So sanh HLS va DASH

Tieu chi	HLS (Apple)	DASH (MPEG)
Developer	Apple	MPEG consortium (open standard)
Container	`.ts` (Transport Stream)	`.mp4` fragments
Manifest	`.m3u8`	`.mpd` (XML)
DRM support	FairPlay	Widevine, PlayReady
Browser support	Safari native, others via JS	Chrome, Firefox, Edge
Latency	10-30s (can giam voi Low-Latency HLS)	3-10s
YouTube dung?	Co (cho iOS/Safari)	Co (cho Android/Chrome)

Thuc te: YouTube dung ca hai — HLS cho Apple ecosystem, DASH cho phan con lai. Player tu dong chon protocol phu hop.

3.2.3 Adaptive Bitrate Flow

sequenceDiagram
    participant User as User/Player
    participant CDN as CDN Edge
    participant Origin as Origin/Storage

    User->>CDN: GET /video/abc123/manifest.m3u8
    CDN-->>User: Manifest (list of resolutions + segment URLs)

    Note over User: Bandwidth = 5 Mbps → Chon 1080p

    User->>CDN: GET /video/abc123/1080p/seg_001.ts
    CDN-->>User: Segment 1 (1080p)

    User->>CDN: GET /video/abc123/1080p/seg_002.ts
    CDN-->>User: Segment 2 (1080p)

    Note over User: Bandwidth giam xuong 1.5 Mbps<br/>(vao thang may)

    User->>CDN: GET /video/abc123/480p/seg_003.ts
    CDN-->>User: Segment 3 (480p) ← Tu dong ha resolution

    User->>CDN: GET /video/abc123/480p/seg_004.ts
    CDN-->>User: Segment 4 (480p)

    Note over User: Bandwidth phuc hoi 8 Mbps<br/>(ra khoi thang may)

    User->>CDN: GET /video/abc123/1080p/seg_005.ts
    CDN-->>User: Segment 5 (1080p) ← Tu dong tang resolution

    Note over User: Pre-fetch: Player tai truoc<br/>2-3 segments tiep theo

3.2.4 Pre-fetching Strategy

Player thong minh khong chi tai segment hien tai ma con tai truoc (pre-fetch) cac segment tiep theo:

Strategy	Mo ta	Khi nao dung
Eager pre-fetch	Tai truoc 3-5 segments	WiFi/unlimited data, user dang xem lien tuc
Conservative pre-fetch	Tai truoc 1-2 segments	Mobile data, user hay skip
No pre-fetch	Chi tai segment hien tai	Bandwidth cuc thap, user chua nhan Play
Quality pre-fetch	Tai truoc resolution thap, replace bang resolution cao khi ranh	Bandwidth bien dong

Aha Moment: Pre-fetching la ly do ban thay thanh buffer chay truoc vi tri dang xem tren YouTube. No giup video tiep tuc chay ngay ca khi mang cham dot ngot trong 2-3 giay.

3.3 CDN Architecture — Da tang (Multi-tier)

3.3.1 CDN la gi va tai sao la chi phi #1?

CDN (Content Delivery Network) la mang luoi server phan tan tren toan cau, cache noi dung gan user nhat co the → Tuan-03-Networking-DNS-CDN.

Voi video streaming, CDN bandwidth chinh la chi phi lon nhat:

Hang muc chi phi	% tong chi phi (uoc tinh)	Ghi chu
CDN bandwidth	50-70%	Chi phi #1, khong the tranh
Storage (S3/GCS)	10-15%	Video goc + transcoded
Transcoding compute	10-15%	CPU-intensive
Metadata infra (DB, cache)	5-10%	Tuong doi re
Network (non-CDN)	3-5%	Inter-region transfer
Other (monitoring, etc.)	2-5%	—

3.3.2 CDN Multi-tier Architecture

graph TD
    subgraph "Tier 1 — Edge (PoP)"
        E1[Edge HCM<br/>Cache: Hot content<br/>TTL: 24h]
        E2[Edge Ha Noi<br/>Cache: Hot content<br/>TTL: 24h]
        E3[Edge Bangkok<br/>Cache: Hot content<br/>TTL: 24h]
        E4[Edge Tokyo<br/>Cache: Hot content<br/>TTL: 24h]
        E5[Edge SF<br/>Cache: Hot content<br/>TTL: 24h]
    end

    subgraph "Tier 2 — Regional"
        R1[Regional Singapore<br/>Cache: Warm content<br/>TTL: 7 days]
        R2[Regional US-West<br/>Cache: Warm content<br/>TTL: 7 days]
        R3[Regional EU-West<br/>Cache: Warm content<br/>TTL: 7 days]
    end

    subgraph "Tier 3 — Origin"
        O1[Origin Storage<br/>US-East<br/>All content]
        O2[Origin Storage<br/>EU-Central<br/>All content]
    end

    E1 & E2 & E3 -->|Cache miss| R1
    E4 -->|Cache miss| R1
    E5 -->|Cache miss| R2

    R1 -->|Cache miss| O1
    R2 -->|Cache miss| O1
    R3 -->|Cache miss| O2

    O1 <-->|Cross-region<br/>replication| O2

3.3.3 Push vs Pull CDN

Kieu	Cach hoat dong	Uu diem	Nhuoc diem	Dung cho
Push	Server chu dong day content len CDN truoc khi co request	Luot xem dau tien cung nhanh (da co cache)	Ton storage, day noi dung khong ai xem	Video trending, popular creators
Pull	CDN chi fetch content khi co request dau tien (cache miss)	Tiet kiem storage, chi cache content thuc su can	Request dau tien cham (phai fetch tu origin)	Long-tail content (video it view)

Chien luoc ket hop: Video moi tu popular creators (>100K subs) → Push len edge ngay. Video tu user thuong → Pull khi co nguoi xem.

3.3.4 Long-tail Content Strategy

Thuc te, phan bo view cua YouTube tuan theo Power Law (quy luat luy thua):

Top 1% video chiem ~80% tong view → day len CDN edge, cache lau
Top 10% video chiem ~95% tong view → cache o regional CDN
Bottom 90% video (“long-tail”) chiem ~5% tong view → serve tu origin, khong cache

Tier	Content type	Cache policy	Chi phi/view
Edge (expensive)	Top 1% — trending, viral	Cache 24h, push proactively	Thap nhat (da cache)
Regional (moderate)	Top 10% — popular, recent	Cache 7 ngay, pull-based	Trung binh
Origin (cheapest storage)	Bottom 90% — long-tail	Khong cache tren CDN	Cao nhat (fetch tu origin)

Aha Moment: Neu ban cache tat ca video tren CDN edge → chi phi CDN se gap 10-20 lan. Bi quyet la chi cache nhung gi thuc su duoc xem nhieu. Day la ly do YouTube chi tieu hang ty USD/nam cho CDN ma van phai toi uu lien tuc.

3.4 Metadata System

3.4.1 Video Metadata DB

Truong	Type	Giai thich
video_id	UUID / Snowflake ID	Primary key, globally unique
title	VARCHAR(500)	Tieu de video
description	TEXT	Mo ta, co the dai
uploader_id	BIGINT (FK)	Nguoi upload
status	ENUM	uploading, processing, ready, failed, deleted
duration_seconds	INT	Thoi luong
file_size_bytes	BIGINT	Kich thuoc file goc
upload_time	TIMESTAMP	Thoi gian upload
available_resolutions	JSON	[“240p”, “360p”, “480p”, “720p”, “1080p”]
manifest_url	VARCHAR(1000)	URL toi HLS/DASH manifest
thumbnail_urls	JSON	Mang URL thumbnails
view_count	BIGINT	So luot xem (eventual consistent)
like_count	BIGINT	So luot thich
category	VARCHAR(100)	The loai
tags	JSON	Tags cho search
is_public	BOOLEAN	Video cong khai hay rieng tu
drm_protected	BOOLEAN	Co bao ve DRM khong

Chon DB: MySQL voi Vitess (sharding layer) hoac CockroachDB — vi metadata can strong consistency (khong duoc hien thi video chua transcode xong).

Sharding key: video_id — phan bo deu, query chinh la theo video_id.

3.4.2 User Activity Data

Data type	Volume	Storage	Giai thich
Watch history	Cuc lon	Cassandra / BigTable	Moi user xem ~20 video/ngay
Search history	Lon	Elasticsearch	Full-text search
Like/dislike	Lon	MySQL (counter) + Kafka	Event-driven update
Watch progress	Cuc lon	Redis (TTL) → Cassandra	”Tiep tuc xem tu 15:32”
Recommendation data	Cuc lon	BigQuery / Data warehouse	ML pipeline input

Luu y: View count la eventually consistent. YouTube khong dam bao view count chinh xac tuc thoi — no duoc aggregate theo batch (moi 5-10 phut) de giam write load.

3.4.3 Search Index

Video metadata duoc index vao Elasticsearch de ho tro full-text search:

Truong duoc index	Boost weight	Giai thich
title	10x	Quan trong nhat
description	3x	Phu, nhung co nhieu keyword
tags	5x	Creator tu gan tag
channel_name	7x	Search theo kenh
captions/subtitles	1x	Tim trong phu de (cuc manh)

3.5 Error Handling & Reliability

3.5.1 Resumable Upload

Van de	Giai phap	Chi tiet
Mat mang giua upload	Chunked upload + checksum	Client ghi nho chunk cuoi thanh cong, retry tu do
File corrupt	MD5/SHA256 checksum per chunk	Server verify tung chunk, reject neu sai
Upload timeout	Pre-signed URL co TTL (thuong 24h)	Qua han → client request URL moi
Duplicate upload	Idempotency key (hash cua file)	Neu cung file → khong upload lai
Server crash	Upload state luu trong durable storage	Khoi dong lai → doc state, tiep tuc

3.5.2 Transcoding Error Handling

Loi	Xu ly	Retry policy
Segment encode fail	Retry segment do (khong phai ca video)	3 lan, exponential backoff (1s → 2s → 4s)
Worker crash	Task quay lai queue, worker khac pick up	Heartbeat timeout = 30s
Out of memory	Retry tren worker co nhieu RAM hon	Scale up worker class
Corrupt input	Mark video = “failed,” thong bao user	Khong retry — user can upload lai
Timeout (encode qua lau)	Chia segment nho hon, retry	Dynamic segment sizing

3.5.3 Streaming Error Handling

Van de	Giai phap
CDN edge down	DNS failover sang edge khac (automatic)
Segment 404 (chua cache)	CDN pull tu regional → origin
Bandwidth drop	ABR tu dong ha resolution
Video bị xoa giua luc xem	Graceful error message, goi y video khac
Region-specific outage	Cross-region CDN failover

3.5.4 Pre-signed URL Security

Buoc	Chi tiet
1. Client request upload	Gui authentication token + file metadata
2. Server validate	Kiem tra token, quota, file type allowed
3. Generate pre-signed URL	URL chua: bucket, key, expiration, signature
4. Client upload truc tiep	Upload thang len S3 voi pre-signed URL
5. URL het han	Sau 1-24h, URL khong con su dung duoc
6. Playback URL	Cung dung pre-signed URL voi TTL ngan (vd 6h) → chong share link trai phep

3.6 Cost Optimization — Chien luoc song con

3.6.1 CDN Cost Optimization

Day la phan quan trong nhat ve mat business — vi CDN chiem 50-70% chi phi:

Chien luoc	Tiet kiem	Chi tiet
Long-tail tu origin	30-40% CDN cost	90% video it view → serve tu origin (S3 bandwidth re hon CDN)
Regional CDN selection	10-20%	Dung CDN re hon o region it traffic (VD: dung ISP CDN cho VN thay vi Akamai)
Codec optimization	20-30% bandwidth	VP9/AV1 cho popular video → giam bandwidth 20-30% per view
Off-peak transcoding	40-60% compute	Transcode video khong gap vao luc 2AM-6AM, dung spot instances
Intelligent caching	15-25% CDN cost	Cache TTL based on popularity — hot video cache lau, cold video cache ngan
Peer-to-peer (P2P)	10-30% CDN cost	Cho phep viewers share segments voi nhau (WebRTC)
Multi-CDN	10-15%	Dung nhieu CDN provider, route traffic theo gia + performance

3.6.2 Storage Cost Optimization

Chien luoc	Chi tiet
Tiered storage	Video moi → S3 Standard. Sau 30 ngay → S3 Infrequent Access. Sau 1 nam → S3 Glacier
Delete old resolutions	Video >2 nam, khong ai xem → xoa resolution thap (240p, 360p), giu 720p + 1080p
Lazy transcoding	Video moi: chi encode 720p + 1080p truoc. 240p, 360p, 4K → chi encode khi co request
Deduplication	Detect video trung lap (fingerprinting) → khong luu 2 ban

3.6.3 Transcoding Cost Optimization

Chien luoc	Chi tiet
Spot instances	Dung AWS Spot / GCP Preemptible cho P2, P3 jobs → re hon 70-90%
Reserved capacity	P0, P1 jobs chay tren reserved instances → dam bao availability
Off-peak scheduling	P3 jobs (re-encode) chi chay 2AM-6AM khi gia compute re nhat
Right-sizing	Video ngan (<1 phut) → worker nho. Video dai (>1h) → worker lon voi nhieu CPU
Hardware acceleration	Dung GPU (NVENC) hoac dedicated hardware (AWS MediaConvert) cho encode nhanh hon 5-10x

4. Capacity Estimation — Uoc luong chi tiet

4.1 Assumptions

Thong so	Gia tri	Giai thich
DAU	5,000,000	5M daily active users
Avg watch time/day	30 phut	Engagement trung binh
Avg video bitrate (streaming)	2.5 Mbps	Trung binh giua 480p va 1080p
Upload users/day	0.1% DAU	5,000 creators upload/day
Avg upload video duration	10 phut	Trung binh
Avg upload video size (original)	1.5 GB	1080p, 10 phut
Transcoding expansion factor	3x	1 video goc → ~3x storage (nhieu resolution)
Video segment size	4 giay	Cho HLS/DASH

4.2 Storage Estimation

Upload storage moi ngay (video goc):

U pl o a d / d a y = 5, 000 \times 1.5 GB = 7, 500 GB = 7.5 TB / d a y

Transcoded storage moi ngay (nhieu resolution):

T r an sco d e d / d a y = 7.5 TB \times 3 = 22.5 TB / d a y

Tong storage moi ngay:

T o t a l s t or a g e / d a y = 7.5 + 22.5 = 30 TB / d a y

Storage 1 nam:

T o t a l s t or a g e / ye a r = 30 TB \times 365 = 10, 950 TB \approx 11 PB / ye a r

Chi phi storage (AWS S3 Standard ~ $0.023/GB/month):

M o n t h l y s t or a g e cos t = 11, 000 TB \times $0.023/ GB = $253, 000/ m o n t h (ye a r 1 a vg)

Luu y: Con so nay chua tinh tiered storage (chuyen video cu sang S3 IA/Glacier se giam 50-70%).

4.3 Bandwidth Estimation — Streaming

Concurrent viewers (peak):

C o n c u rre n t v i e w er s_{p e ak} = D A U \times \frac{30 min}{1440 min} \times 3 (p e ak f a c t or) \approx 5 M \times 0.021 \times 3 = 312, 500

Bandwidth peak:

B an d w i d t h_{p e ak} = 312, 500 \times 2.5 M b p s = 781, 250 M b p s \approx 781 G b p s

Aha Moment: 781 Gbps la con so khong lo. Day la ly do YouTube can mang luoi CDN toan cau voi hang ngan edge server — khong mot data center don le nao co the serve luong bandwidth nay.

Bandwidth trung binh:

B an d w i d t h_{a vg} = \frac{781 G b p s}{3} \approx 260 G b p s

Data transfer moi ngay:

D ai l y d a t a t r an s f er = 5 M \times 30 min \times 60 s \times 2.5 M b p s = 22, 500, 000 G b = 2, 812 TB / d a y \approx 2.8 PB / d a y

Chi phi CDN bandwidth (AWS CloudFront ~ $0.02/GB o scale lon):

C D N cos t / d a y = 2, 812 TB \times $0.02/ GB = 2, 812, 000 \times $0.02 = $56, 240/ d a y

C D N cos t / m o n t h = $56, 240 \times 30 = $1, 687, 200/ m o n t h \approx $1.7 M / m o n t h

Day la chi phi CDN cho 5M DAU. YouTube voi 2 ty DAU → nhan len ~400x (nhung duoc discount volume). Day la ly do CDN cost la #1 expense.

4.4 Transcoding Compute Estimation

So video can transcode moi ngay:

Vi d eos / d a y = 5, 000

Trung binh moi video can encode ra 6 resolution × 2 codec = 12 variants:

T r an sco d in g j o b s / d a y = 5, 000 \times 12 = 60, 000 j o b s

Thoi gian encode trung binh 1 variant (10 phut video):

E n co d e t im e / v a r ian t \approx 15 p h u t (H .264), 60 p h u t (V P 9)

Tong compute time (gia dinh 70% H.264, 30% VP9):

T o t a l co m p u t e = 60, 000 \times (0.7 \times 15 + 0.3 \times 60) = 60, 000 \times 28.5 = 1, 710, 000 CP U - min u t es / d a y

= 28, 500 CP U - h o u rs / d a y

So workers can (moi worker 1 vCPU, chay 20h/day):

W or k ers = \frac{28 , 500}{20} = 1, 425 w or k ers

Chi phi (AWS c5.xlarge spot ~ $0.05/h):

C o m p u t e cos t / d a y = 28, 500 \times $0.05 = $1, 425/ d a y \approx $42, 750/ m o n t h

4.5 Tong hop chi phi

Hang muc	Chi phi/thang	%
CDN bandwidth	$1,700,000	72%
Storage (S3)	$253,000	11%
Transcoding compute	$42,750	2%
Metadata infra (DB, cache, search)	$150,000	6%
Network (non-CDN)	$120,000	5%
Monitoring, ops, other	$100,000	4%
Tong	~$2,365,750	100%

Validation: CDN chiem 72% — dung voi industry benchmark (50-70%). Con so nay xac nhan rang CDN cost optimization la uu tien #1.

5. Security — Bao mat Video Platform

5.1 DRM (Digital Rights Management)

DRM bao ve noi dung co ban quyen (phim, show, music video) khoi bi sao chep trai phep.

DRM System	Platform	Su dung boi
Widevine (Google)	Chrome, Android, Chromecast	YouTube, Netflix, Disney+
FairPlay (Apple)	Safari, iOS, Apple TV	YouTube (tren iOS), Apple TV+
PlayReady (Microsoft)	Edge, Xbox, Windows	YouTube (tren Edge), Netflix

Cach DRM hoat dong:

Buoc	Chi tiet
1. Encrypt video	Moi segment duoc encrypt bang AES-128 key
2. License server	Key duoc luu tren License Server (khong gui cung video)
3. Client request license	Player gui device info + authentication → License Server
4. License Server validate	Kiem tra: user co quyen xem? Device hop le? Region cho phep?
5. Tra ve license	License chua decryption key, co TTL (vd 24h)
6. Client decrypt & play	Player decrypt segments in-memory, khong luu file giai ma

Luu y: DRM khong phai bulletproof — luon co cach bypass (screen recording). Nhung no tang chi phi cho piracy, du de bao ve cho da so use case.

5.2 Content Copyright Detection (Content ID)

YouTube dung he thong Content ID de phat hien vi pham ban quyen:

Buoc	Chi tiet
1. Fingerprinting	Moi video duoc tao “fingerprint” (audio + visual) khi upload
2. Database matching	Fingerprint duoc so sanh voi database cua copyright holders
3. Match found	Neu match → tu dong: block video, cho chay nhung chia doanh thu, hoac chi track
4. Appeal process	Uploader co the khieu nai (dispute) neu la fair use

Ky thuat	Giai thich
Audio fingerprinting	Chromaprint / Shazam-like — trích xuat melody, rhythm, tao hash
Visual fingerprinting	Perceptual hashing — so sanh frame-by-frame, chiu duoc crop/resize
Temporal matching	Khong chi match toan bo video ma con match 1 doan (vd: 30 giay nhac nen)

5.3 Abuse Prevention

Loai abuse	Giai phap
Illegal content	ML-based scanning (CSAM detection, violence detection) truoc khi publish
Spam upload	Rate limiting per user, CAPTCHA cho upload, file type validation
View botting	Anomaly detection: IP clustering, behavioral analysis, khong dem view tu bot
DDoS	CDN-level protection (Cloudflare, AWS Shield), rate limiting → Tuan-09-Rate-Limiter
Account takeover	2FA, suspicious login detection, session management

5.4 Pre-signed URL cho Upload va Playback

Use case	TTL	Giai thich
Upload URL	1-24h	Du thoi gian upload file lon, nhung khong qua lau de bi leak
Playback manifest URL	6-12h	Du cho 1 session xem, het han → client request moi
Playback segment URL	1-6h	Ngan hon manifest, vi segment URL duoc refresh lien tuc
Thumbnail URL	30 ngay	Thumbnail cong khai, khong can bao mat cao

5.5 Watermarking cho Leak Tracking

Loai watermark	Chi tiet	Dung khi
Visible watermark	Logo/text hien thi tren video	Ngan sao chep re-upload
Invisible watermark	Embed thong tin user vao pixel data (khong nhin thay)	Tracking leak source
Forensic watermark	Moi user nhan version video hoi khac nhau	Xac dinh chinh xac ai leak

Forensic watermarking: Netflix dung ky thuat nay — moi subscriber nhan video voi watermark khac nhau (invisible). Neu video bi leak, phân tich watermark → biet chinh xac account nao leak.

6. DevOps & Monitoring — Van hanh Video Platform

6.1 Transcoding Pipeline Monitoring

Metric	Mo ta	Alert threshold
Queue depth	So job dang cho trong transcoding queue	> 10,000 jobs → scale up workers
Processing time p99	Thoi gian encode 1 video (p99)	> 2h cho video 10 phut → investigate
Failure rate	% job bi fail	> 2% → alert, > 5% → page on-call
Worker utilization	CPU usage cua transcoding workers	< 30% → scale down, > 80% → scale up
Queue wait time	Thoi gian job cho trong queue truoc khi duoc pick up	P0 > 1 phut → alert ngay

6.2 CDN Performance Monitoring

Metric	Mo ta	Target
Cache hit ratio	% request duoc serve tu cache (khong can fetch origin)	> 95% cho edge, > 85% cho regional
Origin offload	% traffic KHONG phai di ve origin	> 90%
Edge latency (TTFB)	Time to First Byte tu CDN edge	< 50ms
Cache fill rate	Toc do content duoc cache lan dau	Monitor de detect cache stampede
Bandwidth per PoP	Bandwidth tai moi CDN edge	Dung de capacity planning
Error rate per PoP	% request loi tai moi edge	> 0.1% → investigate edge health

6.3 Streaming Quality Monitoring

Metric	Mo ta	Target	Anh huong UX
Buffering ratio	% thoi gian user phai doi buffer	< 1%	Cao nhat — user roi di neu buffer nhieu
Video Start Time (VST)	Thoi gian tu nhan Play → frame dau tien	< 2s	User mat kien nhan sau 3s
Rebuffer frequency	So lan buffer/gio xem	< 0.5 lan/gio	Moi lan buffer = trai nghiem xau
Resolution distribution	% user xem o moi resolution	Trend tracking	Giam 1080p → mang co van de
Bitrate switches	So lan player doi resolution/phut	< 2 lan/phut	Nhieu lan doi = mang khong on dinh
Playback failure rate	% video khong play duoc	< 0.1%	User khong xem duoc = worst case

6.4 Error Rate per Region

Region	Metric can theo doi	Hanh dong
Vietnam	Buffering ratio, CDN latency	Neu cao → them edge o VN hoac dung ISP peering
SEA	Cache hit ratio	Neu thap → tang cache capacity o Singapore
US/EU	Playback failure rate	Neu cao → check CDN provider health
Emerging markets	Resolution distribution	Neu da so 240p/360p → toi uu cho low bandwidth

6.5 Cost Dashboard

Dashboard	Metric	Update frequency
CDN Cost per View	Tong CDN cost / tong views	Hourly
CDN Cost per Region	CDN cost breakdown theo region	Daily
Transcoding Cost per Video	Avg cost de transcode 1 video	Daily
Storage Growth Rate	TB added/day, projected cost in 30 days	Daily
Cost per DAU	Tong infra cost / DAU	Weekly
CDN Provider Comparison	Cost + performance cua moi CDN provider	Weekly

6.6 Alerting Tiers

Tier	Severity	Response time	Vi du
P0 — Critical	Toan bo streaming down	5 phut	CDN origin unreachable, DB master down
P1 — Major	1 region bi anh huong	15 phut	Edge HCM down, transcoding queue stuck
P2 — Minor	Performance degrade	1 gio	Cache hit ratio < 90%, encode time tang 2x
P3 — Warning	Anomaly detected	Next business day	Cost spike, unusual traffic pattern

7. Mermaid Diagrams — Tong hop

7.1 Upload Pipeline — End to End

graph TD
    A[Client] -->|1. Request pre-signed URL| B[API Server]
    B -->|2. Validate auth + quota| B
    B -->|3. Return pre-signed URL| A
    A -->|4. Upload chunks truc tiep| C[Blob Storage<br/>S3/GCS]
    C -->|5. All chunks received<br/>Trigger event| D[Message Queue<br/>Kafka]
    D -->|6. Dispatch| E{Priority Router}

    E -->|P0| F[Critical Queue]
    E -->|P1| G[High Queue]
    E -->|P2| H[Normal Queue]
    E -->|P3| I[Low Queue]

    F & G & H & I --> J[DAG Scheduler]

    J --> K[Video Splitting]
    K --> L[Parallel Encoding<br/>H.264 / VP9 / AV1<br/>Multiple resolutions]
    K --> M[Audio Extraction<br/>AAC / Opus]
    K --> N[Thumbnail Generation]
    K --> O[Watermark Overlay]

    L --> P[Segment Merging<br/>Per resolution/codec]
    P --> Q[Manifest Generation<br/>HLS .m3u8 / DASH .mpd]
    M --> Q
    Q --> R[Transcoded Storage<br/>S3/GCS]
    R --> S[CDN Origin Push]
    S --> T[CDN Edge Distribution]

    Q --> U[Metadata DB Update<br/>status = ready]
    U --> V[Notification Service<br/>Video san sang!]
    V --> A

7.2 Streaming Architecture — Full Flow

sequenceDiagram
    participant Client as Client (Player)
    participant DNS as DNS / GSLB
    participant Edge as CDN Edge
    participant Regional as CDN Regional
    participant Origin as CDN Origin
    participant Storage as Blob Storage
    participant Meta as Metadata Service
    participant Cache as Redis Cache

    Client->>Meta: GET /api/video/abc123 (video info)
    Meta->>Cache: Lookup video metadata
    Cache-->>Meta: Cache hit ✓
    Meta-->>Client: Video info + manifest URL

    Client->>DNS: Resolve CDN hostname
    DNS-->>Client: Nearest edge IP (GSLB)

    Client->>Edge: GET manifest.m3u8
    Edge-->>Client: Manifest (resolution list)

    loop Moi segment (2-10s video)
        Client->>Edge: GET /1080p/seg_N.ts
        alt Cache HIT
            Edge-->>Client: Segment (from cache)
        else Cache MISS
            Edge->>Regional: Fetch segment
            alt Regional cache HIT
                Regional-->>Edge: Segment
            else Regional cache MISS
                Regional->>Origin: Fetch segment
                Origin->>Storage: Read from blob
                Storage-->>Origin: Raw segment
                Origin-->>Regional: Segment (cached)
            end
            Regional-->>Edge: Segment (cached)
            Edge-->>Client: Segment (cached for next user)
        end
    end

    Note over Client: ABR: neu bandwidth giam<br/>→ chuyen sang resolution thap<br/>cho segment tiep theo

7.3 Adaptive Bitrate Decision Flow

graph TD
    A[Player bat dau] --> B[Request manifest file]
    B --> C[Nhan danh sach resolutions<br/>240p / 360p / 480p / 720p / 1080p / 4K]
    C --> D[Do bandwidth hien tai]
    D --> E{Bandwidth level?}

    E -->|"> 8 Mbps"| F[Chon 4K<br/>~15 Mbps bitrate]
    E -->|"5-8 Mbps"| G[Chon 1080p<br/>~5 Mbps bitrate]
    E -->|"2.5-5 Mbps"| H[Chon 720p<br/>~2.5 Mbps bitrate]
    E -->|"1-2.5 Mbps"| I[Chon 480p<br/>~1 Mbps bitrate]
    E -->|"0.5-1 Mbps"| J[Chon 360p<br/>~0.5 Mbps bitrate]
    E -->|"< 0.5 Mbps"| K[Chon 240p<br/>~0.25 Mbps bitrate]

    F & G & H & I & J & K --> L[Tai segment N voi resolution da chon]
    L --> M[Play segment N]
    M --> N{Tai segment N+1?}
    N --> O[Do lai bandwidth]
    O --> P{Bandwidth thay doi?}

    P -->|"Giam > 20%"| Q[Ha resolution<br/>cho segment N+1]
    P -->|"Tang > 30%"| R[Tang resolution<br/>cho segment N+1]
    P -->|"On dinh"| S[Giu nguyen resolution]

    Q & R & S --> T[Pre-fetch 2-3 segments tiep]
    T --> L

7.4 CDN Multi-tier Decision Flow

graph TD
    A[User request video segment] --> B[DNS/GSLB<br/>Route to nearest Edge]
    B --> C{Edge cache<br/>co segment?}

    C -->|HIT| D[Serve tu Edge<br/>Latency: 5-20ms]
    C -->|MISS| E{Regional cache<br/>co segment?}

    E -->|HIT| F[Serve tu Regional<br/>Cache len Edge<br/>Latency: 20-50ms]
    E -->|MISS| G{Origin cache<br/>co segment?}

    G -->|HIT| H[Serve tu Origin<br/>Cache len Regional + Edge<br/>Latency: 50-200ms]
    G -->|MISS| I[Fetch tu Blob Storage<br/>Cache len Origin + Regional + Edge<br/>Latency: 100-500ms]

    D --> J[User xem video]
    F --> J
    H --> J
    I --> J

    J --> K{Video popular?}
    K -->|"Top 1%"| L[Proactive push<br/>to all edges]
    K -->|"Top 10%"| M[Cache dai han<br/>o regional]
    K -->|"Bottom 90%"| N[Ngan cache TTL<br/>hoac khong cache]

8. Aha Moments & Pitfalls

8.1 Aha Moments — Nhung dieu bat ngo

#	Aha Moment	Giai thich
1	CDN cost la #1 expense	Khong phai server, khong phai storage — CDN bandwidth chiem 50-70% tong chi phi. Moi quyet dinh thiet ke phai xoay quanh viec giam CDN cost.
2	Video khong duoc “gui” — no duoc “cat thanh lat mong va dut”	Streaming = gui hang ngan segment nho, moi segment 2-10 giay. Player quyет dinh resolution cho tung segment. Khong phai download toan bo roi play.
3	Adaptive Bitrate la key cua UX	Khong co ABR, user phai tu chon resolution. Chon sai → buffer. ABR tu dong dieu chinh → video luon chay, chi doi chat luong. Day la ly do YouTube hiém khi buffer.
4	Long-tail strategy quyet dinh loi nhuan	90% video chi co 5% view. Neu cache tat ca tren CDN → pha san. Bi quyet: chi cache 10% popular content, 90% con lai serve tu origin (re hon 10x).
5	Transcoding la CPU-heavy, phu hop spot instances	Transcoding la batch job, co the retry. Spot instances re hon 70-90% → tiet kiem hang trieu USD/nam. Chap nhan duoc vi transcoding la async, khong anh huong user truc tiep.
6	Encode 1 lan, serve trieu lan	Chi phi encode VP9/AV1 gap 10-50x H.264, nhung tiet kiem 20-30% bandwidth moi lan xem. Voi video 10 trieu view → tiet kiem khong lo.
7	DAG architecture cho transcoding	Khong phai “1 job = 1 video.” 1 video = nhieu task doc lap (split, encode, merge, thumbnail). DAG cho phep parallel, retry tung task, va priority scheduling.
8	Pre-signed URL la must-have	Upload: tranh API server lam bottleneck. Playback: kiem soat ai duoc xem, bao lau, chong share link trai phep.

8.2 Pitfalls — Nhung loi thuong gap trong interview

#	Pitfall	Giai thich	Cach tranh
1	Khong tach Upload va Streaming	Hai luong nay co yeu cau hoan toan khac nhau. Thiet ke chung → khong toi uu duoc.	Luon ve 2 pipeline rieng biet
2	Quen CDN cost	Noi “dung CDN” ma khong ban ve cost optimization → interviewer nghi ban khong hieu scale thuc te	Luon de cap long-tail strategy, multi-CDN, codec optimization
3	Chi noi “transcode video”	Khong giai thich DAG, parallel processing, priority queue → interviewer nghi ban chi biet surface level	Mo ta DAG: split → encode tung segment → merge. Noi ve priority va spot instances
4	Quen Adaptive Bitrate	Noi “stream video cho user” ma khong noi ve ABR → interviewer se hoi “neu mang cham thi sao?”	Luon noi ve HLS/DASH, ABR, segment-based streaming
5	Khong noi ve resumable upload	Voi file 2-50GB, upload thuong bi gian doan. Khong co resumable = UX te	Noi ve chunked upload, checksum, retry
6	Thiet ke monolithic	Gop tat ca vao 1 service → khong scale duoc tung phan	Upload service, transcoding service, streaming (CDN), metadata service — tach rieng
7	Quen error handling	”Video upload xong, transcode xong, xem duoc” — qua ly tuong. Thuc te: network fail, disk full, codec error	Noi ve retry, exponential backoff, dead letter queue, alerting

8.3 Interview Tips — Chien luoc trinh bay

Buoc	Thoi gian	Noi gi
1. Clarify	3-5 phut	Hoi: upload + streaming? Scale? Resolution? Paid vs free?
2. High-level	5-7 phut	Ve 2 pipeline: Upload (pre-signed URL → storage → transcode → CDN) + Streaming (CDN → ABR)
3. Deep dive	15-20 phut	Chon 2-3 topic de deep dive: transcoding DAG, CDN multi-tier, ABR. Khong co lam tat ca
4. Cost & Scale	5 phut	Back-of-envelope cho storage + bandwidth. Nhan manh CDN cost la #1
5. Wrap up	3 phut	Noi ve DRM, error handling, monitoring. Goi y extensions: live streaming, recommendation

Pro tip: Khi interviewer hoi “Design YouTube,” ho khong mong doi ban thiet ke tat ca. Ho muon thay ban chon dung van de de deep dive va giai thich ro rang tai sao.

9. Wrap Up — Step 4: Extensions & Trade-offs

9.1 Possible Extensions (neu co thoi gian)

Extension	Mo ta ngán
Live Streaming	RTMP ingest → transcode real-time → HLS/DASH out. Khac VOD: latency quan trong, khong co pre-transcode
Recommendation Engine	Collaborative filtering + content-based. Input: watch history, like, search. Output: suggested videos
Comment System	Threaded comments, real-time update, spam detection. Sharding by video_id
Analytics Dashboard	Creator analytics: views, watch time, demographics. Dung data warehouse (BigQuery)
Multi-language	Auto-generated captions (Speech-to-Text), translation, subtitle management
Monetization	Ad insertion (pre-roll, mid-roll), subscription tiers, super chat
Offline Viewing	Download video cho xem offline. DRM van phai hoat dong offline (license pre-fetched)

9.2 Trade-offs Summary

Quyet dinh	Option A	Option B	YouTube chon
Transcoding timing	Eager (encode tat ca resolution ngay)	Lazy (encode khi co request)	Eager cho popular codecs, Lazy cho rare resolutions
CDN strategy	Push (day len CDN truoc)	Pull (CDN fetch khi can)	Push cho popular, Pull cho long-tail
Storage	1 region	Multi-region	Multi-region voi cross-region replication
Codec priority	Chi H.264 (nhanh, compatible)	Multi-codec (H.264 + VP9 + AV1)	Multi-codec — tiet kiem bandwidth cho popular content
Upload	Qua API server	Pre-signed URL truc tiep len storage	Pre-signed URL — giam tai API server
Metadata consistency	Strong consistency	Eventual consistency	Strong cho video status, Eventual cho view count
Queue architecture	1 queue chung	Multi-queue voi priority	Multi-queue — P0/P1/P2/P3 rieng biet

9.3 Lien ket noi bo (Internal Links)

Topic	Link	Lien quan
CDN va DNS	Tuan-03-Networking-DNS-CDN	CDN multi-tier, DNS-based routing, GSLB
Cache Strategy	Tuan-06-Cache-Strategy	CDN cache policy, Redis metadata cache, cache invalidation
Message Queue	Tuan-08-Message-Queue	Transcoding queue, priority queue, event-driven architecture
Database Sharding	Tuan-07-Database-Sharding-Replication	Metadata DB sharding by video_id, replication cho read
Rate Limiter	Tuan-09-Rate-Limiter	Upload rate limiting, API protection
Monitoring	Tuan-13-Monitoring-Observability	Streaming quality metrics, alerting tiers
Security	Tuan-15-Data-Security-Encryption	DRM, encryption, pre-signed URL
Back-of-envelope	Tuan-02-Back-of-the-envelope	Capacity estimation methodology

“YouTube khong phai la 1 he thong — no la hang chuc he thong lam viec cung nhau. Upload pipeline, transcoding DAG, CDN network, metadata service, recommendation engine… Moi phan co the la 1 bai system design rieng. Bi quyet trong interview la biet focus vao dung phan quan trong nhat: CDN cost optimization va adaptive bitrate streaming.”

lthieu's notes

Explorer

Case-Design-YouTube-Video-Streaming