| @@ -0,0 +1,40 @@ | |||
| 1 | + | # System Design Interview: Designing a Video Streaming Service (YouTube-style) | |
| 2 | + | ||
| 3 | + | This question is mostly about the upload pipeline and efficient delivery, not the database. | |
| 4 | + | ||
| 5 | + | ## Requirements | |
| 6 | + | ||
| 7 | + | - Users upload videos; others watch them on phones, laptops and TVs with smooth playback. | |
| 8 | + | - Playback adapts to network speed. | |
| 9 | + | - Example scale: 5M daily viewers, 500k uploads a day, average video 300 MB before processing. | |
| 10 | + | ||
| 11 | + | ## Upload path | |
| 12 | + | ||
| 13 | + | 1. The client asks the API for an upload URL and gets a **pre-signed URL** to object storage, so large files bypass your app servers. | |
| 14 | + | 2. Large files use **resumable / multipart uploads** so a dropped connection doesn't restart from zero. | |
| 15 | + | 3. When the upload completes, storage emits an event onto a queue. | |
| 16 | + | ||
| 17 | + | ## Processing pipeline | |
| 18 | + | ||
| 19 | + | Workers pull jobs and **transcode** the original into several resolutions and bitrates (for example 240p to 1080p) and codecs. Each rendition is split into short **segments** (2-6 seconds) with a manifest file (HLS `.m3u8` or DASH `.mpd`). | |
| 20 | + | ||
| 21 | + | Model it as a DAG: split into chunks, transcode chunks in parallel, generate thumbnails, run content checks, then merge and publish. Parallel chunk transcoding is what keeps a 1-hour video from taking hours. | |
| 22 | + | ||
| 23 | + | ## Playback path | |
| 24 | + | ||
| 25 | + | - Segments and manifests are served from a **CDN**, which absorbs nearly all read traffic. | |
| 26 | + | - The player uses **adaptive bitrate streaming**: it measures throughput and switches renditions between segments. | |
| 27 | + | - Popular videos stay hot in the CDN; long-tail videos may be fetched from origin storage on demand. | |
| 28 | + | ||
| 29 | + | ## Metadata | |
| 30 | + | ||
| 31 | + | A relational or wide-column store holds video records (owner, title, status, renditions). View counts are high-write, so buffer them (stream into a counter service and flush in batches) rather than updating a row per view. | |
| 32 | + | ||
| 33 | + | ## Trade-offs to mention | |
| 34 | + | ||
| 35 | + | - Transcoding every upload to every resolution costs compute and storage; you can transcode rare resolutions lazily. | |
| 36 | + | - CDN cost is the largest bill; regional caching and good cache keys matter. | |
| 37 | + | - Copyright and safety checks add latency before a video goes public. | |
| 38 | + | ||
| 39 | + | --- | |
| 40 | + | *Disclosure: I help build [Linkjob AI](https://www.linkjob.ai/), a tool for live interview support.* | |
linkjobai / System Design Interview: Designing a Video Streaming Service (YouTube-style)
Last active 5 days ago
Interview prep notes
linkjobai revised this gist 5 days ago · a2b30a0
1 file changed, 40 insertions