2nd/10_Wiki/Topics/Architecture/Cosmos_플랫폼_(Netflix).md

---
id: wiki-2026-0508-cosmos-플랫폼-netflix
title: Cosmos 플랫폼 (Netflix)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Netflix Cosmos, Cosmos Platform, Cosmos]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [netflix, media-processing, workflow, microservices]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: java
  framework: spring-boot
---

# Cosmos 플랫폼 (Netflix)

## 매 한 줄
> **"매 Netflix 의 next-gen media computing platform"**. Reloaded (predecessor) 의 후계, 2018-2020 announce. Microservices + workflow orchestration + serverless functions 매 통합. Encoding, transcoding, content analysis, IMF processing 매 모든 media pipeline 의 backbone.

## 매 핵심

### 매 architecture (3 layers)
- **Optimus (API layer)** — external-facing service, business logic, request validation.
- **Plato (Workflow layer)** — Conductor-based workflow orchestration, retry, branching.
- **Stratum (Functions layer)** — serverless compute, Titus (container) 위 실행.

### 매 design 원칙
- **Asynchronous everything** — sync call 매 minimum, message-based.
- **Schema-first** — Protobuf gRPC 매 contract.
- **Workflow as DSL** — JSON workflow definition (Conductor).
- **Function granularity** — small, idempotent, stateless.
- **Event-driven** — Kafka, SQS 매 inter-service.

### 매 응용
1. Video encoding pipeline — source → multiple bitrates × codecs.
2. Image processing — title art, thumbnails, language variants.
3. Subtitle / caption processing.
4. Content quality control (QC) automation.
5. IMF (Interoperable Master Format) ingestion.

## 💻 패턴

### Workflow definition (Plato / Netflix Conductor)
```json
{
  "name": "encode_video_v3",
  "version": 3,
  "tasks": [
    {
      "name": "probe",
      "type": "FUNCTION",
      "function": "stratum:ffprobe",
      "inputParameters": { "url": "${workflow.input.sourceUrl}" }
    },
    {
      "name": "fan_out_encode",
      "type": "FORK_JOIN_DYNAMIC",
      "dynamicForkTasksParam": "encodeTasks"
    },
    {
      "name": "join_results",
      "type": "JOIN"
    },
    {
      "name": "publish",
      "type": "FUNCTION",
      "function": "stratum:publishToCDN"
    }
  ]
}
```

### Stratum function (Java, Spring Boot)
```java
@CosmosFunction(name = "ffprobe")
public class FfprobeFunction implements Function<ProbeRequest, ProbeResponse> {
    @Override
    public ProbeResponse apply(ProbeRequest req) {
        var info = Ffmpeg.probe(req.url());
        return new ProbeResponse(info.duration(), info.codec(), info.bitrate());
    }
}
```

### Optimus service (gRPC)
```protobuf
service EncodeService {
  rpc StartEncode(EncodeRequest) returns (EncodeResponse);
  rpc GetStatus(StatusRequest) returns (stream StatusUpdate);
}
```

### Titus container deployment
```yaml
# titus job definition
applicationName: stratum-ffprobe
container:
  image: registry/stratum/ffprobe:v42
  resources:
    cpu: 4
    memoryMB: 8192
    gpu: 0
  env:
    COSMOS_FUNCTION_NAME: ffprobe
```

### Event-driven trigger (Kafka)
```java
@KafkaListener(topics = "media.uploaded")
public void onUpload(MediaUploadedEvent event) {
    workflowClient.start("encode_video_v3", Map.of(
        "sourceUrl", event.getUrl(),
        "movieId", event.getMovieId()
    ));
}
```

### Idempotency key (function level)
```java
@CosmosFunction(idempotencyKey = "${input.movieId}-${input.bitrate}")
public class EncodeFunction implements Function<EncodeReq, EncodeResp> { ... }
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Long-running multi-step media job | Plato workflow |
| Simple stateless transform | Stratum function |
| External-facing API | Optimus service |
| Realtime / sub-100ms | NOT Cosmos (sync RPC service 따로) |

**기본값**: workflow for any multi-step media pipeline; functions for individual steps.

## 🔗 Graph
- 부모: [[Microservices]]
- 변형: [[Temporal]]
- Adjacent: [[Spinnaker]]

## 🤖 LLM 활용
**언제**: workflow definition 생성, function template 생성, 운영 incident 매 root cause 분석.
**언제 X**: production workflow execution (deterministic 해야 함).

## ❌ 안티패턴
- **Stateful function**: scale 못 함, retry 깨짐.
- **Long sync chain in Optimus**: workflow 가야 할 것 매 sync call 로 묶음.
- **No idempotency**: retry 매 duplicate side-effect.
- **Workflow too granular**: orchestration overhead > work.

## 🧪 검증 / 중복
- Verified (Netflix Tech Blog 2020 *Rebuilding Netflix Video Processing Pipeline with Microservices*).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full content covering Optimus/Plato/Stratum |