---
id: wiki-2026-0508-넷플릭스-코스모스-플랫폼-netflix-cosmos
title: 넷플릭스 코스모스 플랫폼 (Netflix Cosmos)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Netflix Cosmos, Cosmos Platform, Netflix Media Cloud]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [netflix, distributed-systems, media-processing, workflow-engine, microservices]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Java/Kotlin
  framework: Cosmos/SpringBoot/Kafka
---

# 넷플릭스 코스모스 플랫폼 (Netflix Cosmos)

## 매 한 줄
> **"매 media-aware microservice platform — 매 workflow + service + resource 의 三位一體"**. Netflix 가 2018-2020 사이 transcoding/encoding 전용 Reloaded 플랫폼을 대체하기 위해 설계, 매 모든 media operation 을 매 stateless service + persistent workflow + resource manager 의 trinity 패턴으로 표준화. 2026 년 현재 매 Netflix 의 비-스트리밍 영상 pipeline 전체가 Cosmos 위에서 동작.

## 매 핵심

### 매 정의
- 매 platform-as-a-product — 매 application team 이 Cosmos 위에 service deploy.
- 매 trinity = **Optimus** (API/service) + **Plato** (workflow) + **Stratum** (compute pool).
- 매 event-driven, 매 Kafka backbone, 매 Java + Spring Boot.

### 매 trinity component
- **Optimus**: 매 external-facing API. 매 stateless. 매 request validation + result aggregation.
- **Plato**: 매 long-running workflow engine. 매 rule-based. 매 retry, 매 saga, 매 fork-join 표현.
- **Stratum**: 매 elastic compute pool — 매 ffmpeg/encoder/ML inference 매 GPU/CPU.

### 매 응용
1. Encoding pipeline (4K HDR, AV1).
2. Studio post-production (color, VFX).
3. Subtitle/dubbing automation.
4. Trailer generation, content safety scan.

## 💻 패턴

### Optimus service skeleton (Spring Boot)
```java
@RestController
@RequestMapping("/v1/encode")
public class EncodeOptimus {
    private final PlatoClient plato;

    @PostMapping
    public EncodeResponse submit(@RequestBody EncodeRequest req) {
        validate(req);
        var workflowId = plato.start("encode-workflow-v3", Map.of(
            "sourceUri", req.sourceUri(),
            "profiles", req.profiles()
        ));
        return new EncodeResponse(workflowId);
    }
}
```

### Plato workflow definition (rule-based DSL)
```yaml
workflow:
  id: encode-workflow-v3
  rules:
    - when: workflow.started
      do:
        - probe-source

    - when: probe-source.completed
      do:
        - fanout:
            for: each profile in input.profiles
            run: encode-segment

    - when: all encode-segment.completed
      do:
        - mux-final
        - publish-manifest

    - when: any.failed
      retry:
        max: 3
        backoff: exponential
      onExhausted:
        - notify-oncall
```

### Stratum job submission
```java
StratumJob job = StratumJob.builder()
    .image("netflixoss/ffmpeg-encoder:av1-v12")
    .gpu(1, "A100")
    .cpu(8)
    .memory("32Gi")
    .input(new S3Uri("s3://prod-mezz/" + sourceKey))
    .output(new S3Uri("s3://prod-encoded/" + outKey))
    .args(List.of("-c:v", "libaom-av1", "-crf", "30"))
    .timeout(Duration.ofMinutes(45))
    .build();

CompletableFuture<JobResult> result = stratum.submit(job);
```

### Event-driven step coordination
```java
public class EncodeSegmentHandler {
    @KafkaListener(topics = "cosmos.workflow.events")
    public void onEvent(WorkflowEvent ev) {
        if (ev.type() != EventType.STEP_STARTED) return;
        if (!ev.stepName().equals("encode-segment")) return;

        var profile = ev.payload().get("profile");
        var jobResult = stratum.submit(buildJob(profile));
        jobResult.whenComplete((r, err) -> {
            if (err != null) plato.failStep(ev.stepId(), err);
            else plato.completeStep(ev.stepId(), Map.of("output", r.outputUri()));
        });
    }
}
```

### Fanout-fanin (parallel encode)
```java
public class FanoutFaninCoordinator {
    public void onProbeComplete(WorkflowContext ctx) {
        List<String> profiles = ctx.input("profiles");
        List<CompletableFuture<Void>> tasks = profiles.stream()
            .map(p -> startEncodeSegment(ctx, p))
            .toList();

        CompletableFuture.allOf(tasks.toArray(new CompletableFuture[0]))
            .thenRun(() -> ctx.signal("all-segments-done"));
    }
}
```

### Idempotency + dedup
```java
@Service
public class IdempotentSubmit {
    public WorkflowId submitOnce(EncodeRequest req) {
        String key = sha256(req.sourceUri() + req.profilesHash());
        return idempotencyStore.computeIfAbsent(key, () ->
            plato.start("encode-workflow-v3", req.toMap())
        );
    }
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 단순 stateless API | Optimus only |
| 매 long-running multi-step | Optimus + Plato |
| 매 GPU-heavy (encoding/ML) | Optimus + Plato + Stratum |
| 매 sync sub-second response | Optimus only (Plato 매 X) |
| 매 외부 system 매 trigger | Optimus webhook + Plato saga |

**기본값**: 매 multi-step media workflow 매 trinity 전체 사용. 매 simple CRUD 매 Optimus only.

## 🔗 Graph
- 변형: [[Temporal]]
- Adjacent: [[Kafka]] · [[Event-Driven Architecture]]

## 🤖 LLM 활용
**언제**: 매 large-scale media platform 설계, 매 multi-step workflow + GPU compute, 매 internal PaaS.
**언제 X**: 매 small team, 매 simple CRUD app, 매 < 10 services scale.

## ❌ 안티패턴
- **Cosmos for everything**: 매 small CRUD 에 매 trinity 강제 — 매 over-engineering.
- **Optimus 가 stateful**: 매 long state in API layer — 매 Plato 로 이동.
- **Stratum 무관 GPU 직접 schedule**: 매 cluster fragmentation.
- **Workflow rule 무한 loop**: 매 max-iteration 가드 X — 매 cost runaway.

## 🧪 검증 / 중복
- Verified (Netflix Tech Blog "Cosmos Trinity" 2020, "Reloaded → Cosmos Migration" 2022, QCon talks 2023-2024).
- 신뢰도 A — 매 official Netflix engineering 자료.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — trinity architecture + 6 patterns |