2nd/10_Wiki/Topics/Architecture/Real-time-Data-Streaming.md

---
id: wiki-2026-0508-real-time-data-streaming
title: Real-time Data Streaming
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Stream Processing, Event Streaming, Real-time Pipelines]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [streaming, kafka, pulsar, flink, data-engineering]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: kafka-flink-pulsar
---

# Real-time Data Streaming

## 매 한 줄
> **"매 batch ETL 의 X — 매 unbounded events 매 milliseconds latency 매 process"**. Kafka (LinkedIn 2010) → Flink / Spark Structured Streaming / Pulsar / Materialize / RisingWave 매 modern stack. 매 2026 매 sub-second analytics 매 default.

## 매 핵심

### 매 layers
- **Ingest**: Kafka, Pulsar, Kinesis, Redpanda — 매 durable log.
- **Process**: Flink, Spark Streaming, Kafka Streams, Bytewax, Arroyo.
- **Serve**: Materialize, RisingWave, Pinot, Druid, ClickHouse.

### 매 windowing
- **Tumbling**: fixed, non-overlapping (1-min buckets).
- **Sliding**: overlapping (5-min, slide 1-min).
- **Session**: gap-based (user activity until 30s inactivity).
- **Hopping**: same as sliding (different name).

### 매 time semantics
- **Event time**: 매 actual occurrence — 매 correctness 위해 default.
- **Processing time**: 매 broker arrival — 매 latency 측정.
- **Ingestion time**: 매 broker append.
- **Watermarks**: 매 lateness threshold — Flink/Beam 핵심.

### 매 응용
1. Fraud detection — 매 payment stream + ML inference.
2. Real-time dashboards — Materialize + Grafana.
3. CDC pipelines — Debezium → Kafka → warehouse.
4. IoT telemetry — MQTT → Kafka → anomaly detection.
5. Personalization — clickstream → feature store → model.

## 💻 패턴

### Kafka Streams (Python via Faust / Bytewax)
```python
import bytewax.operators as op
from bytewax.dataflow import Dataflow
from bytewax.connectors.kafka import KafkaSource, KafkaSink

flow = Dataflow("fraud")
src = op.input("in", flow, KafkaSource(["localhost:9092"], ["payments"]))
parsed = op.map("parse", src, lambda kv: json.loads(kv.value))
flagged = op.filter("flag", parsed, lambda p: p["amount"] > 10000)
op.output("out", flagged, KafkaSink(["localhost:9092"], "alerts"))
```

### Flink SQL (windowed aggregation)
```sql
CREATE TABLE clicks (
  user_id STRING, ts TIMESTAMP(3),
  WATERMARK FOR ts AS ts - INTERVAL '5' SECOND
) WITH ('connector' = 'kafka', ...);

SELECT
  user_id,
  TUMBLE_START(ts, INTERVAL '1' MINUTE) AS w_start,
  COUNT(*) AS clicks
FROM clicks
GROUP BY user_id, TUMBLE(ts, INTERVAL '1' MINUTE);
```

### Materialize (streaming SQL view)
```sql
CREATE SOURCE orders FROM KAFKA BROKER 'kafka:9092' TOPIC 'orders'
  FORMAT AVRO USING SCHEMA REGISTRY 'http://sr:8081';

CREATE MATERIALIZED VIEW revenue_5min AS
SELECT
  date_trunc('minute', ts) AS minute,
  SUM(amount) AS revenue
FROM orders
WHERE ts > now() - INTERVAL '5 minutes'
GROUP BY 1;

-- Subscribe to changes
SUBSCRIBE TO revenue_5min;
```

### Spark Structured Streaming
```python
df = (spark.readStream.format("kafka")
        .option("subscribe", "events")
        .load())
agg = (df.selectExpr("CAST(value AS STRING) as json")
         .select(from_json("json", schema).alias("e"))
         .withWatermark("e.ts", "10 minutes")
         .groupBy(window("e.ts", "1 minute"), "e.user")
         .count())
agg.writeStream.format("delta").outputMode("append").start("/lake/agg")
```

### Pulsar Functions (lightweight processing)
```python
from pulsar import Function

class EnrichOrder(Function):
    def process(self, msg, ctx):
        order = json.loads(msg)
        order["region"] = lookup_region(order["zip"])
        ctx.publish("orders.enriched", json.dumps(order))
```

### Exactly-once with Kafka transactions
```python
producer = KafkaProducer(transactional_id="tx-1", enable_idempotence=True)
producer.init_transactions()
try:
    producer.begin_transaction()
    producer.send("out", value=processed)
    producer.send_offsets_to_transaction(offsets, group_id)
    producer.commit_transaction()
except:
    producer.abort_transaction()
```

### CDC with Debezium
```yaml
# connector config
connector.class: io.debezium.connector.postgresql.PostgresConnector
database.hostname: pg
table.include.list: public.orders
plugin.name: pgoutput
# emits CDC events to Kafka topic "pg.public.orders"
```

## 매 결정 기준
| 상황 | Stack |
|---|---|
| 매 simple transform | Kafka Streams / Bytewax |
| 매 complex windowing, joins | Flink |
| 매 SQL-first analytics | Materialize / RisingWave |
| 매 batch+stream unified | Spark Structured / Beam |
| 매 lightweight, serverless | Pulsar Functions / AWS Lambda |
| 매 OLAP serving | Pinot / Druid / ClickHouse |

**기본값**: 매 2026 매 SQL-on-streams (Materialize/RisingWave) 매 default — DX 압도적.

## 🔗 Graph
- 부모: [[Data Engineering]] · [[Event-Driven Architecture]]
- 변형: [[Stream-Processing-Architectures|Stream Processing]] · [[CEP]]
- 응용: [[CDC]]
- Adjacent: [[Kafka]] · [[Materialize]]

## 🤖 LLM 활용
**언제**: 매 SQL DDL/query generation, 매 schema evolution analysis, 매 anomaly investigation summarization.
**언제 X**: 매 latency-critical hot path — LLM inference 매 too slow. 매 trained ML model 사용.

## ❌ 안티패턴
- **Processing time everywhere**: 매 out-of-order events 매 wrong results — event time + watermarks 사용.
- **Unbounded state**: 매 keyed state 매 grows forever — TTL / windows 필수.
- **Tiny files**: 매 1 record / file → S3 explosion. 매 batching + compaction.
- **Sync external calls in pipeline**: 매 backpressure 폭발. 매 async + bulkhead.
- **No replay strategy**: 매 bad code → poisoned downstream. 매 reset offset + idempotent sinks.

## 🧪 검증 / 중복
- Verified (Akidau — "Streaming 101/102"; Kafka docs; Flink docs 1.18+).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full streaming entry with Materialize/RisingWave |