Chapter 11: Stream Processing

Core Concepts

Batch vs Stream Processing

Messaging Systems

Message Brokers:

• Store messages temporarily
• Producer-consumer pattern
• Examples: RabbitMQ, ActiveMQ

Log-Based Messaging:

• Append-only log
• Kafka, Kinesis
• Durable, replayable

---

Log-Based Messaging

How It Works

1. Producer appends message to log
2. Consumer reads from log
3. Messages persisted until consumed
4. Multiple consumers can read same log

Advantages

• Durability: Messages survive crashes
• Replayability: Re-process old messages
• Scalability: Partition log across machines

Consumer Groups

• Group of consumers share partition
• Each partition consumed by one consumer
• Rebalancing when consumers join/leave

---

Change Data Capture (CDC)

Concept

• Capture changes from database
• Stream changes to other systems
• Keep derived data in sync

Implementations

Trigger-Based:

• Database triggers capture changes
• Write to change log
• Simple but has overhead

Log-Based:

• Read database's write-ahead log
• No overhead on database
• More complex implementation

Log Compaction

• Remove old versions of keys
• Keep only latest value
• Reduce storage, maintain current state

---

Event Sourcing

Concept

• Store all events, not just current state
• Current state derived from event log
• Immutable event history

Commands vs Events

• Command: Request to do something
• Event: Record that something happened

Event Sourcing Benefits

• Complete audit trail
• Time-travel debugging
• Derived views from same events

---

Stream Processing

Stream Tables

• Continuous query on event stream
• Results updated in real-time
• Join stream with table (enrichment)

Stream-Stream Join

• Join two event streams
• Time-windowed (events within time range)
• Complex state management

Window Operations

Tumbling Window: Fixed-size, non-overlapping Hopping Window: Fixed-size, overlapping Sliding Window: Time-based, continuous

Exactly-Once Processing

• Each event processed exactly once
• Challenging in distributed systems
• Checkpointing and transactional processing

---

Stream Analytics

Time-Window Aggregation

• Count events per time window
• Compute averages, sums
• Real-time dashboards

Pattern Detection

• Complex event processing (CEP)
• Detect sequences of events
• Fraud detection, monitoring

---

Fault Tolerance

Checkpointing

• Periodically save state
• Restart from checkpoint on failure
• Used by Flink, Spark Streaming

Microbatching

• Process small batches of events
• Checkpoint between batches
• Used by Spark Streaming

Transactional Writes

• Atomic updates to output
• Exactly-once semantics
• Achieved via transactions

---

Key Takeaways

1. Log-based messaging (Kafka) is the foundation for stream processing
2. Change Data Capture keeps derived data in sync with databases
3. Event Sourcing provides complete history and auditability
4. Windowing handles unbounded streams
5. Exactly-once processing is achievable but complex
6. Stream and batch processing are converging (Lambda/Kappa architectures)