Kafka Memory Management¶

Memory architecture and optimization for Apache Kafka brokers and clients.

Memory Architecture Overview¶

JVM Heap¶

Heap Components¶

Component	Description	Memory Impact
Request buffers	Incoming/outgoing request data	Proportional to connections
Metadata cache	Topic/partition metadata	Proportional to partitions
Index structures	In-memory index pointers	Proportional to partitions
Producer state	Idempotent producer tracking	Proportional to producers × partitions
Group coordinator	Consumer group state	Proportional to groups/members

Heap Footprint Estimates (Repository Guidance)¶

Area	Rule of Thumb
Broker heap per partition replica	~1-2 MB
Controller metadata heap	~5 GB for typical clusters

Heap Sizing (Repository Guidance)¶

Cluster Size	Partitions	Heap Size
Small	< 1,000	4-6 GB
Medium	1,000-10,000	6-8 GB
Large	10,000-50,000	8-12 GB
Very Large	> 50,000	12-16 GB

JVM Configuration¶

# Example JVM settings (tune per workload)
export KAFKA_HEAP_OPTS="-Xms6g -Xmx6g"

# GC settings (example values)
export KAFKA_JVM_PERFORMANCE_OPTS="-server \
  -XX:+UseG1GC \
  -XX:MaxGCPauseMillis=20 \
  -XX:InitiatingHeapOccupancyPercent=35 \
  -XX:+ExplicitGCInvokesConcurrent \
  -XX:G1HeapRegionSize=16M \
  -XX:MetaspaceSize=96m \
  -XX:MinMetaspaceFreeRatio=50 \
  -XX:MaxMetaspaceFreeRatio=80"

OS Page Cache¶

Page Cache Role¶

Kafka relies heavily on the OS page cache for performance. The page cache stores recently accessed disk data in RAM.

Page Cache Sizing¶

Rule of thumb: Reserve at least as much RAM for page cache as data you want to keep "hot" (typically last few hours of data).

Page Cache = Total RAM - JVM Heap - OS Overhead

Example:
  Total RAM: 64 GB
  JVM Heap: 6 GB
  OS/Other: 2 GB
  Page Cache: ~56 GB available

Monitoring Page Cache¶

# Check memory usage
free -g

# Check page cache usage
cat /proc/meminfo | grep -E "Cached|Buffers|MemFree|MemTotal"

# Monitor disk I/O (high I/O = cache misses)
iostat -x 1

Buffer Pools¶

Network Buffer Pool¶

Kafka uses buffer pools to reduce garbage collection overhead.

# Broker network buffer sizing
socket.send.buffer.bytes=102400
socket.receive.buffer.bytes=102400
socket.request.max.bytes=104857600

Default Memory/Buffer Limits (Kafka Defaults)¶

Component	Setting	Default
Broker	`socket.send.buffer.bytes`	102400
Broker	`socket.receive.buffer.bytes`	102400
Broker	`socket.request.max.bytes`	104857600
Producer	`buffer.memory`	33554432
Producer	`batch.size`	16384
Producer	`linger.ms`	0
Producer	`max.block.ms`	60000
Consumer	`fetch.min.bytes`	1
Consumer	`fetch.max.bytes`	52428800
Consumer	`fetch.max.wait.ms`	500
Consumer	`max.partition.fetch.bytes`	1048576
Topic	`index.interval.bytes`	4096

Producer Buffer Pool¶

# Producer buffer configuration (example tuning)
buffer.memory=33554432           # 32MB total buffer pool
batch.size=16384                 # 16KB per batch
linger.ms=5                      # Wait time for batching

Consumer Buffer Pool¶

# Consumer fetch sizing
fetch.min.bytes=1                # Minimum bytes to fetch
fetch.max.bytes=52428800         # Maximum per fetch (50MB)
max.partition.fetch.bytes=1048576  # Per partition (1MB)

Memory Pressure Scenarios¶

Heap Exhaustion¶

Symptoms: - OutOfMemoryError - GC taking > 10% of time - Request latency spikes

Causes:

Cause	Solution
Too many partitions	Reduce partitions or increase heap
Large metadata cache	Reduce topic count
Producer state buildup	Reduce idempotent producers
Memory leak	Update Kafka version

Page Cache Exhaustion¶

Symptoms: - High disk read I/O - Consumer latency increases - await time in iostat high

Causes:

Cause	Solution
Heap too large	Reduce heap, leave more for cache
Too much data	Add more brokers
Random access patterns	Improve consumer patterns

Garbage Collection¶

G1GC Tuning¶

# Recommended G1GC settings
-XX:+UseG1GC
-XX:MaxGCPauseMillis=20
-XX:InitiatingHeapOccupancyPercent=35
-XX:G1HeapRegionSize=16M

Parameter	Purpose
`MaxGCPauseMillis`	Target pause time (20ms recommended)
`InitiatingHeapOccupancyPercent`	When to start concurrent GC
`G1HeapRegionSize`	Region size (16M for larger heaps)

GC Monitoring¶

# Enable GC logging
-Xlog:gc*:file=/var/log/kafka/gc.log:time,tags:filecount=10,filesize=100M

# Monitor GC
jstat -gc <pid> 1000

# Analyze GC log
# Look for: pause times, frequency, throughput

GC Monitoring Targets (Repository Guidance)¶

Use GC pause time, frequency, and throughput as trend indicators rather than fixed SLAs.

Direct Memory¶

Off-Heap Buffers¶

Kafka uses direct memory for network I/O operations.

# Configure direct memory limit (example)
-XX:MaxDirectMemorySize=2g

Memory-Mapped Files¶

Index files use memory-mapped I/O:

# These files are memory-mapped
# .index - offset index
# .timeindex - timestamp index

Client Memory Management¶

Producer Memory¶

# Total memory for buffering
buffer.memory=33554432

# Memory allocation behavior
max.block.ms=60000  # Block when buffer full

Memory calculation:

Required memory = buffer.memory +
                  (partitions × batch.size overhead) +
                  compression buffers

Consumer Memory¶

# Fetch sizing
fetch.max.bytes=52428800
max.poll.records=500