nodetool setconcurrency¶

Sets the maximum concurrent operations for read, write, or counter write thread pools.

Synopsis¶

nodetool [connection_options] setconcurrency <type> <value>

Description¶

nodetool setconcurrency modifies the maximum number of concurrent operations that can execute simultaneously for a specific operation type. This controls how Cassandra's thread pools handle incoming requests and directly impacts throughput, latency, and resource utilization.

Understanding Cassandra's Threading Model (SEDA)¶

Cassandra uses a Staged Event-Driven Architecture (SEDA), where different types of operations are handled by dedicated thread pools (stages). Each stage has:

A thread pool with a configurable number of worker threads
A queue for requests waiting to be processed
Metrics for monitoring active, pending, and completed tasks

Concurrency Types¶

Type	Thread Pool	Default	Purpose
`read`	ReadStage	32	Local read operations (single-partition and range queries)
`write`	MutationStage	32	Local write operations (inserts, updates, deletes)
`counter_write`	CounterMutationStage	32	Counter increment/decrement operations

Non-Persistent Setting

This setting is applied at runtime only and does not persist across node restarts. After a restart, concurrency reverts to the settings in cassandra.yaml.

To make changes permanent, update cassandra.yaml:

concurrent_reads: 32
concurrent_writes: 32
concurrent_counter_writes: 32

Arguments¶

Argument	Description
`type`	Concurrency type: `read`, `write`, or `counter_write`
`value`	Maximum concurrent operations (number of threads)

Examples¶

View Current Concurrency¶

nodetool getconcurrency

Set Read Concurrency¶

nodetool setconcurrency read 64

Set Write Concurrency¶

nodetool setconcurrency write 64

Set Counter Write Concurrency¶

nodetool setconcurrency counter_write 32

When to Adjust Concurrency¶

Scenario 1: Read Latency High with Pending Reads¶

Symptoms: - nodetool tpstats shows pending tasks in ReadStage - Read latencies increasing - CPU not fully utilized

Diagnosis:

# Check for pending reads
nodetool tpstats | grep -E "Pool Name|ReadStage"

Example output showing a problem:

Pool Name                    Active   Pending      Completed   Blocked
ReadStage                        32       245        1523456         0

Action: Increase read concurrency:

nodetool setconcurrency read 64

# Verify
nodetool getconcurrency
nodetool tpstats | grep ReadStage

Scenario 2: Write Throughput Limited¶

Symptoms: - Write operations queuing (pending in MutationStage) - Disk I/O not saturated - Application seeing write timeouts

Diagnosis:

nodetool tpstats | grep -E "Pool Name|MutationStage"

Action: Increase write concurrency:

nodetool setconcurrency write 64

Scenario 3: CPU Saturation¶

Symptoms: - CPU at 100% utilization - High context switching - Latencies spiking under load

Diagnosis:

# Check CPU
top -H -p $(pgrep -f CassandraDaemon)

# Check thread activity
nodetool tpstats

Action: Consider reducing concurrency if over-threaded:

# Too many threads can cause contention
nodetool setconcurrency read 24
nodetool setconcurrency write 24

Scenario 4: High-Core-Count Servers¶

Symptoms: - Server has 64+ CPU cores - Default concurrency (32) underutilizes hardware - Throughput plateaus despite available resources

Action: Scale concurrency with core count:

# For a 64-core server
nodetool setconcurrency read 64
nodetool setconcurrency write 64

Scenario 5: Heavy Counter Workload¶

Symptoms: - Counter operations are slow - CounterMutationStage has pending tasks

Diagnosis:

nodetool tpstats | grep -E "Pool Name|CounterMutationStage"

Action:

nodetool setconcurrency counter_write 48

Metrics to Monitor¶

Thread Pool Statistics¶

The primary tool for monitoring concurrency is nodetool tpstats:

nodetool tpstats

Key columns to watch:

Column	Meaning	Healthy Value
Active	Threads currently processing requests	< max concurrency
Pending	Requests waiting in queue	Should be 0 or very low
Completed	Total completed operations	Increasing over time
Blocked	Requests rejected due to full queue	Must be 0

Interpreting tpstats Output¶

Pool Name                    Active   Pending      Completed   Blocked
ReadStage                        32       245        1523456         0
MutationStage                    28         0        2845123         0
CounterMutationStage              2         0          45123         0

Analysis: - ReadStage: Active=32 (at max), Pending=245 (queuing) → Consider increasing read concurrency - MutationStage: Active=28, Pending=0 → Healthy, no changes needed - CounterMutationStage: Active=2, Pending=0 → Healthy, low counter activity

Warning Signs¶

Observation	Problem	Recommendation
Pending > 0 consistently	Thread pool undersized	Increase concurrency
Blocked > 0	Queue overflow, requests dropped	Increase concurrency urgently
Active = max, high latency	May need more threads or disk is bottleneck	Check disk I/O first
Active low, CPU high	Too many context switches	May need to decrease concurrency

Monitoring Script¶

#!/bin/bash
# monitor_concurrency.sh - Watch thread pool health

while true; do
    clear
    echo "=== $(date) ==="
    echo ""
    echo "--- Current Concurrency Settings ---"
    nodetool getconcurrency
    echo ""
    echo "--- Thread Pool Stats ---"
    nodetool tpstats | head -20
    echo ""
    echo "--- CPU Usage ---"
    top -bn1 | head -5
    echo ""
    echo "--- Latencies ---"
    nodetool proxyhistograms | head -15
    sleep 10
done

JMX Metrics¶

For detailed monitoring, key JMX metrics:

Metric Path	Description
`org.apache.cassandra.metrics:type=ThreadPools,path=request,scope=ReadStage,name=ActiveTasks`	Active read threads
`org.apache.cassandra.metrics:type=ThreadPools,path=request,scope=ReadStage,name=PendingTasks`	Queued reads
`org.apache.cassandra.metrics:type=ThreadPools,path=request,scope=MutationStage,name=ActiveTasks`	Active write threads
`org.apache.cassandra.metrics:type=ThreadPools,path=request,scope=MutationStage,name=PendingTasks`	Queued writes

Impact on Cluster and Clients¶

Increasing Concurrency¶

Positive effects: - Higher throughput (more requests processed in parallel) - Lower latencies (less time waiting in queue) - Better utilization of multi-core CPUs

Potential negative effects: - Increased memory usage (more threads = more stack space) - Higher CPU contention if over-subscribed - More pressure on disk I/O - Potential for increased GC pressure

Decreasing Concurrency¶

Positive effects: - Lower memory footprint - Reduced CPU contention - More predictable latencies under overload

Potential negative effects: - Lower throughput - Requests queue up faster - Risk of blocked requests if queue fills

Client Impact¶

Concurrency Change	Client Experience
Too low	Timeouts, slow responses, connection pool exhaustion
Optimal	Consistent low latencies, high throughput
Too high	May see latency spikes if resources over-subscribed

Recommended Values¶

General Guidelines¶

Server Type	CPU Cores	Recommended Read	Recommended Write
Small (4-8 cores)	4-8	16-32	16-32
Medium (16-32 cores)	16-32	32-64	32-64
Large (64+ cores)	64+	64-128	64-128

Storage-Based Considerations¶

Storage Type	Read Concurrency	Write Concurrency	Notes
HDD	16-32	32-64	Reads limited by seek time
SATA SSD	32-64	32-64	Balanced I/O
NVMe SSD	64-128	64-128	Can handle high parallelism

Formula Approach¶

A common starting point:

concurrent_reads = 16 × number_of_drives
concurrent_writes = 8 × number_of_cpu_cores

For example, with 8 cores and 4 SSDs: - Read concurrency: 16 × 4 = 64 - Write concurrency: 8 × 8 = 64

Tuning Process

Start with defaults (32/32/32)
Monitor tpstats for pending tasks
If pending consistently > 0, increase by 50%
Monitor CPU and memory impact
Repeat until balanced

Relationship to Other Settings¶

Native Transport Requests¶

# cassandra.yaml
native_transport_max_threads: 128  # Threads handling client connections

The native transport threads hand off work to the stage pools. If native_transport_max_threads is high but concurrency is low, requests will queue.

Concurrent Compactors¶

concurrent_compactors: 4  # Separate from read/write concurrency

Compaction has its own thread pool and doesn't compete with read/write concurrency settings.

Memory Implications¶

Each thread requires stack space:

Memory per thread ≈ 256KB (default stack size)
64 threads ≈ 16MB stack space
128 threads ≈ 32MB stack space

High concurrency values increase overall heap pressure indirectly.

Cluster-Wide Configuration¶

Apply to All Nodes¶

#!/bin/bash
# set_concurrency_cluster.sh

READ_CONCURRENCY="${1:-32}"
WRITE_CONCURRENCY="${2:-32}"# Get list of node IPs from local nodetool status


nodes=$(nodetool status | grep "^UN" | awk '{print $2}')

echo "Setting concurrency across cluster..."
echo "Read: $READ_CONCURRENCY, Write: $WRITE_CONCURRENCY"
echo ""

for node in $nodes; do
    echo "=== $node ==="
    ssh "$node" "nodetool setconcurrency read $READ_CONCURRENCY"
    ssh "$node" "nodetool setconcurrency write $WRITE_CONCURRENCY"
    ssh "$node" "nodetool getconcurrency"
    echo ""
done

Making Changes Permanent¶

# cassandra.yaml
concurrent_reads: 64
concurrent_writes: 64
concurrent_counter_writes: 32

Troubleshooting¶

Pending Tasks Not Decreasing After Increase¶

# Check if disk is the bottleneck
iostat -x 1 5

# If disk is at 100% util, concurrency won't help
# Consider:
# - Faster storage
# - Better data model (fewer reads/writes)
# - More nodes

High CPU After Increasing Concurrency¶

# Check for excessive context switching
vmstat 1 5

# If 'cs' (context switches) is very high, reduce concurrency
nodetool setconcurrency read 32

Blocked Tasks Appearing¶

# Blocked means queue overflow - serious issue
nodetool tpstats | grep -E "Pool Name|Blocked"

# Immediate actions:
# 1. Increase concurrency
nodetool setconcurrency read 128

# 2. Check for resource bottlenecks
iostat -x 1 3
top -H

# 3. Consider if cluster is undersized

Changes Not Taking Effect¶

# Verify change applied
nodetool getconcurrency

# Check logs for errors
tail -100 /var/log/cassandra/system.log | grep -i concurrency

# Try again
nodetool setconcurrency read 64

Best Practices¶

Concurrency Guidelines

Start conservative - Begin with defaults, increase based on metrics
Monitor continuously - Watch tpstats before and after changes
Balance resources - Don't set concurrency higher than available CPU cores
Consider storage - HDDs need lower concurrency than SSDs
Test under load - Validate changes during realistic traffic
Make permanent - Update cassandra.yaml after validating
Apply cluster-wide - Keep settings consistent across nodes

Common Mistakes

Setting concurrency very high without monitoring impact
Ignoring disk I/O when tuning (disk is often the bottleneck)
Different settings on different nodes (causes imbalanced load)
Forgetting to persist changes (lost on restart)
Not monitoring blocked tasks (indicates dropped requests)

Command	Relationship
getconcurrency	View current concurrency settings
tpstats	Monitor thread pool statistics
proxyhistograms	View read/write latency distributions
info	General node information

nodetool setconcurrency¶

Synopsis¶

Description¶

Understanding Cassandra's Threading Model (SEDA)¶

Concurrency Types¶

Arguments¶

Examples¶

View Current Concurrency¶

Set Read Concurrency¶

Set Write Concurrency¶

Set Counter Write Concurrency¶

When to Adjust Concurrency¶

Scenario 1: Read Latency High with Pending Reads¶

Scenario 2: Write Throughput Limited¶

Scenario 3: CPU Saturation¶

Scenario 4: High-Core-Count Servers¶

Scenario 5: Heavy Counter Workload¶

Metrics to Monitor¶

Thread Pool Statistics¶

Interpreting tpstats Output¶

Warning Signs¶

Monitoring Script¶

JMX Metrics¶

Impact on Cluster and Clients¶

Increasing Concurrency¶

Decreasing Concurrency¶

Client Impact¶

Recommended Values¶

General Guidelines¶

Storage-Based Considerations¶

Formula Approach¶

Relationship to Other Settings¶

Native Transport Requests¶

Concurrent Compactors¶

Memory Implications¶

Cluster-Wide Configuration¶

Apply to All Nodes¶

Making Changes Permanent¶

Troubleshooting¶

Pending Tasks Not Decreasing After Increase¶

High CPU After Increasing Concurrency¶

Blocked Tasks Appearing¶

Changes Not Taking Effect¶

Best Practices¶

Related Commands¶