Cassandra Performance Benchmarking

Measure and validate Cassandra cluster performance.

cassandra-stress

The official benchmarking tool included with Cassandra.

Basic Usage

# Write test (1M operations)
cassandra-stress write n=1000000 -rate threads=50

# Read test
cassandra-stress read n=1000000 -rate threads=50

# Mixed workload (50% read, 50% write)
cassandra-stress mixed ratio\(write=1,read=1\) n=1000000 -rate threads=50

Connection Options

cassandra-stress write n=1000000 \
    -node 192.168.1.10,192.168.1.11,192.168.1.12 \
    -rate threads=100 \
    -mode native cql3 \
    -transport "truststore=/path/to/truststore.jks truststore-password=pass" \
    -log file=results.log

Custom Schema

# Create YAML profile
cat > user_profile.yaml << 'EOF'
keyspace: test_ks
keyspace_definition: |
  CREATE KEYSPACE test_ks WITH replication = {'class': 'NetworkTopologyStrategy', 'dc1': 3};

table: users
table_definition: |
  CREATE TABLE users (
    user_id uuid PRIMARY KEY,
    name text,
    email text,
    created_at timestamp
  )

columnspec:
  - name: user_id
    size: fixed(36)
  - name: name
    size: gaussian(10..50)
  - name: email
    size: gaussian(20..50)
  - name: created_at

insert:
  partitions: fixed(1)
  batchtype: UNLOGGED

queries:
  read_by_id:
    cql: SELECT * FROM users WHERE user_id = ?
    fields: samerow
EOF

# Run with profile
cassandra-stress user profile=user_profile.yaml ops\(insert=3,read_by_id=7\) n=1000000

Key Metrics to Measure

Throughput

Op rate       : Operations per second
Partition rate: Partitions written/read per second
Row rate      : Rows affected per second

Latency

Latency mean  : Average latency
Latency median: 50th percentile
Latency 95th  : 95th percentile (target < 50ms)
Latency 99th  : 99th percentile
Latency max   : Maximum observed latency

Benchmark Scenarios

Write Performance

# Sequential writes
cassandra-stress write n=5000000 cl=LOCAL_QUORUM \
    -rate threads=200 \
    -node node1,node2,node3

# Expected: 20,000-100,000 ops/sec per node (varies by hardware)

Read Performance

# Random reads
cassandra-stress read n=5000000 cl=LOCAL_QUORUM \
    -rate threads=200 \
    -node node1,node2,node3

# Expected: 10,000-50,000 ops/sec per node

Mixed Workload

# 70% read, 30% write (typical application)
cassandra-stress mixed ratio\(read=7,write=3\) n=5000000 cl=LOCAL_QUORUM \
    -rate threads=200

Counter Workload

cassandra-stress counter_write n=1000000 \
    -rate threads=50

YCSB (Yahoo Cloud Serving Benchmark)

Alternative benchmark suite for database comparison.

Set up

# Clone YCSB
git clone https://github.com/brianfrankcooper/YCSB.git
cd YCSB

# Build Cassandra binding
mvn -pl site.ycsb:cassandra-binding -am clean package

Run Workloads

# Load data
./bin/ycsb load cassandra-cql -s -P workloads/workloada \
    -p hosts="node1,node2,node3" \
    -p recordcount=1000000

# Run workload A (50% read, 50% update)
./bin/ycsb run cassandra-cql -s -P workloads/workloada \
    -p hosts="node1,node2,node3" \
    -p operationcount=1000000

YCSB Workloads

Workload	Read	Update	Insert	Scan	RMW
A	50%	50%	-	-	-
B	95%	5%	-	-	-
C	100%	-	-	-	-
D	95%	-	5%	-	-
E	-	-	5%	95%	-
F	50%	-	-	-	50%

Interpreting Results

Good Performance Indicators

✓ Op rate consistent across nodes
✓ p99 latency < 100ms
✓ No timeouts or errors
✓ Stable throughput over time
✓ GC pauses < 200ms

Problem Indicators

✗ High latency variance
✗ Increasing latency over time
✗ Error rate > 0.1%
✗ Uneven distribution across nodes
✗ GC pauses > 500ms

Benchmark Best Practices

1. Warm up cluster - Run shorter test first
2. Use realistic data - Match production data patterns
3. Test at scale - Use production-like data volumes
4. Run multiple times - Average results across runs
5. Monitor during tests - Watch GC, I/O, CPU
6. Test failure scenarios - One node down, network partition

Sample Results Interpretation

Results:
Op rate                   :   45,231 op/s
Partition rate            :   45,231 pk/s
Row rate                  :   45,231 row/s
Latency mean              :    4.4 ms
Latency median            :    2.1 ms
Latency 95th percentile   :   12.3 ms
Latency 99th percentile   :   35.2 ms
Latency 99.9th percentile :   89.4 ms
Latency max               :  245.1 ms
Total partitions          : 5,000,000
Total errors              :        0

Analysis:

• 45K ops/sec is good for 3-node cluster
• Median 2.1ms is excellent
• p99 at 35ms is acceptable
• No errors indicates stability

---

Next Steps

• cassandra-stress - Full tool reference
• Query Optimization - Optimize slow queries
• Key Metrics - Production monitoring