ScyllaDB vs. Aerospike - A brief comparison
Aerospike’s NoSQL database platform has earned an exemplary reputation for delivering exceptional — and predictable — performance at scale, superior uptime, and strong data consistency at a much lower total cost of ownership (TCO) than other alternatives. We benchmarked our platform against ScyllaDB and revealed that Aerospike provides:
ScyllaDB vs. Aerospike - A brief comparison
Aerospike’s NoSQL database platform has earned an exemplary reputation for delivering exceptional -- and predictable -- performance at scale, superior uptime, and strong data consistency at a much lower total cost of ownership (TCO) than other alternatives. We benchmarked our platform against ScyllaDB, an Apache Cassandra replacement said to deliver “ultra-low latency and extreme throughput.” Our tests revealed that Aerospike performed 7x better at 15% of the cost.
Benchmark summary
We tested Aerospike 4.7.0.5 Community Edition and ScyllaDB Open Source version 3.1.2 using the Yahoo Cloud Serving Benchmark (YCSB) with a 50/50 mix of reads and writes on 4 TB of unique data consisting of 2 billion objects sized at 2K each. With a replication factor of 2, the total user data was 8 TB. Our environment included three Dell® R730xd rack-mounted servers, each with dual sockets and a total of 256 GB of DRAM. Each server utilized four 3.2 TB Micron 9300 Max NVMe SSDs for data and ran Centos Linux version 7.6. We tuned both DBMSs appropriately.
Table 1 summarizes the results when both systems were under maximum load. Aerospike demonstrated extraordinary speed at scale, providing 7.17x better throughput and 99th percentile read/write latencies that were 3.2x lower than ScyllaDB.
We also compared throughput and latencies when each system was operating a more typical production load of 70% of maximum. Again, as Table 2 shows, Aerospike delivered more than 7.2x the throughput with 2.9x lower data access latencies at the 99th percentile than ScyllaDB.
But raw performance is only one consideration. TCO is another. We calculated the TCO of running Aerospike and ScyllaDB for 3 years with a database of 14 billion keys (28 TB of unique user data) and 20% annual data growth rate. We considered equipment, utilities, space, and labor costs, assuming an annual administrator’s salary of $85,706.
The result? As shown in Table 3, Aerospike was 7x cheaper than ScyllaDB. ScyllaDB’s total cost for 3 years was just over $4.6 million. Aerospike’s cost was less than $650,000 -- a savings of more than 85%.
Let’s turn now to the detailed results of the YCSB benchmark runs.
Load results (Insert rate)
To populate our databases, we used a single instance of YCSB with 250 threads. As Table 4 shows, Aerospike loaded data 7.6x quicker than ScyllaDB. We took this measurement during YCSB’s load phase, which is not part of the workload phase and simulates a 100% write workload. Unlike Aerospike, ScyllaDB periodically performs compactions. When such operations are taken into account, Aerospike performed 7.9x better.
Read results (Throughput and latency)
Aerospike serviced reads more quickly and with greater predictability than ScyllaDB regardless of whether both systems were running at maximum load or 70% of capacity. Let’s examine throughput results first.
Aerospike provided 7.17x better throughput at maximum and 7.2x better at 70% capacity, as shown in Figs. 1 and 2. Aerospike’s maximum read throughput was 753,000 OPS with a 99th percentile read latency of 6.5 ms compared to ScyllaDB’s 104,950 OPS with a 99th percentile read latency of 21ms. At 70% load, Aerospike serviced 527,100 reads per second with a 99th percentile 1.2 ms latency compared to ScyllaDB’s 73,465 reads with a 99th percentile 3.5 ms latency.
In addition to throughput, predictable read latency is critical in operational scenarios. Figs. 3 and 4 explore how the two systems fared when we measured latencies at the 99th percentile for reads. Not only did Aerospike deliver 3.2x lower read latencies at maximum capacity and 2.9x lower read latencies at 70% capacity, it did so with far fewer variances in response times. Simply put, Aerospike’s performance was much more predictable -- an important consideration for mission-critical applications.
Update results (Throughput and latency)
Aerospike exhibited significant performance advantages over ScyllaDB for updates as well. Indeed, Aerospike delivered faster throughput with lower (and more predictable) latencies than ScyllaDB when both systems were under maximum load as well as when both systems were operating at 70% capacity.
As Fig. 5 indicates, Aerospike’s maximum update throughput was 753,000 OPS with a 4.5 ms 99th percentile latency compared to ScyllaDB’s 104,950 OPS with a 99th percentile latency of 14.3 ms. Fig. 6 depicts results when operating at 70% capacity; Aerospike serviced 527,100 writes per second with a 1.5 ms 99th percentile latency compared to ScyllaDB’s 73,465 writes with a 4.6 ms 99th percentile latency.
In addition to better update throughput, Aerospike delivered lower and more predictable update latencies than ScyllaDB, as shown in Figs. 7 and 8. We measured latencies at the 99th percentile for updates and found that Aerospike delivered 3.2x lower update latency at maximum capacity and 2.9x lower update latency at 70% capacity. Furthermore, in both situations, Aerospike’s latencies were far more uniformly low than ScyllaDB’s.
What does this mean for you?
Our comparative benchmark with ScyllaDB shows Aerospike’s clear performance advantages: 7.17x greater throughput, 3.2x lower read/write latencies, and 7.9x greater insert throughput when running at maximum load. Aerospike is predictably performant and more easily scalable than ScyllaDB at 7x lower TCO. Aerospike requires fewer clusters, and these clusters can process more transactions at lower latencies. That reduces the complexity and costs of supporting your applications.
About Aerospike
Aerospike is the global leader in next generation, real-time NoSQL data solutions at scale. Its patented Hybrid Memory Architecture™ unlocks the full potential of modern hardware to deliver incredible value from vast amounts of data, powering critical applications for Airtel, Baidu, Banca d’Italia, Nielsen, PayPal, Snap, Verizon Media, Wayfair, and many other firms.
Get started with Aerospike
For high-performance, scalable data management and ultra-low latency, ideal for handling massive datasets and real-time applications.
