Originally from the User Slack
@Antony_Mithun: Hi everyone, we’re evaluating databases for a low-latency use case and considering ScyllaDB vs. Aerospike.
We want to understand ScyllaDB’s expected read/write latency (P50/P99) without having to conduct benchmarks ourselves.
The docs don’t specify an SLA for latency, and I couldn’t find clear numbers in this ScyllaDB vs. Aerospike doc.
Meanwhile, Aerospike claims different results in their benchmark.
Can someone clarify what kind of latency we can expect under high-throughput workloads? Does ScyllaDB have an official SLA for latency?
@dor: Hi Antony, latency depends on a ton of factors - your hardware, workload, schema, ram, .. In general, you can expect very low single digit p99 latency from Scylla. P50 below 1ms easy. Our latency is general on par with Aerospike, their might even be better for reads. However our throughput is better and so is the data model/featureset and elasticity and scalability
@Robert: hi, the latency more depends on Your infrastructure, invest money, use case and it requirements etc., than on DB engine technology. For example network latency between nodes and ųService-nodes - which also depends on level of consistency. If Your application can enable any flag like SEP?
Infra part - if You contain RAM to handle all data stored on disk in scylla case You can avoid IOPS - all data will be cached. But it cost a lot, maybe You want to have some ratio 1:5 (ram/disk)?
Imo it’s a long story to just simple answer, but for my perspective on bare metal which has 1:2 ratio ram/disk I’m having <1ms p99 with 100k req/sec.
@Antony_Mithun: Thank you for your quick and very helpful feedback. We currently use Redis, which provides sub-millisecond latency for most requests (I can get back to you with the exact p99 and p50 values and request counts). However, with our new features, we expect to store significantly more data, which could impact the current Redis performance (especially with AOF). Even 5ms of latency would be considered too high for us. Therefore, we were exploring options that could offer low latency with a hybrid disk-backed setup. Initially, we considered Bigtable but were concerned that it might not perform well in terms of latency.
As you guys pointed out, it ultimately depends on the infrastructure setup, so it seems we’ll need to experiment and determine the best solution ourselves. Thanks again for your input!
@dor: AOF is indeed awful. There shouldn’t be a problem to go below 5ms p99