Redpanda: A new storage engine with Kernel bypass technologies for 10x lower tail latencies
HTML-код
- Опубликовано: 22 сен 2024
- “Redpanda: A new storage engine built from the ground up for Kafka-API compatibility with Kernel bypass technologies for 10x lower tail latencies” by Alexander Gallego
In this talk we’ll cover how we built a new storage engine from scratch using a thread per core architecture for predictable tail latencies. We’ll dive deep into the challenges of adapting a well known protocol (Kafka-API) on top of this new storage engine and the performance gains of optimizing for modern hardware.
Event page: performancesum...
Slack QA channel: performancesum...
this was brilliant. thank you very much. next stop ScyllaDB
he stated that intra-datacenter network latencies nowadays between machines was on par with inter-NUMA latencies. that’s just categorically untrue. 100s of nanoseconds vs 100s of microseconds. that’s 3 orders of magnitude.
i mentioned for a 'contended' resource. I've measured in the milliseconds.
but to add, you *can* in fact do a network call in low double digit microseconds. It's easy to measure, just hook up 2 servers back to back with some SPF link or smth and DPDK and you'll see. All of this is quite easily verifyable really.
Why I never see throughput benchmark for this XD.
"Developers love Kafka api" - are you serious? After a year working with Kafka, I know some configuration choices and design problems I still shudder at
Oh hell yes. Completely agree.
lossy compression here. what ppl want is their existing apps to go faster w/ no code changes. the partitioning scheme of un ordered collection with totally ordered sub collections is pretty handy as a modeling. what you point out is the heavy weight nature of partitions which is true, but the mental model is helpful.
Any tutorial on this?
vectorized.io has an extensive documentation for Redpanda - vectorized.io/docs.
Anybody really cares about latency what to use a RPC and streaming? If I can tolerate 18ms I did not see why I can not tolerate 100ms.
well, spending $1m vs $10m on compute makes a good case for preferring one over the other ;)
What happened to Concord.io?
sold it to akamai.