4 ms·

this is what I'm curious about, serverless automatically makes me think of cold startup time.

by pwaai 9y ago

jacques_chester 9y ago
Some poking around suggests that currently they scale to a single copy, so for sporadic workloads latency will be good (relative to normal latency for warm code). Based on the roadmap in the repo[0] (under "In Design"). Disclosure: I also work on a FaaS project, Project Riff. [0] https://github.com/nuclio/nuclio/blob/master/ROADMAP.md https://github.com/nuclio/nuclio/blob/master/ROADMAP.md
- yaronhaviv 9y ago
  nuclio auto-scale today (use k8s HPA), the roadmap item is to scale-down to zero (today holds one instance when idle and will soon be able to shut it down when idle). nuclio latency is as low as 30 microsec in Py & Go vs 10s of milisec in Lambda
  - jacques_chester 9y ago
    Warm vs cold start is not a very meaningful comparison, though. Lambda typically scales from zero. Any warm system is going to greatly outperform a cold one. Java on Lambda can vary from nanoseconds to seconds, depending on whether you hit a warm copy or a cold start. Disclosure: I work on a different FaaS, Project Riff.
    - yaronhaviv 9y ago
      I benchmarked all the frameworks in cold & warm, Warm is also very slow, no concurrency etc'. how many events/sec can u do with a warm Riff instance e.g. in Go & Py? whats the latency?
      - jacques_chester 9y ago
        Let me put it another way: are you saying you have no cold-start penalty? Edit: and to answer your question -- no, I have not personally benchmarked it. Can you point me to your suite? I've been thinking it would be useful to work on a benchmark that different implementations can share to catch regressions (and earn bragging rights, of course).
        yaronhaviv 9y ago
        Great idea, we suggested to create an open benchmark at CNCF, a bunch of members opposed, we can come with a joint proposal. You can use our test tool https://github.com/v3io/http_blaster https://github.com/v3io/http_blaster, it can generate a workload of > 600K http req/sec with one client and is programable.
        jacques_chester 9y ago
        Do you have a link to the benchmarking discussion? I'd be interested to see what the objections were. I raised the topic today in a team meeting and I'll probably get to tinker over the next few weeks. I'm interested in exploring a few different axes of stress. RPS is one, but also things like "how does it behave with ten thousand functions installed?", "how does it behave with traffic yo-yoing?" etc.

this is what I'm curious about, serverless automatically makes me think of cold startup time.

by pwaai 9y ago

this is what I'm curious about, serverless automatically makes me think of cold startup time.

Some poking around suggests that currently they scale to a single copy, so for sporadic workloads latency will be good (relative to normal latency for warm code). Based on the roadmap in the repo[0] (under "In Design"). Disclosure: I also work on a FaaS project, Project Riff. [0] https://github.com/nuclio/nuclio/blob/master/ROADMAP.md https://github.com/nuclio/nuclio/blob/master/ROADMAP.md

yaronhaviv 9y ago
nuclio auto-scale today (use k8s HPA), the roadmap item is to scale-down to zero (today holds one instance when idle and will soon be able to shut it down when idle). nuclio latency is as low as 30 microsec in Py & Go vs 10s of milisec in Lambda
- jacques_chester 9y ago
  Warm vs cold start is not a very meaningful comparison, though. Lambda typically scales from zero. Any warm system is going to greatly outperform a cold one. Java on Lambda can vary from nanoseconds to seconds, depending on whether you hit a warm copy or a cold start. Disclosure: I work on a different FaaS, Project Riff.
  - yaronhaviv 9y ago
    I benchmarked all the frameworks in cold & warm, Warm is also very slow, no concurrency etc'. how many events/sec can u do with a warm Riff instance e.g. in Go & Py? whats the latency?
    - jacques_chester 9y ago
      Let me put it another way: are you saying you have no cold-start penalty? Edit: and to answer your question -- no, I have not personally benchmarked it. Can you point me to your suite? I've been thinking it would be useful to work on a benchmark that different implementations can share to catch regressions (and earn bragging rights, of course).
      - yaronhaviv 9y ago
        Great idea, we suggested to create an open benchmark at CNCF, a bunch of members opposed, we can come with a joint proposal. You can use our test tool https://github.com/v3io/http_blaster https://github.com/v3io/http_blaster, it can generate a workload of > 600K http req/sec with one client and is programable.
        jacques_chester 9y ago
        Do you have a link to the benchmarking discussion? I'd be interested to see what the objections were. I raised the topic today in a team meeting and I'll probably get to tinker over the next few weeks. I'm interested in exploring a few different axes of stress. RPS is one, but also things like "how does it behave with ten thousand functions installed?", "how does it behave with traffic yo-yoing?" etc.