Lucian Maly

Hi Mark,

In the context of generative AI, how can PyTorch be leveraged to effectively address issues like bias and fairness in the generated outputs? I have not seen this subject in your book's TOC.

Thanks

I would use Pytorch for

Research

Rapid prototyping

Projects that require flexibility and ease of use

And Tensorflow for

Large-scale deployments

Production environments

Projects requiring high performance and scalability

It makes sense to use Pytorch for these smaller projects in the book...

The branch of AI that focuses on creating entirely new content is rapidly evolving (e.g. every version of Midjourney is exponentially better than previous).
What are your thoughts on the future of generative AI and its potential impact on society? I already see some impact on the Art industry...

In the book, you mentioned building a music generator using GANs and Transformers. Can you elaborate on the specific techniques used in this project, please?

Hi Mark, welcome to the Code/Java Ranch. Glad you are here :-)

Hi Wouter,


Bit of a philosophical question for you - How can technical knowledge be used to foster creativity in software engineering?


Thank you for your time.

Many thanks, I am looking forward to this. Thank you Shai for hanging out, it was great!

Hi Shai,

What strategies do you use for debugging in a Kubernetes environment and how do you ensure that debugging does not negatively impact the performance and availability of your production clusters?

Many thanks

Hi Shai,

Do you have much experience with Prometheus? Based on TOC, there is a chapter about Observability and also Kubernetes, so I thought I'd ask. Sometimes I have trouble understanding what metric attributes (counter vs gauge vs histogram vs summary) do I need for my troubleshooting.

For example if I want to know "how many requests took somewhere between 0.08s and 0.1s on node nodeX", it is not immediately obvious if i should use e.g. request_latency_seconds_bucket{le="0.1"} or quantiles?

Is there a golden rule for troubleshooting when using Prometheus metrics?

Best,

Hi Shai,

Would you mind sharing some of the auxiliary tools you use for debugging and which ones are good for what? I typically use strace ltrace dtrace

Best,

Welcome, Shai Almog! Enjoy your stay :-)

Hi Jamie,

Thank you for your reply. We have a local cluster that gathers absolutely all telemetry (we do not filter out anything!) and then we alert only on few important things - this took a while to get right. I like your point of view from the SRE perspective, we tend to ignore that sometimes.

Just FYI community seems to be recognizing 3 different types of "important" metrics and that matches wit what you are saying:

1. The USE (Utilization / Saturation / Errors): great for low-level metrics and often used in the context of performance engineering and root cause analysis = SDLC school

2. The RED (Rate / Errors / Duration): focus on the number of requests served, the number of failed requests, and how long requests take; useful for many application-level cases = SRE school

3. USE+RED combined - latency, traffic, errors, and saturation etc.

Cheers

Hi Jamie,

Metrics-style telemetry is about using numbers - e.g. value of rates, timers and so on, to get the feedback about what’s going on in your system. But How do you know what metrics should you be looking for (for example in Kubernetes, there are hundreds of different metrics). What's your golden rule?

Many thanks for hanging out on CodeRanch.

Welcome, Jamie!

Fantastic, thank you for promoting me to Ranch Foreman