Robson Dobzinski’s Post

Day 14: Continuing the CPU tracker scenario using Syftbox, today's lesson was learning more about the aggregator API. The Aggregator API aggregates CPU metrics from various distributed clients while ensuring privacy through differential privacy techniques. The API's workflow consists of: Discovery: Identifies and validates active participants. Data Collection: Collects protected CPU metrics, ensuring they are up-to-date. Aggregating & Visualization: Computes statistics and visualizes them on a dashboard The aggregator preserves privacy through: Exclusive use of differentially private metrics Time-limited data retention Aggregation of recent measurements only No exposure of individual peer data Implementation can be found on my GitHub: https://lnkd.in/gAVE8xmC #30daysofflcode

We are very excited to announce the release of deployKF 0.1.5 🚀 🚀 🚀 ! ➤ https://lnkd.in/gB8nT4TV Don’t let the patch version fool you, this is a significant (but backward-compatible) update for all users. Here are the headline features: ➤ Kubeflow Pipelines V2 support (with full backwards compatibility for V1 pipelines) ➤ Updated default Kubeflow Notebooks images (much newer versions of everything like TensorFlow and PyTorch) ➤ Improved Dashboard UI ➤ Significant security and bug fixes

On average 1 in 2 findings in your container image comes from libraries in your code, contextual reachability should help, but is not there...or is it? Something exciting coming up! In the meantime, Brian, could you please fix the 300 critical container image findings? Thanks! #aspm #containersec #appsec

A super simple packet sniffing script leveraging Scapy and Cloudshark.

Last two weeks I've been optimizing our code base. Came up with this as a side product: https://lnkd.in/dU7FGhy2 The long story: When working with heterogeneous compute, most profilers don't do the job at finding the hotpots in your code. Actually, there tend to be "cold spots" unless you can keep all devices busy crunching numbers. This is a tricky equation to balance, if you have a typical CPU + GPU setup, the main job of the CPU will most likely be to keep the GPU busy + maintaining various "book keeping" tasks. However, the CPU and GPU need to be synchronized from time to time, and if the GPU is not kept busy, this can easily become a cold spot where the CPU code just blocks on a fence waiting for the GPU finish, instead of providing more data for it to munch upon. The problem with most profilers is that they do not measure time spent in I/O, mutex locks, and other sorts of inter-device communication. The simplest, and admittedly ugly strategy, is to use the debugger and randomly press Ctrl-C while running the code, inspect the stack trace, and continue. Doing this a couple of times can quickly show your "extreme spots". Traditionally this is known as the "poor mans debugger", or PMP for short. The repository linked above takes inspiration from this approach and automates it some scripts, such that it actually becomes a descent profiler, which I call p4. This has been extremely valuable to profile our code. By consistently applying p4 I have managed to refactor a CPU intensive algorithm with several fence related performance issues to the GPU and improved performance by more than 10x while doing so. Hope this comes to use for others as well, please let me know if it helps you, and please suggest improvements to the profiler if you have any!

Our 2.3.0 release of flask-caching is out! We can now easily tell whether responses were served from cache by checking response headers (hit indicator). It's a simple change, yet quite useful for debugging, monitoring cache performance and enhancing overall transparency of request-response cycle. New contributors are more than welcome, check https://lnkd.in/dUMnenRx to learn more about the project.

Run Daft Distributed in 60 seconds! In this first part, you’ll see in 30 seconds how to install daft-launcher, configure a cluster and also spin up a cluster. Watch the full video: https://lnkd.in/d_NTftJY Try it yourself: https://lnkd.in/g4tbpWDB

The more you understand the internals of a framework before using it, the more time you'll save when running it in production. Learn the internals by studying the source code.

Say goodbye to unnecessary software in your container images. Discover the benefits of going distroless ✨ 📺 Watch: https://bit.ly/3UoQA3K 📰 Read: https://bit.ly/3Uk3Czj #DockerBlog

#supermaven with VSCode and Cursor with Sonnet has been a great for productivity—calling it a 10x boost might be an understatement. However, the reliance on closed models and the issue of data being sent to third parties has been a major drawback, preventing it from becoming a go-to workflow. But now, with Void(YC24) available under the MIT license, there’s finally seems to be hope. Check it out: https://lnkd.in/gaivvfqF