Good morning #KubeCon! We are excited to meet all of you and talk about performance! 🚀 Stop by our booth at 1360 next to the Demo Theatre to learn more about CPU, GPU, and memory profiling in production! 🧑‍💻

Is there some global load-balancing product out there that's less expensive than Google Cloud's load balancer? ($0.008/GiB)

You thought docker is causing a lot of wasted space on your disk? Try working on a moderately sized Rust codebase that produces 10 binaries 🫠

Never thought I'd say this, but for the first time in a decade, I'm working on a database that can just be queried with SQL, and it's kind of nice 😅

[email protected] has really outdone himself with the animated parts of this blog post!

I was disappointed with the Unifi lineup from Ubiquiti for some time, but the new products are looking pretty awesome! (APs, 10GbE fiber gateway, 10GbE switches with PoE++) Mikrotik is lagging behind for the first time in some time for me!

1% across the board and significantly more for some workloads! Mad props to Canonical for doing the non flashy but impactful important work that only few companies on the planet can! discourse.ubuntu.com/t/introducin...

We made a bit of a dumb thing early on at Polar Signals where we accidentally created a critical GCS bucket as multi-region. But today I learned about Bucket relocation (launched in 2024). Very timely! docs.cloud.google.com/storage/docs...

I noticed that cross-zone traffic has once again crept up on us at Polar Signals, so I dusted off kubezonnet, and implemented a few new features/bug fixes as I was troubleshooting. I present v0.2.0, which we just used to cut 62% of cross-zone traffic. Real money saved. github.com/polarsignals...

We're on the HN front page with our post on the NVIDIA GPU profiler we created. Please help us spread the word and comment if you have any questions!

It's absolutely crazy what computers are capable of when we produce software that works with hardware in mind.

The NVIDIA RTX Pro 4000 Blackwell with 24GB is looking quite interesting as a workstation GPU as an alternative to a Spark DGX. Obviously not quite apples to apples with different memory speed and sizes but still, 24gb is large enough to do interesting things with.

I had a great time talking about GPU profiling at Ubuntu Summit! Thanks everyone for joining!

"We're seeing more Rust recently", "Although 95 % of CUDA workflows use Python" @brancz.com from @polarsignals.com at #UbuntuSummit2510 #UbuntuSummit

If you're at #UbuntuSummit Thursday and Friday, come hear me speak about how we created the world's first open-source low-overhead always-on #NVIDIA #CUDA Profiler! discourse.ubuntu.com/t/infrastruc... #performance #GPU

I could not be more excited about this. To our knowledge, the world's first open-source low-overhead always-on GPU profiler! As always, zero-instrumentation and powered by #eBPF!

Quick one-hour weekend experiment, and I'm already visualizing our CI infrastructure at Polar Signals as distributed traces in Jaeger! Already gave me insight that I didn't know before (our Rust tests are the critical path, not release binary build times)!

It's been that kind of a day ...

I don't know why I have to rediscover this over and over, but reducing CI time is such an incredible productivity booster; every second shaved off is worth the investment!

When did we go from "computers are not great at making use of all those cores" to "a lot of software scales linearly with the number of cores you give it". Not complaining, just feels like yesterday that we said the former statement a lot.

I have several services that use as much network bandwidth as is available to them on Kubernetes. But I need to prioritize or at least guarantee some amount of bandwidth to a certain component. What are my options other than separate node pools?

Took advantage of the lunch break to replay this great talk by Alfonso Subiotto (@asubiotto.com) making the case for a State Machines based approach to Deterministic Simulation Testing, being used at @polarsignals.com, very interesting! Their inspiration for this: sled.rs/simulation

Figure out when the deploy happened 👀

So cool and parca.dev getting mentioned out of nowhere. Love to see it!

Google Mail moved the mark as unread button in the dropdown, and I hate it.

It's awesome to see the arrow project get support like this! medium.com/@QuantStack/...

Despite the ecosystem fragmentation, I'm excited to see more innovation in this space. For now we've firmly settled on Vortex, but happy to see that there are ideas being exchanged!

I have two workloads on the same Kubernetes node, and they are using as much CPU as is available to them on the system, but one workload is a background/batch task whereas the other is serving traffic. What’s the best way to strictly prioritize the serving process in Linux/Kubernetes?

Sure, there can be leaps that I don't see coming, but at the moment we've thoroughly come to the point where claude code/cursor works really well when we understand and describe the problem well, but leaving the thinking to it, completely and utterly fails. Does anyone else see it this way?

We open sourced a read-through cache for S3, called Cachey github.com/s2-streamsto...

I was at a complete loss for words when the message reached me this morning. Cherish the time you have with your loved ones. Han, my friend, you will be dearly missed. I will never forget the memories we’ve built together. github.com/cncf/memoria...

We're excited to be at @rustconf.com 2025! Let's talk about profiling Rust with zero compromise! #rustconf25

We recently made some performance improvements to our Flame Graph by implementing backend pre-computation and viewport culling. Read more in our latest blog post: www.polarsignals.com/blog/posts/2...

Another step closer to making our next-gen storage engine generally available to our customers. Today's milestone: It's now on par with our old storage, while we still have tons of low-hanging and well-understood fruit!

What do you think of our (possibly controversial) take?

Do I know anyone who has an AMD 9955HX? I’d love to know how much faster it is at compiling rust than my M1 Max.

When you think you're just gonna fix a small bug and it turns out to take a week to understand and fix, but then you can't even ship it right away because it requires a data migration, which takes <retention period> of time. 🫠

Ok, I have a rant I have to let go of. If you generate a change to an open-source project fully with AI, didn't read, review, understand, and questioned it, then at least have the decency to say this on the PR description. You're stealing people's time by making them review it for you.

Kyaml is literally valid jsonnet. Can’t make this stuff up, y’all have had this (and more) since 2016 🤣 thenewstack.io/kubernetes-i...

Such an awesome feature! Catch a heap profile just before your Go process is just about to be killed by the OOM killer. (truly hooking into the kernel's OOMkiller with eBPF!)

Is there a good AI product that can learn from all slack-connect channels, discord and docs that I can then use to pre-prepare messages for customer support? I don't want the support itself to be autonomous, I want a human to approve or have the ability edit the message before the customer sees it.

JS isn’t slow, your code is. the way we disregard performance and exepect to be fast is one of our biggest problems.

We're excited to share that the Parca Agent now works offline! Store CPU profiles locally when disconnected, upload when ready. Because not all computing happens in data centers. Read how we built it: www.polarsignals.com/blog/posts/2...

What do folks use for Kubernetes configuration these days? (I do not accept Helm as the answer!)

Vortex is incredibly exciting, we’ve already made plans months ago to make it the primary format we will be storing our data in on object storage! www.linuxfoundation.org/press/lf-ai-...

New storage format RFC in the works 👀 Early preview:

A much requested feature finally shipping! Now you can finally understand why your process is taken off the CPU, whether for timing reasons, for I/O reasons, or any of the many other possibilities!

AI has gotten so good at coding it's frustrating how bad it is at novel, rare, low-level, or hard problems. Some days I'm flying with it, other days it's just complete and utter trash.

We make use of reproducibility properties a lot at Polar Signals. Here's how we do it for our next.js frontend(s).

Yeah, trust me, you want #continuousprofiling. 50% CPU improvement with a ~5-line code change. I often hear people say "we're not at that level of optimizing". I've not found one company this was true for, and you're missing out on opportunities like these.