Aaron Stannard

Founder and CEO @Petabridge @getsdkbin @textforgeapp
OSS: @AkkaDotNet and me on YouTube:

[Aaron Stannard @Aaronontheweb]

@mhmd_azeez A little rhetorical flair from me lol

[Aaron Stannard @Aaronontheweb]

A recurring anti-pattern I see some of our customers pursue just to try to save on their cloud infrastructure bills: "auto-scaling is magic"

Take a complex business application and then tell your engineers to use Karpenter / whatever to run it with AWS spot instances, which assumes that you can just migrate loads freely between instances without any penalty

But wait.... 🤔

Oops - this application uses Akka .NET actors to manage long-running, computationally expensive jobs.

Scaling the system on-demand causes that work to get rebalanced to new nodes and none of the jobs were designed to be easily parked / resumed on-demand, because el cheapo auto-scaling wasn't a requirement originally, so the jobs get killed mid-progress and resumed on their new nodes.

So when the system starts actually getting used heavily, the auto-scaling kicks in and _makes the jobs restart and take way longer than they used to_ in the eyes of customers / stakeholders. This is a giant performance degredation!

So the next dumb step we take is trying to stop _current_ jobs from moving and only start new ones on the auto-scaled nodes. Sounds good, right? But _which jobs created the load that required auto-scaling in the first place_? THAT WOULD BE THE CURRENTLY EXECUTING ONES. So what we see is a bunch of new instances get auto-scaled, and they stay idle-ish while the old nodes stay busy! Effectively an auto-scaling stalemate that works only when we get lucky with incoming requests.

So that doesn't work - next up is trying to design for auto-scaling properly: re-entrant jobs, work-pulling instead of push, rebalance notifications, etc. This is an engineering effort that adds significantly more moving parts and infrastructure, requires a much higher standard of testing and validation, and will take a LONG time to develop.

Engineering leadership, again, looks for a short cut: what if we just force the rebalancing to happen immediately instead of gradually, so jobs that will be moved get moved instantly? Well, this causes chaos inside the cluster because everything gets blown up and restarted all at the same time as new nodes come online, rather than gradually + gently metered out. A thundering herd problem.

The morals of this story:

1. Auto-scaling works but only if you design properly for it. Not all workloads behave the same.
2. You'd probably be better off just buying servers or reserving EC2 instances than trying to add all of this complex machinery.

[Aaron Stannard @Aaronontheweb]

@DanielW_Kiwi @zanehengsperger Good question

[Aaron Stannard @Aaronontheweb]

https://t.co/cW1NmQdW8t - I know you're curious!

[Aaron Stannard @Aaronontheweb]

I am so pleased with what Netclaw can render in Slack - this is such an improved viewing experience over TUIs

[Aaron Stannard @Aaronontheweb]

I need to send some upstream patches to SlackNet to get table rendering to work

[Aaron Stannard @Aaronontheweb]

@damianGray Keep some .GGUFs of popular OSS models downloaded because getting that shit banned is where this might go next

[Aaron Stannard @Aaronontheweb]

Between the data center protests and now the USGOV being able to decide which models can / can't be released, AI companies only have themselves to blame.

Fear-based messaging (devs replaced in 6 months, Mythos is a nuclear weapon) was always going to result in backlash long-term

[Aaron Stannard @Aaronontheweb]

@Mike_Preston17 it does - the built-in DuckDuckGo search is flaky due to throttling. I ran with a paid Brave Search API key for a while but setting up a self-hosted SearXNG instance has been the best option for me so far. Netclaw supports all 3, defaults to DDG.

[Aaron Stannard @Aaronontheweb]

@Mike_Preston17 Check out https://t.co/oULkeBKSSl - I've done 1.3-1.4b tokens with it since about April using local models

[Aaron Stannard @Aaronontheweb]

Persuasion and selling people on win-win opportunities was always the better long-term option, but it's not "turbo growth" enough for hucksters or the autists at Anthropic who genuinely believe their own bullshit

[Aaron Stannard @Aaronontheweb]

@colossus_cap @sunny051488 More like a Hobson’s choice when it comes to home ownership

[Aaron Stannard @Aaronontheweb]

You're cute, ChatGPT

[Aaron Stannard @Aaronontheweb]

@IAmAdamReed LOL I’m basically glueing MCP OAuth flow in front of browser auth myself for this pig

[Aaron Stannard @Aaronontheweb]

Good lord, I have to write another MCP server because a service I use hates OAuth2 standard workflows

[Aaron Stannard @Aaronontheweb]

@WescheNex1q what's the template for running it?

[Aaron Stannard @Aaronontheweb]

@petabridge already working on a new beta while we're at it....

[Aaron Stannard @Aaronontheweb]

I need to add optional parameters to dotnet-slopwatch - that'd probably be the most opinionated rule we have but LLMs can't resist doing it as a "safe fallback" https://t.co/whkfbCfLrf

[Aaron Stannard @Aaronontheweb]

An amazing product I use every week to make graphics for this, that, and the other just got even better. Templates pair great with LLMs and their MCP server btw.

(Jeff, I still owe you a blog post and I'm sorry)

[Jeff Needles @jsneedles]

Yesterday I shipped one of the biggest updates to @htmlcsstoimage in years.

It was months of devs but many years of dreams.

🥁🥁🥁

We have a Template editor now! Huzzah!

It's the most frontend-y thing I've ever done.

Now: enjoy

https://t.co/SvHfpKQwU9

[Aaron Stannard @Aaronontheweb]

recorded some YouTube shorts from our last @AkkaDotNET community standup last week

[Aaron Stannard @Aaronontheweb]

NetClaw, human-in-the-loop approvals #netclaw #openclaw #akkadotnet https://t.co/CHYIpjkJzL