I’m curious as to which tools and technologies you all are using to keep track of all those services you are deploying, whether it be resource tracking, network traffic, logs, traces, or uptime.
As a bonus question, how have you organized your network or your services to reduce the overhead of implementing observability?
I’ve got grafana and Prometheus, but I’m currently working on a generic log pulling script that will then feed to a local ollama instance to summarize the logs and let me know if there are any potential issues I should be aware of. Grafana is great for current state, but I’m a big believer in watching the logs to see trends, but I’m too lazy to do it by hand.
Once that’s up and running and steady, I’ll probably tie it into ntfy to actually alert me in real time
Loki?
Cool, sounds like a fun project.
Not sure if you have been looking into Jev or other System 1 models, but it might be a good fit since it can output a structured type-safe output. Most of the automation I’ve tried in the past with LLMs failed due to the output structure being hallucinated (ignoring my requested structure). But it sounds like System 1 models might fix that.
I’ve been looking at an open source model called Laya a little and it might work well for your project. Sounds like you’d want structured output.
What kind of rig do you have ollama running on?
Jev and ollaya are on my list to eventually add. I’m running ollama through an RTX 3080 with 10 gigs of RAM. Just because it’s what I had lying around. I was mostly using it for gaming but it works for local model use. I’m also planning on installing Hermes in a bit once the log manager project is done, but that’s going to take some careful tinkering because I’m technically underpowered for it. I’m probably just going to have some gate checks up to make sure nothing loads during a gaming session.
My main goal is to completely get off the frontier models if possible because I have concerns about the environmental impacts, especially. As a shameless pirate I mind intellectual theft less, but I really do disagree with the environmental impact, especially when a optimized local model plus a bit of scripting and filtering. Does the job 90% of the time
But we’ll see, I’m still tweaking everything, but I think I’m about 90% done with the log normalizer to do most of the heavy lifting act on the logs programmatically, to keep the context window for my local models as optimized as possible.
It’s definitely been a good fun learning project
Oh, one of the first things that I implemented was a hallucination checker. If it cites any data that’s not actively in one of the corpus I’ve supplied, it automatically gets flagged and filtered out. But most of what I do can be handled with internal documentation and material I explicitly provide.