I did this with the panther file generated by running a trace on android and Gemini advanced seemed impressively capable and able to analyze the large file, and ouput a pretty technical breakdown of everything; pointing out culprit apps (often gApps) that are misfiring and causing latency issues etc. Was curious if anyone else has any experience and or opinions on this
I haven't given it massive logs but I've given error logs to ChatGPT and it's helped me figure out subtle issues very nicely. Gives me a few ideas of what could be wrong, ways to diagnose them, and even commands to run to test or fix things. I've been shocked how well it works.
I worked on this project:<p><a href="https://www.cs.dartmouth.edu/~dfk/research/aslam-kerf-news/aslam-kerf-news.pdf" rel="nofollow">https://www.cs.dartmouth.edu/~dfk/research/aslam-kerf-news/a...</a><p>which had a machine learning component that automatically helped refine syslog SQL queries to find anomolies. More info here:<p><a href="https://www.cs.dartmouth.edu/~dfk/research/project/kerf/index.html" rel="nofollow">https://www.cs.dartmouth.edu/~dfk/research/project/kerf/inde...</a><p>Before LLM's.
The company [0] I work for gave a talk on this [1]. We're going a bit beyond analyzing logs because we have more contextual information about running software so we can compare different application state over time and infer whether it appears the application is experiencing an "incident".<p>0. <a href="https://noop.dev" rel="nofollow">https://noop.dev</a><p>1. <a href="https://www.youtube.com/watch?v=PxZx71xfIl8" rel="nofollow">https://www.youtube.com/watch?v=PxZx71xfIl8</a>
My friend (former MSFT principal engineer) is actually developing this idea as a startup. He has a MVP. If interested, I can connect you two. Email in profile.
Had a system where logs were watched to trigger various sound samples (bird chirps for firewall probes, etc). There was a pattern recognition component, when the entropy on that came up there was a "hmmmm?" sound.
I struggle with dyslexia which makes it very hard to read through logs, so I often troubleshoot error logs, to summarize the logs and break them down by meaning.<p>I wouldn't ever do this in an automated way. However in person, it not only makes me solve problems maybe five times faster, but there are also many errors that I would give up on, which I can now solve easily.<p>Just the ability to turn a log into a human readable narrative does something remarkable. After I've done it once or twice, then my brain provides the narrative itself, and I can read the raw log without the need for assistance.<p>This suggests that not only does the AI assist me, but it assists my learning as well.<p>I also struggle with Man pages, in particular finding an option to do something that I need to do. I end up having to read through dozens of options getting more and more confused, sometimes skipping the option that I needed and not finding it. Now I can just dump the entire man page into an llm, describe what I need to do, and ask it to highlight the sections that I need to read. Or I can just ask it questions.