Hi! I'm Larry, founder and CTO at Zebrium. I used to build bespoke machine data analytics platforms for software and appliance vendors. These systems were really useful, but it took a lot of work to make and maintain them. So, I took a couple years to myself to learn, apply, and develop the ML chops I needed to build such a system automatically. That's what Zebrium is all about: applying a range of techniques at different scales to achieve near-100% structured log data coverage, with no supervision or pre-structuring, from as little as 100K of real-world log data.
Our software builds a catalog of event types and parameters for each of a set of log types (for us, a "stack" corresponds to a set of log types). With this context, it finds "anomalies" through the logs, using a model built on a set of features. These features include simple things like severity and first/rare occurrence, and complex things like change in rate/periodicity, cross-event-type and cross-stream correlation, NLP topic, and timeseries features. We've trained our model on data from a few dozen stacks, including dozens of real and interesting "anomalies" the operators would have liked to see uncovered.
Our software seems to spit out "good stuff" most of the time, and tends not to want to "ring the pager" when nothing breaks. I think it's become quite useful - but I'm biased. :) We want to continue to make it better, so we want to get feedback on what happens when people submit log files with interesting problems that could have been spotted within the logs. Let us know what we get right and wrong.
Our SaaS service is in private beta, but anyone can try our log anomaly detection. You can upload up to 5 logs at once (more data and related files improve accuracy) and get a report listing your anomalies, the reasons for the anomalies, and a visualization of the event patterns within your logs. The report is sent to you by email (this is why we ask for your email address). The service is free and you can use it as many times as you want (limit 500MB of logs each use). Please try it (www.zebrium.com/anom-detector) and let us know what you think.
Co-founder here.
Understand your concern and perspective Brad.
In the near term this is a SaaS offering, so the best we can do to address these concerns is to offer strong security controls (a short list on our site, longer list on request) backed up by external certification (underway).
Plus the credibility and track records of the team behind it.
I’m going to echo the GP - this does look like an interesting tech, but it must work in a privacy-preserving fashion. Logs contain loads of sensitive info, especially if they are of a verbose/debug type, which is what will probably be used to capture failures. But in SaaS form, where it wants raw logs, it’s a no-go even for the eval purposes.
Gavin from Zebrium here. Again, we understand this concern. All data is encrypted in-flight and at rest and we have a lot of security controls in place (see our website). We also have the option of a dedicated VPC assigned to a single customer. Beyond this, we have a unique capability: since our machine learning structures and types everything, we have a feature that lets us hash (or delete) any sensitive information in logs (and we can record match to find other places this info occurs that you might not even be aware of). Not dodging the fact that we're SaaS, just pointing out what we do.
5 comments
[ 3.8 ms ] story [ 37.8 ms ] threadOur software builds a catalog of event types and parameters for each of a set of log types (for us, a "stack" corresponds to a set of log types). With this context, it finds "anomalies" through the logs, using a model built on a set of features. These features include simple things like severity and first/rare occurrence, and complex things like change in rate/periodicity, cross-event-type and cross-stream correlation, NLP topic, and timeseries features. We've trained our model on data from a few dozen stacks, including dozens of real and interesting "anomalies" the operators would have liked to see uncovered.
Our software seems to spit out "good stuff" most of the time, and tends not to want to "ring the pager" when nothing breaks. I think it's become quite useful - but I'm biased. :) We want to continue to make it better, so we want to get feedback on what happens when people submit log files with interesting problems that could have been spotted within the logs. Let us know what we get right and wrong.
Our SaaS service is in private beta, but anyone can try our log anomaly detection. You can upload up to 5 logs at once (more data and related files improve accuracy) and get a report listing your anomalies, the reasons for the anomalies, and a visualization of the event patterns within your logs. The report is sent to you by email (this is why we ask for your email address). The service is free and you can use it as many times as you want (limit 500MB of logs each use). Please try it (www.zebrium.com/anom-detector) and let us know what you think.
Give me a tool that I can run locally, and I might be interested.