Cookie preferences

Essential storage keeps iPulse AI reliable. Optional analytics and LinkedIn ad measurement stay off unless you accept. Privacy Policy

Skip to main content

What can a time-series model tell you about a five-minute spike?

Two servers trigger the same alert. Which one would you investigate first? A hands-on guide to reading signals with OpenTSLM.

Matviy Kruglov4 min read
A magnifying glass enlarges a brief spike in a blue ink trace while a red trace peaks a little later.
OpenTSLMTime seriesModel evaluationForecasting

Two servers each spend five minutes above 80% CPU. On the first, all five minutes come together. On the second, they arrive one at a time, ten minutes apart. The monitoring dashboard gives you the same total. Would you send the same person to investigate both?

You already know something the total leaves out. One machine had a sustained burst. The other kept coming back for more. You can ask whether either pattern matches a scheduled job, but the measurements alone can’t tell you which job ran. That gap between what a trace shows and what you still need to look up is a good place to meet OpenTSLM.

How do you ask a line a question?

A time series is a set of measurements in time order. Swap two readings and you may change the story without changing a single value. A time-series language model lets you ask a question about that order in ordinary language.

Take the two servers. Give the model one CPU reading per minute for an hour, tell it the units, and ask it to identify the longest uninterrupted stretch above 80%. For our invented examples, you can score the answer yourself. Five minutes on the first server. One minute on the second. Now ask which process caused the load. A process name would be a guess until you supply a process log.

OpenTSLM-Flamingo has a numerical encoder that turns the readings into a compact representation. As the language model writes its answer, a mechanism called cross-attention lets it consult that representation alongside your question. This gives a phrase like “the second burst” a sequence of measurements to refer to. Read the OpenTSLM research paper.

What would you check before clearing a disk?

In an AWS walkthrough coauthored by Matviy Kruglov, iPulse AI’s strategic advisor, a workplace computer starts hanging. Free disk space is falling and requests for disk access are piling up. Imagine being the person on the support call. Which observation would make you comfortable proposing a cleanup?

Nexthink’s example gives the support agent a practical next step. It checks the space actually available and which temporary files qualify for removal. The employee gets to see and approve the proposed cleanup. Afterwards, the agent checks the free space and asks whether the computer feels faster. Those last two checks can disagree. If the disk has room again and the application still hangs, the support team has learned where to look next. Read Matviy and his coauthors’ walkthrough on AWS.

How could a better score fool you?

In one experiment in the walkthrough, the authors fine-tuned OpenTSLM on multiple-choice questions about software incidents. Accuracy on 135 held-out questions rose from 5.0% to 45.9%. Before training, the model often returned an empty answer. Afterwards, it picked an option every time.

Suppose you were hiring this model for the support desk. You would want two counts side by side. How often did it attempt the question? When it attempted one, how often was it right? Learning to answer in the required format matters. The headline score alone can’t tell you how much of the gain came from reading the traces better. If the actual job is to explain an incident in a sentence, give it that job in the evaluation too.

What happens when you zoom out?

Try a thought experiment. A server sits at 20% utilization for 55 minutes and at 80% for the other five. Its hourly average is 25%. Someone who sees only that average has no way to reconstruct when the burst occurred. Even with all the original readings, finding a short event in a long recording can be difficult. Researchers tested that problem directly in TS-Haystack.

The same issue comes up in market research. Imagine a share opens at 100, dips to 92 and closes at 100. A second share stays at 100 all day. Give a model just the opening and closing prices and ask which share fell sharply at lunchtime. Both records are consistent with either answer. Minute-by-minute prices could settle the timing. To investigate a cause, you would also want the news or exchange notices available at that time. To test a forecast, you would need a later period the model had never seen.

Try it on a chart you know

Pick one event you can locate by eye. Write down its start and end before you ask the model anything. Give it the readings and their units, then ask which interval supports its answer. Repeat with a wider window. Finally, try a quiet interval and see whether it invents a disturbance. A pencil, a spreadsheet and a chart you know well are enough to make the answer checkable.

Go back to those two servers. The most useful answer might be a request for one more record. Which log would let you tell a routine five-minute backup from a problem worth waking someone for?

Meet Matviy Kruglov, coauthor of the AWS walkthrough.

Evidence register

Empirical sources behind this publication

Concept publication record

This owned edition preserves the stable iPulse AI concept record, including visible corrections and source lineage. It is editorial analysis, not formal research, a guarantee, or personalized investment advice.

Follow iPulse AI concepts

Get research highlights by email.

Selected findings, articles and updates through our Substack newsletter.

Subscribe on Substack