New Relic’s forecasting is surprisingly good at predicting future capacity needs, but only if you feed it the right kind of historical data.
Let’s see it in action. Imagine you’re running a web service and you want to know how many more application servers you’ll need in a month. You’ve been tracking your request volume and latency in New Relic.
Here’s a query that might show you current request volume:
SELECT count(*) FROM Transaction SINCE 1 week ago FACET appName
And here’s one for latency:
SELECT average(duration) FROM Transaction SINCE 1 week ago FACET appName
New Relic’s forecasting engine looks at these time-series metrics. It identifies trends, seasonality, and even anomalies. If your request volume has been steadily increasing by 5% week-over-week, and you have a holiday coming up that historically doubles traffic, the forecasting will try to account for both. It’s not just a simple linear extrapolation; it uses statistical models like ARIMA (AutoRegressive Integrated Moving Average) to capture more complex patterns.
The core problem this solves is proactive capacity planning. Instead of reacting to performance degradation when it happens (which is often too late), you can anticipate it. This means buying hardware, scaling up cloud instances, or optimizing code before users start complaining about slow load times or errors. You get to see a projection like: "Based on current trends and historical patterns, my-app-production will likely exceed 80% CPU utilization within 2 weeks if no action is taken."
The key levers you control are:
- Data Granularity: How often are you reporting metrics? If you only report every 5 minutes, you’re missing finer-grained fluctuations. Higher frequency data (e.g., every 1 minute) gives the forecasting models more detail to work with.
- Time Window: How much historical data are you feeding it? A week might be too short to capture long-term trends or seasonalities (like annual sales cycles). A year or more is often better.
- Metric Selection: Are you forecasting based on the right metrics? Request count is good, but what about error rates, database connection pool usage, or memory consumption? Forecasting all relevant metrics gives a more complete picture.
- Business Event Tagging: If you can tag your data with significant business events (e.g.,
promo_launch_2023-11-15,black_friday_peak), New Relic can learn how these events impact your metrics, making forecasts more accurate around those periods.
The forecasting models aren’t magic; they’re statistical tools. They assume that past patterns will continue into the future. This means sudden, unprecedented events (like a global pandemic or a viral marketing campaign that completely blows up your traffic overnight) will throw off predictions. The system doesn’t know a new, unpredicted factor has entered the equation. It just sees deviations from established patterns.
When you start using forecasting, you’ll often find that your initial predictions seem off. This is usually because the default settings or the data you’ve been collecting aren’t rich enough. The real power comes from refining the metrics you feed into it and understanding the limitations of statistical prediction when faced with truly novel business or technical changes.
The next step after getting good at forecasting capacity is understanding anomaly detection and how it can flag unexpected deviations from those forecasts.