A month of one-second data holds millions of points, but a trend chart is only a few hundred pixels wide, so drawing every point is both impossible and pointless. Downsampling is how a historian reduces that flood to a manageable number for display or export without losing the features that matter. Done naively it hides spikes; done well it preserves them, and the difference is entirely in the method chosen.
Downsampling Time-Series Data in one line: Downsampling is the retrieval-time reduction of a dense time series to far fewer points so it can be plotted or exported efficiently, typically targeting roughly one representative point per pixel of chart width. Unlike compression, which happens once at storage, downsampling happens on demand for a specific view. Good methods such as min/max buckets or largest-triangle selection preserve peaks and troughs, while naive averaging or plain thinning can erase brief spikes entirely.
Compression and downsampling both reduce data, but they answer different questions at different times. Compression runs once as data is stored, deciding permanently which points enter the archive within a tolerance. Downsampling runs at retrieval, deciding how to render whatever is in the archive for a particular chart or export. The archive keeps its full compressed detail; downsampling just picks a view of it appropriate to the screen or the export size in front of you.
The driving constraint for display downsampling is pixel width. A chart that is eight hundred pixels wide cannot show more than about eight hundred distinct points across the axis, so pushing a million points at it wastes bandwidth, slows rendering, and draws pixels on top of pixels to no benefit. The goal is to hand the chart roughly as many points as it has horizontal pixels, chosen so the reduced series looks the same as the full one at that resolution.
Because downsampling is per-view, zooming changes it. When a user zooms into an hour of a month-long trend, the historian re-downsamples the archive over that narrower window, revealing detail that was averaged away at the full-month zoom. This is why a spike can appear as you zoom in even though it seemed absent zoomed out, and why the method must be chosen so important features survive at every zoom level, not just up close.
The dangerous naive method is plain decimation or bucket averaging: divide the time range into as many buckets as pixels and take every Nth point or the mean of each bucket. Averaging a bucket that contains a brief pressure spike blends the spike into its neighbours and it vanishes from the chart, so an operator scanning the trend never sees the excursion. For monitoring data, hiding transient peaks is exactly the failure you cannot accept.
Min/max downsampling solves this directly. For each time bucket it keeps both the minimum and the maximum value seen in that bucket, so every spike and every dip survives as an extreme point even after the range collapses to a few pixels. It roughly doubles the point count compared to a single-value method, but that is a small price for the guarantee that no peak or trough is silently swallowed. For alarm-relevant process data this is the safe default.
The largest-triangle-three-buckets method takes a different tack, choosing one point per bucket but selecting the point that forms the largest triangle with its neighbours, which favours the points that most define the visual shape of the line. It produces a clean, faithful-looking curve with a single point per bucket, preserving the trend's overall form very well. The trade is that it selects representative points rather than guaranteeing both extremes, so many monitoring tools favour min/max where never missing a spike is the priority and largest-triangle where a clean shape matters most.
In a browser-based SCADA, downsampling is what makes long-range trends feel instant. Asking for a year of a one-second tag would otherwise mean shipping tens of millions of points over the network into a chart, which is slow to transfer and slower to draw. Downsampling on the server reduces that to a few hundred well-chosen points before it ever leaves the cloud, so the trend loads quickly even over a modest connection while still showing the real shape of the data.
For oil and gas operators this matters most when investigating an event across a wide window. A supervisor scanning a week of wellsite pressure needs the brief overpressure spike to be visible at that zoom, then wants to zoom in and see it in full detail. A min/max aware downsampler keeps the spike on the wide view and re-resolves the fine structure on the zoom, so the operator is never misled into thinking nothing happened because a naive average smoothed the peak away.
Merobix downsamples on the server so trends over long periods render fast in the browser without hiding excursions, then re-resolves detail as an operator zooms in. Because the reduction preserves the extremes rather than averaging them out, a transient spike on a remote site's pressure or flow stays visible on the overview, and the full-resolution history remains in the archive for the moment someone needs to look closely at exactly what occurred.
Compression happens once when data is stored, permanently deciding which points enter the archive within a tolerance. Downsampling happens at retrieval, reducing whatever is in the archive to fit a specific chart or export without altering what is stored. Compression is about storage cost; downsampling is about rendering a readable view at a given resolution.
Averaging blends every value in a time bucket into one number, so a brief spike gets diluted by the surrounding normal values and disappears from the chart. Min/max downsampling avoids this by keeping both the highest and lowest value in each bucket, so peaks and dips always survive. For monitoring data where transient excursions matter, min/max is the safer method.
It is a method that keeps one point per time bucket but chooses the point forming the largest triangle with its neighbouring selected points, which favours the points that most define the line's visual shape. It produces a clean, faithful-looking curve with few points. Because it selects representative points rather than guaranteeing both extremes, tools that must never miss a spike often prefer min/max instead.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.