Cloud & Infrastructure
Setting rack temperature thresholds that hold up
A useful monitoring setup starts with knowing where to measure and when to react. This guide walks through placing sensors, choosing alert levels and reviewing results, using a single rack as the worked example.
01
Step one: place sensors where air actually moves
Take a single rack of twelve servers. Put one sensor at the front near the bottom, one at the front near the top, and one at the rear exhaust near the top. Front readings show the air the equipment actually draws in, while the rear reading shows how much heat the rack is adding. Record the equipment manufacturer's stated operating range for intake air before choosing any numbers. Label every sensor with its rack, position and side so an alert later points to a physical spot, not a vague room reading.
02
Step two: set warning and critical levels
Use two tiers rather than one. Choose a warning level comfortably inside the manufacturer's intake range, so someone can investigate while nothing is at risk. Set a critical level closer to the limit, tied to a firm action such as shutting down noncritical workloads. Add a duration, for example ten minutes above warning, so a briefly opened door does not page anyone. Also add a rate-of-rise alert, since a fast climb often signals a failed cooling unit before absolute values look alarming. Write each threshold and its reasoning beside the rack's documentation.
03
Step three: review alerts and adjust the plan
After two weeks, review every alert from the example rack. Note which fired from real heat, which came from sensor placement, and which were cleared before anyone acted. Move a sensor that sits in a draft, and lengthen the duration where short spikes repeat harmlessly. Compare top and bottom front readings; a wide gap can suggest recirculation or missing blanking panels. Keep a short log of each change, its date and the owner. Repeat the review after any new hardware or cooling change, so thresholds follow the room rather than a stale assumption.