How to Maintain a CDU: A Practical Maintenance Guide for Liquid-Cooled Data Centers

The rear service panel of a 4U in-rack coolant distribution unit on the factory floor: paired stainless quick disconnects for return and supply lines, dual power inlets labeled POWER A and POWER B, a dry-contact ALARM OUTPUT terminal block, and leak-detection ports. Every interface you can see in this photo is a maintenance item — and every one of them has a story about what happens when it is ignored.

A coolant distribution unit is the one box in the white space whose failure stops every rack it feeds. Fans can degrade gracefully; a CRAC unit can fail over to its twin. But when the single pump skid that circulates coolant through forty GPUs stalls, the thermal mass above it is measured in seconds of tolerance, not minutes. That asymmetry is why CDU maintenance deserves its own program instead of being folded into the general facility checklist as an afterthought.

The good news: keeping a CDU healthy is not complicated. It is a rhythm of short, boring checks — daily trend sweeps, weekly coolant and leak walks, quarterly redundancy drills, an annual deep service — plus a handful of chemistry rules that never change. This guide walks through that rhythm interval by interval, explains what each check is actually protecting, and ends with a schedule table you can lift straight into your maintenance system.

Start by knowing every service interface on the unit

Maintenance begins before the first task on the calendar. Walk up to your CDU and identify every service interface, exactly as shown in Figure 1. The paired stainless couplings are dry-break quick disconnects — one pair for the return line, one for the supply, and on many units an additional pair bridging the secondary loop to the rack manifold. The two IEC-style inlets labeled POWER A and POWER B feed redundant power from separate branches; lose one and the unit keeps pumping. The green terminal block marked ALARM OUTPUT is the dry-contact path that hard-wires pump faults, over-temperature and leak events into your BMS or EMS without depending on the network. The small ports labeled LEAK DETECTION connect to sensing rope laid under the unit and along the pipework.

Do three things during that first walk-down. Photograph the nameplate and file it with the as-built documentation, because model, cooling capacity and serial number are what you will quote on every spare-part order. Label both power branches at the PDU side so nobody kills the wrong feed during service. And confirm which quick disconnect belongs to which loop — modern dry-break couplings are keyed to prevent cross-mating, but you should know the pairing by heart anyway. Industry checklists rank quick disconnects, their EPDM seals and the leak-detection rope among the most critical spares to hold for exactly these interfaces [1].

Daily: watch the trends, not just the alarms

Alarms tell you when something has already failed. Trends tell you when it is about to. Build a five-minute daily routine around the unit's HMI or your DCIM dashboard, and log it — a written daily record of temperatures, pressures and flows is the single most recommended habit in every operator guide we reviewed [2].

What to sweep every day:

A gradual rise in supply temperature at constant IT load is the classic early signature of a fouled heat exchanger — catch it here and you can plan a chemical cleaning for a maintenance window instead of discovering it during a thermal throttle event [2]. One caution: if your secondary loop runs below room dew point, condensation forms on cold pipework — ASHRAE's liquid-cooling classes impose a 2°C minimum on chilled-fluid temperature relative to room dew point for exactly this reason, so keep an eye on your dew-point calculations; we cover that interaction in our dew point guide on this blog [6].

Weekly and monthly: keep the coolant and the filters honest

Weekly, two walks. First, the coolant: check expansion tank or reservoir level and glance at fluid condition — discoloration, visible particulates or an oily film all deserve a lab sample. Second, the leaks: run a hand and an eye over every fitting, valve stem, pump seal and quick disconnect you can safely reach, and confirm the leak-detection controller shows a healthy, untriggered state [2]. Operator guidance for data-center CDUs converges on weekly coolant level and quality checks, with conductivity and inhibitor concentration monitored monthly — a sudden change in either is the fingerprint of corrosion, dosing drift or biological growth [2].

Monthly, add the filter and strainer routine. Read the differential pressure across every strainer and cartridge and compare it to the clean baseline you recorded at commissioning. A climbing dP means capture, which is exactly what the filter is for — but a filter pushed past its change-out point starts passing particles or collapsing, and either outcome sends contamination downstream toward cold-plate microchannels. In dusty retrofit spaces, check inlet filters every two weeks rather than monthly [2]. Clean Y-strainer baskets quarterly at minimum, and more often on young systems that are still shedding installation debris.

Quarterly: exercise the redundancies before they need to work

Redundant hardware that never runs is redundant hardware you cannot trust. A standby pump that has sat idle for a year may have a seized bearing or a dried seal, and you will find out at the worst possible moment. The quarterly drill list:

• Pump inspection. Listen and feel for vibration or cavitation — a gravelly noise is a bearing or impeller complaint — and trend motor current if the unit exposes it; rising amp draw points at bearing wear or impeller fouling [2]. Inspect shaft seals for weeping [1].
• Redundancy switchover test. Force the lead/lag changeover from the controller and confirm the standby pump picks up the loop cleanly, with flow and supply temperature holding steady. Commissioning specs for well-built units typically require the standby pump to carry load within seconds with less than a 10% flow dip — if your unit takes materially longer today than it did at acceptance, the standby side has a problem [7].
• Valve exercise. Stroke motorized control and isolation valves through their full travel and compare demanded vs. feedback position; deviation grows as actuators wear [1]. A valve that never moves will eventually stick — exercising it is cheaper than replacing it flooded.
• Sensor calibration check. Verify temperature, pressure and flow readings against a reference instrument. Quarterly drift checks are standard practice; miscalibrated sensors do not just mislead you, they quietly detune the control loop [1][7].
• Leak-detection function test. Trigger the sensing rope or spot probe deliberately (a wetted cloth works) and confirm the controller alarms, the dry contact closes, and any configured auto-isolation actually isolates [1]. An untested leak detection loop is a suggestion, not a system.
• Alarm path end-to-end. Fire a test event and verify it lands in the BMS/DCIM through the dry-contact output — the green terminal block in Figure 1 — not just on the local HMI.

Annually: the full service that resets the clock

Once a year, the CDU gets a planned deep service — typically aligned with the facility's broader cooling maintenance shutdown [2]. The annual scope:

• Coolant analysis and refresh. Send a sample to the lab: freeze point by refractometer, pH, inhibitor reserve, and dissolved metals, which are the direct evidence of corrosion in progress [3][4]. Replace the charge per the fluid manufacturer's interval — do not guess [2]. Even well-inhibited fluid has a finite life, and degrading inhibitors accelerate attack on every wetted metal in the loop [4].
• Heat exchanger service. If approach temperature has drifted upward over the year (more on trending below), plan a chemical clean of the plate heat exchanger; many operators fold exchanger cleaning into the annual visit by default [7].
• Electrical and controls. Re-torque power and terminal connections, dust out enclosures and confirm fans are clear, then update controller firmware to the vendor's current release [2][7].
• Full sensor calibration. Annual formal calibration of the temperature, pressure, flow and level sensors, with certificates filed against the asset record [2].
• Insulation and condensation walk. Inspect pipe insulation for gaps and wet spots — wet insulation is both a corrosion point and, in the wrong place, a drip hazard over live equipment [7].

Budget a full day for a rack or row-level unit including refill, venting and a 72-hour observation tail with daily parameter review — the same extended trial that commissioning teams use to prove an installation is stable [7].

Coolant chemistry: what quietly kills CDUs

Most CDU failures that are not mechanical are chemical. The coolant is doing three jobs at once — carrying heat, protecting metals, and suppressing biology — and neglect hits all three. The baseline for cold-plate secondary loops is inhibited propylene glycol (typically PG25 — 25% PG in deionized water): PG is strongly preferred over ethylene glycol on the IT side precisely because it is far less toxic, and a leak inside a GPU rack is a very different incident with a non-toxic fluid [4]. Facility-side outdoor loops sometimes run EG for slightly better thermal performance, but the secondary loop that touches your cold plates should be PG [4].

Glycol concentration follows the coldest spot the loop can see, not the average room temperature. Practical freeze-protection figures for propylene glycol:

Keep at least 3–5°C between your coldest credible operating condition and the measured freeze point of the mix — running tight to the theoretical freeze point is how slush forms, plates block and heat exchangers die [3]. Test the actual concentration twice a year with a refractometer (pre-winter and pre-summer), alongside pH and inhibitor levels [3]. Watch pH over the fluid's life: a healthy inhibited PG mix sits roughly between 7.5 and 9.0, and a falling pH means the inhibitor package is being consumed [4]. And resist the urge to over-concentrate — beyond roughly 60% glycol you gain almost no freeze protection while paying heavily in heat-transfer efficiency and pump head [4].

One more chemistry rule with a metallurgy root: galvanic corrosion. Where two dissimilar metals sit in electrical contact inside a conductive fluid, the less noble one becomes the anode and corrodes preferentially — the classic failing pairs in liquid cooling are aluminum cold plates against copper piping or brass fittings [5]. Control it the boring way: keep the wetted-material list consistent, use isolation fittings where dissimilar metals must meet, and treat any coolant chemistry change — new inhibitor, different glycol brand — as a wetted-materials event, not a procurement event [5]. If your loop is already showing fouling symptoms, our guide to preventing and removing fouling in liquid cooling loops walks the cleaning chemistry in depth.

Quick disconnects and O-rings: the smallest parts with the biggest incident potential

Nothing on a CDU concentrates more failure energy per square centimeter than its quick disconnects. They are mated and unmated during every service event, they carry the full loop pressure, and a slow weep at a QD sits meters above equipment that cannot get wet. Inspect them quarterly: look for seal weep rings, drips on de-mating, and external damage [1]. A dry-break coupling that leaves a drop behind on disconnect is telling you its seal is tired — replace the seal cartridge, not your confidence.

Handle them correctly and they last for years. Mate and unmate slowly and squarely; twist-loading chews O-rings. Cap every open port immediately — a QD dangling open is a contamination intake. Keep mating surfaces clean and lubricated only with the lubricant the coupling vendor specifies. And when you order spares, buy genuine seal kits for the exact coupling model; industry checklists put QD seal kits and EPDM O-ring sets in the CRITICAL spare tier precisely because a “close enough” O-ring in a glycol loop is a leak scheduled for the worst week of the year [1]. For a deeper dive on coupling selection — blind-mate vs. standard dry-break, and why seal material choice matters — see our UQD quick-disconnect selection guide.

The maintenance schedule at a glance

Consolidating the intervals above into a table you can hand to the technician who inherits the program. Interval sources: operator guides [1][2], glycol program guidance [3][4], and industry engineering references for CDU acceptance and operations [7].

Spares: buy back your downtime

A maintenance program without a spares shelf is a wish, not a program. The checklist tiering from industry guides is a good shopping list: filters and cartridges are cheap and critical; a complete EPDM O-ring and gasket kit for your unit's ports and couplings; a mechanical seal kit for each pump model; a level/leak sensing rope section; and at minimum one spare flow or pressure transmitter of each type installed [1]. Pumps themselves sit in the recommended tier — if your deployment grows past a handful of units, holding one spare pump for the fleet beats paying emergency freight twice [1].

Two economics notes. First, the whole coolant charge for a rack-class CDU — roughly 60 to 200 liters — runs on the order of a thousand to a few thousand dollars depending on fluid and concentration [1]; that is rounding error against a single lost training run. Second, consumables expire on the shelf too: O-rings age, glycol inhibitors deplete, gaskets take a compression set. Date-label everything you stock.

Trend your way from reactive to predictive

Every measurement in this guide gets more valuable when you plot it over time. Four trends carry most of the predictive signal:

• Approach temperature. The gap between your facility water supply and the CDU's secondary supply, at a given load. Design approach on a healthy plate exchanger is a few degrees. If that gap grows by 1–2°C at constant load over months, fouling is arriving — schedule the chemical clean before the throttling event, not after.
• Pump speed creep. Variable-speed pumps holding constant flow: rising speed at the same flow means rising system resistance — clogging strainers, closing valves, or plate fouling.
• Filter dP slope. Not the reading — the slope. A steepening slope predicts the change-out date better than any calendar.
• Conductivity drift. Slowly climbing conductivity means ions are entering the loop: corrosion, exhausted DI media, or makeup water mistakes. It is your earliest chemistry alarm.

Operators who log daily and review monthly catch these drifts quarters before they become incidents [2]. Operators who rely on alarms alone meet their heat exchanger failures in production. There is no third category.

When to stop and call the vendor

Some findings exceed what site maintenance should tackle. Call the OEM or your service partner when you see: supply temperature you cannot hold even with a clean filter and stable chemistry; flow below nameplate after strainer cleaning and venting; a redundancy switchover that trips or fails to carry load; repeated leak-detection triggers with no visible wetting (sensing cable contamination or controller fault); conductivity that will not settle after a partial drain and refill; or any seep at a pump seal that returns after a seal-kit replacement. Document what you observed, the trend history, and what you already tried — a good service organization can halve its diagnosis time with that packet.

The bottom line

CDU maintenance is not heroics; it is a rhythm. Daily trend sweeps, weekly coolant and leak walks, monthly chemistry and leak-detection tests, quarterly redundancy drills, an annual deep service — and a spares shelf that matches the critical tier of every part you just inspected. Do the boring checks on schedule and the unit in Figure 1 will do its own boring job for a decade: moving hundreds of kilowatts quietly, at constant temperature, without asking for attention.

Planning a liquid-cooled deployment, or inheriting one with no maintenance history? Request a quote and our engineers will size the cooling chain end to end — CDU, piping, manifolds and quick disconnects — and hand you the commissioning baseline your future maintenance program will stand on. For the surrounding reading: our rack vs. cabinet CDU comparison explains where the redundancy should live in your architecture, and our coolant pipe sizing guide covers the flow arithmetic between the units.

References

1. Resistance Zero, “CDU Install, Inspection & Maintenance Checklist” — component-by-component intervals: quarterly pump seal and vibration inspection, monthly filter dP, quarterly QD and O-ring inspection, glycol-compatible EPDM (not NBR) seals, monthly leak-rope function tests, quarterly sensor calibration, and CRITICAL-tier spares including QD seal kits, filters and coolant charge sizing (60–200 L, $8–25/L). resistancezero.com
2. SHC Refrigeration (TopColdRoom), “How to maintain a CDU cooling distribution unit” — data-center CDU routines: daily logging of temperature and pressure trends, weekly coolant level and quality checks, monthly conductivity and inhibitor monitoring, two-week filter checks in dusty environments, pump vibration and amp-draw trending, quarterly valve and connection checks to manufacturer torque, and annual fluid replacement, sensor calibration and enclosure care. topcoldroom.com
3. Evolution Cooling, “Glycol In Process Cooling: When You Need It, How To Specify It, And How To Avoid Freeze-Ups” — PG concentration vs. freeze protection (20% ≈ −8°C, 30% ≈ −14°C, 35–40% ≈ −18 to −22°C), the 3–5°C safety margin from measured freeze point, refractometer testing twice per year, and partial refresh every 3–5 years in clean systems. evolution-cooling.com
4. Alliance Chemical, “AI GPU Liquid Cooling: Deionized Water, Ethylene Glycol & Propylene Glycol — The Ultimate Liquid Cooling Guide” — PG preferred for secondary/server loops for low toxicity, EG for facility loops; typical data-center concentrations 30–50%; no benefit above ~60% glycol; inhibited vs. uninhibited fluid life; annual fluid testing with pH held at 7.5–9.0; top-ups with DI water only. alliancechemical.com
5. NASA Kennedy Space Center, “Forms of Corrosion” — galvanic corrosion mechanism: dissimilar metals in electrical contact bridged by an electrolyte; the galvanic series; and control by isolating dissimilar metals (insulating gaskets, sleeves, washers and isolation flanges). public.ksc.nasa.gov
6. ASHRAE TC 9.9, “The ASHRAE Thermal Guidelines for Data Centers — Past, Present, and Future” — facility water system liquid cooling classes: W32 (cooling tower primary, 32°C / 89.6°F facility supply) and W40 (40°C / 104°F), with a 2°C minimum water temperature relative to room dew point across classes — the context for supply-temperature setpoints and condensation margins discussed above. dallas-ashrae.org (PDF)
7. Lianli Work, “In-Depth Research Report on Data Center CDU Technology (2026)” — engineering acceptance and O&M reference table: commissioning criteria including leak-detection alarm within 2 s, standby pump pick-up within 5 s with <10% flow fluctuation and secondary conductivity <1 µS/cm at handover; O&M intervals covering daily parameter logging, weekly leak/filter/joint checks, monthly insulation and electrical checks, quarterly strainer cleaning and sensor calibration, semi-annual pump seal replacement and redundancy switchover tests, annual coolant replacement and heat-exchanger chemical cleaning, and biennial exchanger teardown. cn.lianliwork.com