Section

Maintenance

Maintenance is the set of scheduled and reactive tasks that keep a system doing what it was bought to do, and the single most consequential decision is not which tool you buy but which failure modes you choose to prevent and which you accept. Everything else — schedules, spares, monitoring, contracts — follows from that split. Get the split wrong and you either overspend on guarding things that rarely break or you spend your weekends firefighting things that should never have been allowed to fail.

2 articles

Maintenance
01

All in Security

Newest first
How to clean a mechanical keyboard safely
Maintenance

How to clean a mechanical keyboard safely

The mistake is pulling every keycap off at once and dropping them into a bowl of water with dish soap.

10 SepRead →
Cleaning a mechanical keyboard without damage
Maintenance

Cleaning a mechanical keyboard without damage

The safe way to clean a mechanical keyboard is to remove the keycaps, clean the exposed plate and switches dry first, use isopropyl alcohol only on the keycaps and the case, and never pour or spray…

9 SepRead →

Maintenance is the set of scheduled and reactive tasks that keep a system doing what it was bought to do, and the single most consequential decision is not which tool you buy but which failure modes you choose to prevent and which you accept. Everything else — schedules, spares, monitoring, contracts — follows from that split. Get the split wrong and you either overspend on guarding things that rarely break or you spend your weekends firefighting things that should never have been allowed to fail.

Preventive versus reactive: the trade-off nobody avoids

Preventive maintenance is worth doing when the cost of the intervention is lower than the expected cost of the failure multiplied by how often it happens — and when the failure is gradual enough that you can see it coming. Reactive maintenance is the correct choice for everything else, and admitting that is not laziness. A cheap commodity item with no downstream consequences is better run to failure than serviced on a calendar.

The way you tell the difference in practice is by watching for warning signs that arrive before the failure. Bearings give you noise and heat before they seize. Filters give you pressure drop before they clog. Batteries give you rising internal resistance and shorter runtime before they die. If a failure mode does not announce itself, calendar servicing is guesswork dressed up as discipline — you are replacing a part at some arbitrary interval that has nothing to do with its actual condition. If it does announce itself, condition-based servicing beats calendar servicing on both cost and reliability, because you act on evidence instead of the passage of time.

The cost of getting this wrong is asymmetric and that is why people over-maintain. Over-maintaining wastes labour and parts but rarely causes an outage. Under-maintaining a critical asset can take down something far more expensive than the asset itself. So the rational default is to over-maintain the few things whose failure cascades and run the rest to failure. The mistake is applying the critical-asset schedule to the whole estate because it feels safer. It is not safer — it is just more expensive, and it dilutes the attention you should be spending on the things that matter.

There is a third category that catches people out: intrusive maintenance that causes the failure it was meant to prevent. Opening a sealed enclosure, reseating connectors, disturbing a calibrated assembly — each of these introduces risk. If a task's preventive value is small, the act of doing it can be net negative. The honest question before any scheduled task is not "could this fail?" but "does touching it make it more or less likely to fail?"

Condition monitoring: what to measure and when it pays

Condition monitoring pays when three things are true: the failure develops slowly enough to catch, the measurement is cheap relative to the consequence, and you will actually act on the reading. Miss any one and you have bought instrumentation, not maintenance.

The measurements that earn their keep are the ones tied directly to a known degradation mechanism. Vibration on rotating machinery. Temperature on anything that dissipates power. Pressure differential across filters. Runtime hours on consumables. Error counters on anything with a log. The value is not the sensor — it is the threshold and the response. A temperature reading with no defined action attached is decoration.

What goes wrong here is threshold drift. Someone sets an alarm level during commissioning, the asset ages, the baseline shifts, and either the alarm fires constantly until people ignore it or it stops firing when it should. The fix is to revisit thresholds when the asset's behaviour changes and to treat a persistent nuisance alarm as a fault in the monitoring, not in the operator. The other common failure is measuring too much. Every metric you collect is a metric someone has to interpret and a signal that can drown out the one that matters. Start with the single measurement closest to the failure mode you are most worried about and add only when you have a reason.

The limit: condition monitoring is the wrong tool for failures that are instantaneous or for assets where the monitoring cost approaches the replacement cost. A sensor and its installation can easily cost more than the component it watches. In that case, run to failure and keep a spare on the shelf.

Spares, lead times, and the cost of holding inventory

Hold spares for the failures that stop you and for the parts whose lead time exceeds your tolerance for being down. Everything else is inventory you are paying to store, insure, and eventually write off when the design changes. The decision is a straightforward comparison between the cost of holding a part and the cost of waiting for it, and most organisations hold too much because nobody has ever been asked to justify a shelf.

The parts that justify holding are the ones with long lead times, the ones that fail predictably, and the ones that are common across multiple assets so a single spare covers several failures. The parts that do not are the high-value low-failure-rate items — those are usually better covered by a service contract or an overnight arrangement with the supplier than by capital tied up in a box. A spare that sits unused for its whole life is not insurance, it is a bet you lost slowly.

Lead time is the number that drives all of this, and it is the number people most often get wrong because they use the catalogue figure rather than the figure they actually experienced. If a replacement has historically taken longer to arrive than the supplier claims, plan to the longer number. The cost of a spare is not the purchase price — it is the purchase price plus storage plus obsolescence plus the capital that could have been doing something else. That total is what you compare against downtime.

The trade-off is real in both directions. Under-holding spares turns a routine failure into an extended outage while you wait. Over-holding ties up money and floor space and creates the illusion of preparedness. The discipline is to review spares against actual failure history rather than against a fear of what might happen.

Schedules, intervals, and how to set them without guessing

Set intervals from observed failure data and manufacturer guidance, then shorten or lengthen them as the data accumulates. A schedule that has never been revised since commissioning is not a schedule, it is a habit. The first pass uses the manufacturer's recommendation because it is a defensible starting point; the second pass uses your own records because they are the only data that reflects how the asset is actually used.

The mechanics matter more than people expect. A calendar interval and a usage interval give different answers, and the right one depends on what drives wear. Equipment that wears with runtime should be serviced on runtime; equipment that degrades with time regardless of use — seals, lubricants, batteries — should be serviced on the calendar. Using the wrong basis means you either service idle equipment that did not need it or miss the wear on heavily used equipment that did.

What goes wrong is interval creep in both directions. Intervals get extended when a task is skipped and nothing breaks, which is not evidence that the task is unnecessary — it is evidence that the failure is rare. Intervals get shortened after a single failure, which over-corrects for an outlier. The fix is to track failures against the interval that preceded them and adjust on the trend rather than on the last event. A task that has never caught anything in years of records is a candidate for removal; a task that has caught the same failure twice at the same interval is a candidate for shortening.

The cost of a too-short interval is labour and parts and downtime for servicing. The cost of a too-long interval is the failure you did not catch. Neither is free, and the balance point is different for every asset class. There is no universal interval — only an interval that fits the failure history you have and the consequence you are willing to accept.

Documentation, handover, and the knowledge that walks out the door

The maintenance record is the asset's memory, and the test of whether it works is simple: could someone who has never touched this equipment service it correctly using only what is written down? If the answer is no, the documentation is not documentation — it is a personal notebook that happens to live at work.

What belongs in the record is narrower than people think. The failure, what was done, what part was used, how long it took, and what was observed. Not prose, not narrative, not a diary. The value is in being able to answer questions later: how often does this fail, how long does the repair take, which part keeps breaking, has this been tried before. Records that cannot answer those questions are not worth the time spent writing them, and that is the honest reason most maintenance logs are abandoned — they were designed to demonstrate diligence rather than to be read.

The knowledge that walks out the door is the tacit kind: which bolt seizes, which sequence to follow, which diagnostic actually finds the fault. It leaves with the person who learned it the hard way, and it leaves without warning. The only defence is to write it down while it is fresh and to have the person who knows do the writing. A handover document produced by someone who has never done the task is worse than nothing because it looks authoritative and is wrong.

The limit of documentation is that it costs time to produce and time to maintain, and a stale record is actively misleading. The trade-off is between writing enough to be useful and writing so much that nobody keeps it current. The pragmatic answer is to document the things that are hard to rediscover — sequences, torques, part numbers, the reason a workaround exists — and leave the obvious to the obvious.

When to repair, when to replace, and when to do neither

Replace when the cost of the next repair plus the expected cost of the failures after it exceeds the cost of replacement, and when the asset's reliability has degraded to the point that downtime is costing more than the capital. Repair when the failure is isolated and the rest of the asset is sound. Do neither when the asset is no longer needed — decommissioning is a maintenance decision too, and one that is routinely deferred because nobody owns it.

The signal that replacement is due is rarely a single dramatic failure. It is a rising frequency of small ones, a growing list of parts that are no longer available, or a repair that costs a meaningful fraction of a new unit. Once you are sourcing obsolete components, you are maintaining a museum piece and the cost curve only goes one way. The other signal is energy or consumable cost: an old unit that runs fine but consumes far more than a current one is a replacement candidate on running cost alone, independent of reliability.

What goes wrong is the sunk-cost trap. Money already spent on repairs makes replacement feel like writing it off, so the asset gets one more repair, and then one more. The correct comparison ignores what has already been spent and looks only forward: from here, is repairing cheaper than replacing, including the downtime and the risk of the next failure? Past spend is a fact about the past, not an argument about the future.

The trade-off in replacing is disruption and the learning curve on new equipment, which is real and often underestimated. A like-for-like replacement is cheap to absorb; a change of platform brings new failure modes, new spares, new procedures, and a period where nobody knows the machine yet. That cost belongs in the comparison. Replacement is not automatically the modern choice — it is the right choice when the forward numbers say so, and the discipline is to run those numbers honestly rather than to let either sentiment or novelty decide.

Maintenance done well is unglamorous: a short list of things that actually fail, a schedule that gets revised when the data says so, spares sized to lead time rather than to anxiety, and records that someone could actually use. The organisations that get it right are not the ones with the most elaborate programmes — they are the ones that have decided, deliberately, what they are willing to let break.