Book Appointment

Six Monitoring Myths Hospital Software Teams Still Believe

Healthcare Software

Software teams in hospital administration pick up beliefs about monitoring the way waiting rooms pick up old magazines: they accumulate, nobody remembers where they came from, and few are current. Here are six myths about ObservabilityOps, the work of keeping the monitoring and tracing platform itself healthy. The material is non-clinical and covers the systems behind scheduling, billing and staff portals. Each myth gets a short verdict and a practical fix.

Myth 1: “Hoard now, sort out later”

It sounds like caution, yet it may be the costliest belief here. Telemetry that is never queried still occupies disk, memory and network, and it makes the useful signals harder to find. The Ops Atlas files this slogan under anti-patterns. The deciding rarely happens, while the invoice keeps arriving every month.

What to do instead: collect with a purpose. For each stream, write down the question it answers and who asks it. Streams nobody reads go on a deletion shortlist.

Myth 2: “Three pillars means three separate tools”

Metrics, logs and traces are often presented as pillars, and the next step is buying one product per pillar. The CNCF observability working group, in its 2023 whitepaper, prefers the phrase primary signals over three pillars, and the Atlas treats the tool-per-pillar idea as a trap. Signals need to link to each other: a trace identifier in a log line, an exemplar on a metric.

What to do instead: judge tools by how well the signals connect, not by how many boxes they tick.

Myth 3: “A central team ought to hold all the alerts”

This one feels tidy and quietly breaks a team. DORA’s research warns against locking monitoring inside one person or one team, and the Atlas echoes that. A platform owner looks after the plumbing, the standards and the limits; service teams decide which of their own symptoms deserve to wake someone.

What to do instead: split the duties. The owner reviews alert quality and retires unused rules, while each service keeps its own paging decisions.

Myth 4: “Running it ourselves costs nothing, the code is open”

A licence is merely one line of the budget. Running stateful back ends for metrics, logs and traces means upgrades, storage planning, backups and on-call time. The Atlas also points out that several popular tools changed their licence terms, so legal review belongs on the checklist as well, not only the engineering one.

What to do instead: estimate hours and storage before choosing. Even a rough sum, with a named owner, beats an assumption of zero.

Myth 5: “Tracing works once the SDK is installed”

Install is the first ten percent. Traces come apart at message queues that lack span links, at proxies that swallow the trace header, and wherever old and new header formats meet. Meanwhile, context values carried alongside requests may reach outside services when no one screens them.

What to do instead: pick the W3C default headers, test one request across every hop, allow-list the extra keys that may travel, strip them at the boundary and start a new trace where requests come in from the internet.

Myth 6: “Automated analysis will find the problem for us”

Vendor pages promise that machine analysis will surface root causes on its own. The Atlas calls such claims marketing until they are tested on your data. A model fed messy labels, missing trace links and noisy alerts will produce confident noise. Good inputs come first.

What to do instead: run any such feature on a sample of your own telemetry for a month and compare its findings with what your engineers already know.

Why these myths survive

They survive because the pain is delayed. The surveys the Atlas cites, from Grafana Labs, show teams naming complexity, noisy signals and cost as their top frustrations, which is exactly what these six beliefs produce. A 2025 poll put the average observability spend at 17% of compute spend, with a middle value of 10%.

For deeper reading on the neighbouring field of operations, our article about how engineers scale fleets of working robots uses a similar watch-and-maintain lens, and the essay on a joined-up architecture for IT operations helps place this topic among its relatives.

The cost of leaving the myths alone

A hospital software group rarely feels these beliefs as a single event. Instead there is a slow drift: the monitoring server needs a bigger disk every quarter, pages arrive at night that nobody can act on, and engineers begin to ignore the pager because most alerts are noise. The drift is expensive in people, not only in storage. Someone who stops trusting an alert will also fail to trust the one that matters.

A modest habit helps: once a quarter, hold a one-hour review where each myth is checked against what the team actually does. Bring numbers, such as series counts, pages that led to action and monthly cost, and leave with two changes you can finish before the next review.

Four questions to test yourselves

  • Can you name who owns the platform?
  • Do you know which stream costs the most and who reads it?
  • Have you tested a trace across all hops lately?
  • When did you last retire an unused alert?

If you can answer three of four, you are ahead of most teams. The full reasoning, including the scope and the maturity ladder, is set out in the Atlas page on ObservabilityOps. It is not an official standard, so read it as one careful viewpoint, then test it against your own systems.