Anyone who publishes regularly, whether a personal blog, a link directory or a content portal, eventually meets the moment when pages crawl and the dashboard says nothing is wrong. At that point folklore takes over. Somebody says the server needs more memory, somebody else says reboot it, and a third person says the host must be under attack. SysOps, the practice of looking after server hosts, exists partly to replace that folklore with measurement. What follows is a rundown of six stubborn myths, held up against the Ops Atlas notes on each.
The Atlas entry, reachable at its SysOps reference page, defines the field as keeping servers healthy: services, users, logs, scheduled jobs, packages, filesystems, time and performance triage. It also concedes that the label is just a close cousin of system administration. That candour is useful, because it means the advice that follows is not a brand; it is accumulated habit.
Myth 1: “A high load average means the processors are busy”
The load figure is easy to read wrongly. It counts not only tasks that want the processor but also tasks stuck in uninterruptible sleep, which most often means waiting on storage. Picture a content server pulling thousands of thumbnails off a sluggish disk: the load figure can climb to twenty while the processors loaf along, barely working.
What to do instead: look at where tasks are waiting. The Atlas points to the kernel’s pressure stall counters, added in 2018, which separately report how long work has been held up on processor, memory and input-output. One glance tells you which resource is the bottleneck.
Myth 2: “Low free memory means the server is starved”
Linux uses spare memory to cache files, so a healthy machine often shows very little that is labelled free. That is the system being efficient, not distressed. Cached pages are given back instantly when applications need them.
What to do instead: read the available-memory figure. The Atlas names it as the number that matters. If that value is small and swapping is heavy, then you have a real shortage; otherwise, leave it alone.
Myth 3: “Disk utilisation near 100 percent proves the disk is the problem”
The busy-percentage metric was designed for devices that do one thing at a time. Flash drives and RAID sets serve many requests at once, which lets a device announce itself fully busy while still having room to spare. Relying on that number for such hardware appears in the Atlas’s list of habits to avoid.
What to do instead: compare latency and queue depth against what the device normally shows, and keep a baseline so you know what normal is.
Myth 4: “We are in the cloud, so nobody has to do SysOps”
On infrastructure-as-a-service, the provider looks after the physical machine and the control plane. The operating system inside your virtual server remains yours: its patches, its user accounts, its logs, its disk filling up. The Atlas is direct on this point, stating that the guest operating system’s updates are the customer’s responsibility. Provider-side tasks belong to a neighbouring discipline it calls CloudOps; the host itself does not disappear.
The same applies if you run a publishing tool on a rented virtual machine. Anyone following a guide such as the one on publishing an article on DEV Community is renting audience space on someone else’s platform. If you host your own, you also rent the duty to patch it.
Myth 5: “CPU sits at 40 percent, so the slowness must be elsewhere”
This one is half right and therefore dangerous. Averages taken every five minutes hide bursts: a spike that saturates every core for twenty seconds, repeated through the hour, will barely nudge a five-minute graph. Two other causes appear in the Atlas’s notes: steal time on virtual machines, where the hypervisor takes processor cycles away, and saturation somewhere else entirely, such as a database connection pool.
What to do instead: sample at finer intervals during the slow period, check steal time, and follow a quick standard sequence of commands. The Atlas references a well-known ten-command, sixty-second routine published in 2015, plus a method that asks about utilisation, saturation and errors for every resource in turn.
Myth 6: “When in doubt, restart it”
Restarting feels productive, and it often does clear the symptom. It also clears the evidence: memory contents, open connections, the exact state of the stuck process. Restart-first shows up on the Atlas’s anti-pattern list for that reason. A related temptation is to protect a critical process by setting its out-of-memory adjustment to the minimum value; the Atlas notes this simply moves the kill to some other process.
What to do instead: capture first. Save the process list, memory summary, recent journal entries and a pressure reading to a file, then restart. Your future self, and anyone helping you on a forum, will thank you. A good question on a technical forum is one that includes such output, which is the lesson behind the guide to asking a clear question on Stack Overflow.
What keeps the folklore alive
Each of these convictions carries a sliver of accuracy, and that is exactly why it lingers. Load can mean busy processors. Low free memory can mean a shortage. A restart does fix many problems. The error is treating a single number or a single reflex as a verdict. Dashboards invite this, because a graph with one colour line looks like an answer, and most publishers have no time to question it between scheduling posts and answering comments.
The cure involves a routine more than software: before you act on a number, ask which resource it describes, over what interval, and how it compares with the same measure on a quiet day. A couple of minutes spent that way prevents many needless upgrades and knee-jerk restarts, and it costs nothing.
A short routine that replaces the myths
You do not need an elaborate monitoring stack. Four signals cover most of the ground for a small publishing server: disk runway, memory available, pressure stalls and failed scheduled jobs. Check them weekly at first. Write down what normal looks like, so that the next slow day is compared against a record rather than a hunch.
If you share your technical setup publicly, perhaps in a stack write-up, accuracy matters, and the walkthrough on documenting a tool stack on StackShare is a decent prompt to record what you actually run. Doing so also builds the habit the Atlas keeps returning to: recording changes rather than relying on memory.
Folklore is cheap and measurement is only slightly more expensive. When a page next crawls, shrug off all six myths, glance at the signals above, and decide on evidence. Add a line about ‘what we checked’ to the incident note, and the following slowdown will start from knowledge instead of guesswork.