Flaem

Server monitoring and statistics: every second, with real peaks

Flaem measures every service and every node each second and keeps the average, minimum, maximum and percentiles for every period, so a short spike does not vanish into an average. Clients see the charts of their own service, and you also see nodes, hardware, the network and who is taking resources from whom.

  • Measured every second
  • Per-minute average
  • Highest peak in the window

The same minute of CPU work: measured every second the spike is plain to see, in the per-minute average it all but disappears.

01Service resources

The full resource picture of every service

Game servers, websites, VPS and LXC containers share one grid of charts going back 30 days. CPU is split into user, system, I/O wait and steal, so a low average cannot hide the throttling that causes lag.

CPU

User, system, I/O wait and steal, plus CPU pressure.

Memory

Memory breakdown, memory pressure and a count of out-of-memory kills.

Disk

Reads and writes in B/s, IOPS and latency.

Network

In and out in b/s and in packets.

Plan limit

For containers, the time spent throttled by the plan limit.

Inside the VPS

What eats the resources: system services and Docker containers inside the machine, with history.

Per-second precision

The agent measures every second and sends the average and the maximum every 5 seconds, so short bursts of CPU, network and disk do not disappear. An open chart of a VPS, an LXC container or a dedicated server refreshes every second, only while someone is watching.

VPS monitoring in the client panel: CPU, memory, disk and network every second, percentiles, events on the chart and an active alert
02Percentiles

Percentiles and real peaks instead of a smooth average

Every chart and summary shows p50, p90, p95 and p99 for any window, with a p5 to p95 band under the average. Network and disk counters remember the highest 5-second rate in every period, so a monthly chart shows the real peak, not a smoothed average.

p50
half of the measurements are lower
p90
9 in 10 measurements are lower
p95
19 in 20 measurements are lower
p99
only 1 measurement in 100 is higher
The client's question
Clients ask why it lagged last night, not what the average was. Yesterday's p99 tick time answers that with one number.
A gap is a gap
Missing data is a gap on the chart, not a flat line at zero. A machine that reported for 20 minutes out of 6 hours was not idle for the rest.
Honest downsampling
Every bucket keeps the count, sum, minimum, maximum and last value. An hour with 12% average CPU and a 100% maximum still shows the incident.
03Typical week

What is normal: a typical week and period comparison

Every metric has a 168-hour weekly profile and a day-by-hour heatmap: players, CPU, traffic. The chart shows an overlay of the same period a week or a month ago and a grey band for "usually at this time".

For looking, not for alerts
The weekly profile sends no notifications. It answers one question: is this spike normal.
Peak hours
A game server owner sees when their players actually play, and you see the same for the whole node.
Period comparison
This week or month against the previous one, on the same axis.
04Black box

Events on the charts and a black box with full-resolution data

Restarts, crashes with a diagnosis, backups, plan changes, updates, attacks and maintenance are markers on the chart. On an event, a ticket or a press of "keep", Flaem freezes the 10-second data from 30 minutes before to 10 minutes after.

  • Restart
  • Crash with diagnosis
  • Backup
  • Plan change
  • Update
  • Attack
  • Maintenance

The ticket often arrives the next day, when only per-minute averages are left. The frozen window stays for as long as you set, for example 30 days, and shows "before and after" at full resolution.

05Game metrics

Game metrics next to resources: TPS, MSPT, players and entities

On a game server, the game charts share one time axis with CPU, memory, disk and network. A TPS drop and an entity spike show up in the same minute, so "it lags in the evening" turns into a concrete cause.

Minecraft server panel: live usage, the player chart and the console
TPSticks per second, how smoothly the world runs
MSPTmilliseconds per tick, how long one game step takes
Playersnumber of connected players
Entitiesmobs, items and other objects in the world, for example in Minecraft
CPUCPU used by the game server
Disk and networkdisk reads and writes, network in and out

the same minute on every chart

Every game has its own rules for what can be read from it, and overall CPU, RAM, disk and network usage comes with every one of them.

06Node and hardware

Nodes at full resolution and hardware health

You see the node itself, not just the list of services on it. The host reports every core, every disk and every network card every 10 seconds, and hardware sensors warn you before something fails.

Node cores on a heatmap

Hardware health

  • Disk health (SMART)
  • Temperatures
  • Memory errors
  • Fans
  • Power supplies

The alert reaches you before something fails, not after.

Cores
Every core separately, on a heatmap.
Disks
Every disk: IOPS, B/s and latency.
Network cards
b/s, packets, dropped packets and errors on every card.
Connections
The node's connection table.
Disk pools
How full the LVM-thin pools are, metadata included, because full metadata can stop every VPS at once.
Memory
ZFS memory and same-page merging, with how many gigabytes it really saves.
07Noisy neighbours

Who is getting in whose way on a node

In the minutes when steal, pressure or disk latency rises on a node, the panel lists the services that were using the most at the time, with their percentage share. Over 30 days you see, for every service, the minutes it caused and the minutes it suffered.

Minutes causedMinutes suffered

An example node, without numbers

Only the operator sees this. Clients get no view into node load, and what to do about a noisy neighbour stays your decision.

08Sites, databases, network

Websites, databases, dedicated servers and the network

The same rules cover the rest of the offer. Every kind of service has the metrics its owners ask about.

Websites

  • Requests per second
  • p50, p95 and p99 response times
  • 5xx error rate
  • PHP pool state: active processes, queue and slow requests
  • JavaScript apps such as Next.js next to PHP
  • CPU and RAM of the site

"The site is slow" gets an answer: the PHP queue, slow code or traffic.

09Maintenance planning

Maintenance planning with a preview of the impact

When you schedule work on a node, the panel shows that node's weekly map: players, traffic and services. The chosen hour gets a description from history, and the panel suggests the 3 least painful windows to choose from.

MonTueWedThuFriSatSun
Suggestion 1: Wednesday, 3:00-5:00

An example hour description: usually 3 players, 12 clients affected.

The day of the week matters
A Friday is judged against other Fridays, not against the weekly average. The evening before a weekend looks different in games than a Wednesday.
You choose
The panel suggests windows, and the decision stays with the operator.
Clients know in advance
Clients on the node get an announcement about the planned work, and the status page shows it publicly or not, as you decide.
10Retention and capacity

Retention, capacity reports and an oversell calculator

You decide how long to keep each kind of data, and you see what it costs before you change it. The finest data is compressed, so 10-second measurements fit into 7 days.

How far back

7 days

every 10 seconds

30 days

gradually sparser points

a full year

history

Basic
CPU, memory, disk and network of every service
Detailed
CPU breakdown, pressure, latency, inside the VPS
Games
TPS, MSPT, players and entities
Node
cores, disks, network cards and pools
The cost before the change
The retention tab shows how much space every class of data will take before you save the change.
Capacity reports
Plain reports: how much CPU, memory and disk the services use on every node.
Oversell calculator
In the Tools tab, a calculator works out prices, plans and oversell from real usage data rather than declared limits.
11Early warnings

Early warnings for the operator

Some failures are visible long before they happen. Flaem warns you when:

  • the installation has no fresh backup,
  • a node is running out of disk space, with a forecast of when it will run out,
  • a node is running out of memory,
  • the licence is about to expire,
  • the licence server cannot be reached,
  • a node agent is out of date,
  • a node clock has drifted.
Flaem operator dashboard: sales and MRR, live nodes and the issues waiting for a person
On-call view in the phone app: alarms, the queue and node health

The warning goes to the person on call, phone included, and onto the "waiting for a human" strip on the operator dashboard.

12FAQ

Questions about monitoring and statistics

Precision, retention and who sees what.

Every second. The agent sends the average and maximum every 5 seconds, and an open chart of a VPS, an LXC container or a dedicated server refreshes every second while you watch it.

As long as you set for each class of data. 10-second measurements fit into 7 days thanks to compression, and the history goes back a year.

No. Clients see the resources of their own service. Node load and who is getting in whose way are visible to the operator only.

No. The typical week and percentiles are for looking. You get alerts for operator early warnings, hardware health and the thresholds you set yourself.

Yes. A BGP session has a state timeline, flaps, prefix counts and its last error, and a tunnel has its traffic, connection age and restarts.

From the real usage of services on your nodes, collected by Flaem, not from plan limits.

We will show the statistics on data from your nodes

Tell us how many nodes you run and what you sell on them. In the demo we will show service and node charts, the black box and maintenance planning, then work out your licence from the price list.