Home › Topics › Syslog & Log Collection › What to Log

What to Collect From Network Devices — and What to Drop


Once you have a central syslog server, every device on the network is suddenly willing to tell you everything — and "everything" is a trap. Logging policies fail in two directions. Collect too little and the one line that would have explained the outage was never sent. Collect everything and the line was sent, stored, and buried under forty thousand messages about successfully forwarded packets, where nobody will ever see it. Hoarding feels like diligence. It is actually a way of hiding the signal from yourself.


The fix is a written logging policy: a deliberate, per-device-class decision about what to collect, what to drop, and at what level. It should be small enough to fit on one page, honest enough to admit trade-offs. This guide builds that policy with you. It covers the keep-list for each device class with the reasoning spelled out, and the drop-list with equally spelled-out reasoning. It explains the tuning method for making devices comply and the one-week noise review that keeps the whole thing honest. At the end there is a copyable policy table to adapt to your own fleet.


In this guide

  • Collect with intent, not hope
  • The keep-list: firewalls
  • The keep-list: switches, routers, and Wi-Fi
  • The keep-list: servers, UPS, and environment
  • The drop-list, with reasons
  • Setting device levels deliberately
  • The one-week noise review
  • What over-collection actually costs
  • The one-page policy table

Collect with intent, not hope


A useful test for any log source: can you name the question this log answers? "Who changed the switch config" is a question. "Which nights did the UPS see dirty power" is a question. "Maybe it will be useful someday" is not a question — it is hope, and hope at scale becomes a disk full of noise. The keep-lists below all pass the test. Each entry exists because it answers something you will genuinely ask: what changed, who got in (or tried), what failed, and what is about to fail. Those four questions drive nearly every keep decision in this guide.


One reassurance before the pruning starts: dropping a log category is not destroying evidence. Every device below keeps a local buffer, and debug-grade detail can be switched on during an investigation in seconds. The policy decides what flows to the collector every day, not what is knowable in a crisis.


The keep-list: firewalls


The firewall is your most valuable log source — it sits on the boundary where the interesting stories happen. Keep:


  • Authentication events — every admin login, success and failure. The firewall is the crown jewels. You want a complete record of every hand that touched it, and failed attempts against its management interface are an early warning worth counting.
  • Configuration changes — who, when, from where. "What changed" is the first question of every outage, and rule changes are also exactly what an intruder makes.
  • VPN tunnel events — up, down, and authentication results. Tunnel flaps explain branch-office mysteries, and VPN logins are remote doors into your network; both directions of that story matter.
  • HA state changes — if you run a high-availability pair, every failover event. A pair that quietly failed over last month is a pair with one dead member and no spare. And the log line is often the only witness.

Then comes the famous judgment call: denied traffic. Deny logs are genuinely valuable — they show scans, probes, and misconfigured internal hosts — and they are also the single most voluminous thing a firewall can emit. Logging every deny on a busy edge can multiply your total log volume many times over. The honest answer is to scope rather than choose. First, log denies on inbound-from-internet rules, where every hit is at least mildly interesting. Then skip the chatty internal denies. Broadcast and multicast noise from printers and phones being told "no" thousands of times a day teaches you nothing the second thousand times. If your firewall supports sampling or rate-limiting deny logs, that is a fine middle path for the categories you keep.


The keep-list: switches, routers, and Wi-Fi


Switches

The big principle on switches is that not all ports are equal. Link up/down on an uplink, a server port, or an access-point port is gold — each one is infrastructure, and a flap there degrades service for many people at once. Link up/down on ordinary access ports is mostly the sound of laptops docking and people going to lunch. One person undocking is two lines. But two hundred people across five switches is thousands of lines a day of pure commute noise. At fleet scale it becomes the loudest thing in your collector. If the switch lets you scope link logging per port (or suppress it on access ports), do so; if not, collect it and let the noise review below decide. Beyond links, keep spanning-tree topology changes (loops and root changes are network-wide events with subtle symptoms). Keep auth and config events (same reasoning as the firewall). Keep PoE events (a port that stopped powering its phone or camera, or a switch hitting its power budget). Keep port error messages where the platform logs them — a rising error count is a cable failing in slow motion.


Routers

On anything that runs a routing protocol, the prize is adjacency changes — the messages announcing that a routing neighbor relationship went down or came up. One neighbor flap at 03:12 explains a hundred downstream mysteries: the VoIP complaints, the timeout tickets, the monitoring blips. It is the single log line that most often turns "the network was weird last night" into a diagnosis. Alongside adjacencies, keep config changes and auth events, for the same reasons as everywhere else.


Wi-Fi

From wireless controllers and access points, keep AP up/down (an AP that reboots nightly has a story to tell — power, cabling, or firmware). Keep radar and channel-change events. On radar-shared bands, a detection forces the AP to jump channels, which users feel as a mystery stutter. The log is the only explanation you will get. Keep config/auth events too. The usual noise source is per-client chatter: every association, roam, and disconnect for every phone in the building. Client roaming is normal behavior, not an event; leave it out of daily collection and enable it only when chasing a specific client's problem.


The keep-list: servers, UPS, and environment


Servers and hosts

Keep authentication events — and that means successes as well as failures. Failures get all the attention because they suggest guessing, but successes are what forensics runs on. When an account is compromised, the attacker succeeds. The question "where did this account log in, and when" can only be answered by a record of successful logins. Keep service starts and stops — a service that restarted at 02:00 is your explanation for the 02:00 monitoring blip. And one that stopped without starting again is tomorrow's ticket. Keep storage warnings: full disks remain one of the commonest self-inflicted outages, and the operating system warns in the logs well before the crash.


UPS and environment

Tiny volume, outsized value. Keep power events (on battery, off battery, input voltage anomalies). A UPS that hiccups weekly is documenting your building's power quality for free. Keep battery test results. A failed self-test is the difference between a graceful shutdown and a hard crash at the next outage — batteries age quietly. Keep temperature warnings from any device that reports them. The log that says the closet hit high temperature on Friday evening is the one that saves the switch stack over the weekend.


The drop-list, with reasons


Everything above earned its place by answering a question. These four categories reliably fail the test, and dropping them is where your signal-to-noise ratio is won:


  • Debug-level chatter in production. Debug output is a developer narrating internal states; it can multiply a device's log volume enormously while answering no operational question. It is a fine investigation tool and a terrible steady diet. If a device ships with debug flowing to remote syslog, that is a misconfiguration to fix on day one.
  • Per-packet and per-session logs outside investigations. A log line for every forwarded packet or session on a busy device is a firehose that documents normal behavior at maximum verbosity. Traffic questions are better answered by counters and flow data; keep packet-level logging as a scalpel, not a default.
  • The same event logged twice by two devices. When a VPN tunnel drops, both ends log it; when a port flaps, the switch and the attached server may both notice. Where you control both sides, pick the better-placed witness (usually the device with more context) and quiet the other. Duplicate timelines double your reading for zero new information.
  • The accepted-connection firehose. Logging every successful connection to a busy internal service — every file-share open, every intranet page — produces enormous volume describing the system working as designed. Unless a compliance duty specifically demands it, drop it; the auth events you are already keeping cover the "who was there" question at a sane volume.

Tip: the drop-list is about routine central collection, not about capability. Keep the ability to turn any of these on quickly — a debug session, a temporary per-session rule during an incident. The skill is being able to open the firehose on purpose and, crucially, remembering to close it.


Setting device levels deliberately


Policy becomes reality through each device's logging-level knob — and the detail that makes this workable is that most network gear supports a separate level per destination. The console, the local buffer, and the remote collector each get their own severity floor (the ladder from emergency to debug is covered in syslog fundamentals). On many switch and router CLIs the shape looks like:


logging host 10.20.0.9
logging trap informational      ! level for the remote collector
logging console warnings        ! console stays quiet
logging buffered debugging      ! local buffer can stay verbose

One vendor quirk worth knowing: on several CLI families the command that sets the remote syslog level is confusingly named trap. It has nothing to do with SNMP traps. The standard setting for the collector destination is informational-and-up (severity 6 and more urgent): it captures the full narrative — logins, changes, link events — while excluding debug. The local buffer can afford to be more verbose since it costs no network or collector storage. And the console should be near-silent so it is usable during a crisis. Reserve debug for active investigations, scoped to the subsystem you are chasing, with a reminder to turn it off. The fossil record of forgotten debug sessions is written in every log archive.


The one-week noise review


No policy survives contact with real firmware, so the second half of the method is empirical. Let the fleet log for a week, then rank your sources by log volume and confront the top of the list. With the per-host file layout, the ranking is one line:


wc -l /var/log/remote/*/$(date +%F).log | sort -rn | head

  48210 /var/log/remote/ap-lobby/DATE.log
  31007 /var/log/remote/fw-edge-01/DATE.log
   2214 /var/log/remote/sw-core-01/DATE.log
    880 /var/log/remote/srv-file-01/DATE.log
     12 /var/log/remote/ups-rack1/DATE.log

Read that ranking with the right prior: the noisiest source is usually misconfigured, not important. An access point out-shouting the internet-edge firewall is not the most eventful device you own. It is logging client roams at informational, or someone left a debug flag on. So open the loudest file and find the two or three message types making up most of the bulk. A quick way in is to count by the program tag (the fifth whitespace-separated field in classic-format lines):


awk '{print $5}' /var/log/remote/ap-lobby/$(date +%F).log | sort | uniq -c | sort -rn | head -3

  45320 wifid:
   2101 hostapd:
    790 kernel:

One subsystem is producing almost everything — now grep a screenful of its lines and read what it is actually saying. Counting exact lines with uniq -c alone undercounts, because MAC addresses and session numbers make near-identical messages unique. Count by tag or by a keyword grep instead. Then decide each message type deliberately: is this answering a question, or narrating normality?


When the answer is "drop it," practice the mute-with-a-note discipline. Suppress the message at the source or the collector, and record what you muted, where, and why. Use a comment in the device config, a line in the policy table below, anything findable. The difference between a tuned system and a booby-trapped one is exactly this documentation. A silent drop is a small trap for your future self. Two years from now, someone will search for those messages, find nothing, and conclude the event never happened rather than that it was filtered. Mute freely; mute in writing.


What over-collection actually costs


It is worth being precise about why hoarding hurts, because "storage is cheap" is the argument you will hear against every drop decision. Storage is the smallest of the three costs, but run the arithmetic once to see how fast it moves. The collector guide estimated a modest fleet at roughly 50,000 lines and 7–8 MB a day. Now switch on log-every-deny at a busy internet edge: a firewall can emit hundreds of thousands of deny lines a day. At ~150 bytes each, half a million lines is 75 MB. One rule change, and a single device out-writes the rest of the fleet ten to one, shrinking a years-long disk into a months-long one. The same arithmetic runs in reverse, which is the quiet payoff of the drop-list. Halve the volume and you double both your affordable retention window (log retention does the full math) and your search speed. Because that is the second cost: a bloated archive slows every hunt. Greps that took seconds take minutes. If you graduate to an indexing platform, noise inflates its RAM and disk appetite directly, since every kept message gets indexed.


The real cost is attention. Every screen of routine chatter makes the eye a little worse at noticing the line that matters. And every extra thousand lines makes you slightly less likely to look at all. The archive exists to be read by a stressed human at 2 AM during an incident. Every message you collect is a small tax on that person. Collect like you will be that person — because you will.


The one-page policy table


Here is the whole guide as a single living table. Copy it, adjust it to your fleet, and keep it where the collector's documentation lives:


Device classCollectDropRemote level
FirewallAuth (success + fail), config changes, VPN tunnel events, HA state changes, denies on inbound-from-internet rulesInternal broadcast/multicast denies, accepted-connection logging, debugInformational
SwitchLink up/down on uplinks, server and AP ports; STP topology changes; auth and config; PoE events; port errorsAccess-port link flaps (where scoping allows), debugInformational
RouterRouting adjacency changes, config changes, authPer-packet logging, debugInformational
Server / hostAuth successes and failures, service starts/stops, storage warningsAccepted connections on busy services, application debugInformational
UPS / environmentPower events, battery test results, temperature warningsRoutine status heartbeatsInformational
Wi-FiAP up/down, radar and channel changes, config and authPer-client associate/roam/disconnect chatterInformational

Add a fourth dimension in your own copy: a "muted, and why" column per device class, which is where the mute-with-a-note discipline lives. And notice how well this table feeds the next stage. Nearly every "collect" entry here is also an alert candidate, which is exactly where alerting on logs picks up the thread.


Where to go next


Put a quarterly review on the calendar: rerun the volume ranking, re-read the muted list, and prune whatever new noise the latest firmware invented. Vendors add message types far more often than they remove them. That is the right mental model for the whole exercise: logging policy is pruning, not planting. The devices will always volunteer more; your job is to keep cutting it back to the set that answers questions. The rest of the series — parsing what you kept, alerting on it, and deciding how long it lives — continues at Syslog & Log Collection.


Put this into practice with SysPeep Monitor

The collect-with-intent philosophy is exactly how good monitoring works too. SysPeep Monitor applies it out of the box. It auto-discovers devices and collects health and performance over SNMP, WMI, and SSH. Its thresholds, dependencies, and alert rules are designed to keep noise down — so alerts stay meaningful the way your logs now will. Free trial on any Windows machine.

Download Free TrialLearn About SysPeep