Firewall Logs: Alerting on What Matters (Not Everything)
Turn on full logging on a firewall and stand back. Within a day it will be producing more log lines than every server, switch, and application on your network put together. That is because a firewall sits where all traffic crosses, and it can write at least one line about every single conversation. Somewhere in that torrent are a handful of messages that genuinely deserve to wake you up. For example, the standby unit just took over. The session table is nearly full. Someone is guessing the admin password. The rest is weather.
The failure modes here are the two ditches on either side of the road. Ditch one: overwhelmed by volume, you log nothing, alert on nothing, and the firewall's most important announcements scroll into the void. Ditch two: you alert on everything and receive hundreds of notifications a day. You learn within a week to delete them unread. And you miss the one that mattered. That is the same outcome as ditch one with extra suffering. The road between them is a deliberate sorting. There is a short list of events that page a human immediately, and a second tier that gets reviewed on a schedule. A third is merely kept for the day you need to look something up.
This guide builds that sorting for you. You will learn why firewall log volume behaves the way it does, and where the logs should go. You will learn the real-time alert shortlist (it is shorter than you fear), and the analyze-later tier. You will learn the noise-control techniques that keep the whole thing sustainable. You will get a triage table you can adapt to your own network this week.
In this guide
- The volume reality
- First move: ship everything to central syslog
- The real-time alert shortlist
- The analyze-later tier
- Noise control: keeping alerts believable
- Retention sanity
- The triage table
- A worked example: one night in five minutes
The volume reality
Why so much? Because firewall logging scales with traffic, not with problems. A server logs when something notable happens. A firewall configured to log allowed sessions writes a line for every conversation anyone has — every web page, every DNS lookup, every app's background chatter. Point an internet-facing deny rule at a logger and you add the internet's constant background scanning. Strangers probe your address space around the clock, every probe a dutiful little log line. None of this is wrong. Session logs are gold during investigations ("which machine talked to that address, and when?"), and deny logs are how several attack patterns become visible at all. The volume is not a bug; it is the nature of the vantage point.
But it forces a discipline that server logging never taught you: the decision "log it" and the decision "alert on it" must be completely separate. Logging is cheap-ish storage; alerting is human attention, the scarcest resource you manage. A healthy firewall logging setup writes millions of lines nobody reads, keeps them searchable for the day they matter, and converts only a tiny, curated subset into interruptions. Get that separation into your head first and every following decision gets easier.
Should you even log everything? Mostly, yes — with eyes open. The categories worth sending in full include all system, hardware, and HA events (tiny volume, huge value). They include all administrative activity (logins, changes, commits), VPN events, and denied traffic. The genuinely debatable category is allowed session logging, the single biggest volume driver. Here is the honest trade. With it, you can answer "which machine talked to that address last month?" That is the question every serious investigation eventually asks. Without it, you save substantial storage and give up that answer. Most shops that can afford the disk keep it. And the retention section below is how you make affording it a choice rather than a surprise.
First move: ship everything to central syslog
Before sorting a single event, get the logs off the firewall. The box itself is the worst place for them. Onboard storage is small, so local buffers wrap and overwrite within hours or days on a busy unit. Some products lose buffered logs on reboot. That means the crash you most want to investigate is the event most likely to eat its own evidence. And if the firewall is ever compromised, its local logs belong to the attacker. Central collection fixes all three, and it is the difference between "I think it rebooted around midnight" and an answer.
The standard mechanism is syslog — the venerable, universal protocol where devices send each log line to a collector as it happens. Every business-grade firewall speaks it. Setting up the receiving side (a modest server, storage, and a bit of hygiene) is its own project, covered step by step in building a central syslog server. From the firewall's side it is usually one configuration block: collector address, transport, and which log categories to send. Send generously — the sorting happens at the collector, where storage is cheap and search exists. Two practical notes matter while you are in that config. Classic syslog rides on UDP, which can silently drop lines under load. So use the reliable transport option if your product and collector support it. And check the firewall's log timestamps against reality. Logs whose clocks disagree with your other systems turn every investigation into time-zone archaeology. Point the firewall at time synchronization first and the logs become evidence instead of anecdotes.
And once the stream is flowing, monitor the stream itself. A firewall that goes quiet on the collector has not become well-behaved — the forwarding broke, the collector's disk filled, or someone's "temporary" config change stuck. Because a busy firewall never has a genuinely silent hour, silence is a reliable failure signal. Set an alert that fires when no lines have arrived from the firewall within a window that would be impossible during normal operation. It is the cheapest health check in this entire guide, and it guards all the others.
The real-time alert shortlist
Now the part everyone gets backwards. The events worth a real-time alert from a firewall's logs are not the exciting-sounding security ones. For example, "attack blocked!" is the firewall doing its job, at internet scale, forever. The shortlist is about changes to the firewall's own state and safety margins, because those are rare, actionable, and expensive to learn about late. Five classes:
- HA state changes. A failover, a peer going unhealthy, a sync failure. Rare, always news, always worth a page — you are running without a spare from that moment. The full story is in Monitoring Firewall HA Pairs.
- Resource limit warnings. Most firewalls announce their own distress in the log before users feel it. That might be a session table passing a threshold, memory entering a protective mode, or log disk filling. These lines are the box saying "I am about to have a bad day." That is the exact early warning that session table monitoring and the health metrics exist to catch. The log message is your backstop when a spike outruns the polling cycle.
- Configuration changes. Every commit, every rule change, logged with who and when. If your team is small, alert on all of them. Each should match a change you knew about. The exceptions are either process failures or intruders, both of which you want to hear about immediately. On a busier team, alert outside change windows and review the rest.
- Management authentication failures. A failed admin login happens; a burst of them is someone guessing. And a burst followed by a success is a very bad evening starting. Alert on repeated failures against the management interface, on logins from unexpected sources, and on account events like a new administrator appearing.
- VPN tunnel events. When a site-to-site tunnel drops, a whole branch may be offline; when it flaps, users live in a lottery. Tunnel up/down events are clean, low-volume, high-meaning log lines — ideal alert material. (VPN health is its own discipline, with its own pillar: VPN Monitoring.)
Notice what the shortlist has in common: every item is rare under normal conditions. That is the property that makes real-time alerting work. The moment you alert on something that happens routinely, you have built a machine for training yourself to ignore alerts.
The analyze-later tier
The second tier is everything that is meaningful in aggregate but meaningless as individual lines. Deny patterns are the flagship. No single blocked probe from the internet matters. But consider this question: "which internal machines got denied trying to reach the internet, and where were they going?" It is one of the highest-value security questions you can ask, and yesterday's logs answer it. The same goes for top blocked sources, new-and-unusual destination ports, and a quiet rule that suddenly started matching. It also goes for trends in allowed traffic (why did the file server start talking to the guest network this week?).
A workable weekly review answers three questions in ten minutes. Who inside got blocked, going where? Sort outbound denies by internal source; investigate any server on the list and any workstation with an unusual destination spread. What changed? Compare this week's top denied and allowed patterns with last week's; new ports, new destinations, and newly-busy rules all have stories. Did anything touch the management plane? Skim every admin login and config event; each should attach to a face and a reason. Write down one sentence of findings even when the answer is "nothing" — the trail of sentences is what makes the eventual anomaly obvious.
The right cadence for this tier is a scheduled review, not a notification. Use a saved search or small dashboard you look at daily-ish with coffee, weekly at worst. Keep it to a handful of views: denies by internal source, top talkers among the blocked, management-plane events, anything matching rules you have flagged as sensitive. The habit matters more than the tooling. Ten minutes of pattern-reading per week catches things no threshold ever will, because you carry context no rule can encode. When a pattern from this tier proves reliably meaningful — say, outbound denies from server networks — promote it. Give that pattern a counting alert (next section). Move it into the real-time tier deliberately, with evidence, rather than by anxiety.
Noise control: keeping alerts believable
Even a good shortlist can turn hostile without flow control. Log-driven alerts have a failure mode metric alerts do not: one condition can emit thousands of matching lines. A flapping VPN tunnel logs every bounce; a scanning host trips the same deny pattern per probe; an HA pair in trouble narrates continuously. If each line becomes a notification, your phone melts and your brain learns a lesson: "firewall alerts are spam." That is the most expensive lesson in monitoring. Three techniques keep the channel believable:
- Count-based triggers. Alert when a pattern occurs N times within M minutes, not on each occurrence. "Five failed admin logins in ten minutes" is a signal; each individual failure is a typo with ambitions.
- Rate-limiting and grouping. One notification per condition per window — the first tunnel-down line pages you; the next four hundred join the same incident instead of your inbox. Deduplication and grouping deserve real design, and this guide to deduplication and grouping covers the mechanics in depth.
- Maintenance awareness. Planned firewall work generates the exact log lines you alert on — config commits, HA transitions, tunnel resets. Suppress or downgrade during declared windows, and re-arm automatically after, so the suppression cannot be forgotten in place.
One more quiet trap: do not outsource your priorities to the vendor's severity labels. Firewalls ship with their own opinions about what is "critical," and those opinions were formed for every customer at once. Some products stamp alarming severities on mundane events and file genuinely important ones under informational. Build your alert rules on event types you have looked at and classified (the triage table below is the worksheet), using severity only as a starting hint. The first week of any new log-alerting setup should be a calibration week. Run the rules, count what fired, and demote everything that fired without deserving action. That pruning pass is not an admission of failure; it is the step that makes the survivors trustworthy.
Tip: apply a simple test to every log alert you create: "when this fires at 3 AM, what will the on-call person do?" If the honest answer is "read it, shrug, go back to sleep," it belongs in the review tier, not the paging tier. Alerts are instructions to act, not a news service.
Retention sanity
Because the firewall out-logs everything else, it will also dominate your log storage — and retention deserves a decision rather than a default. Here is the tension. Troubleshooting mostly wants recent logs (days to weeks). Investigations of quiet compromises can need months. And some organizations carry compliance obligations that dictate specific periods regardless of taste. The usual resolution is tiering: recent logs hot and searchable, older logs compressed into cheaper storage, oldest deleted on a schedule you chose on purpose. Compression is unusually kind to firewall logs — millions of near-identical lines squeeze dramatically — which makes keeping more history cheaper than it first appears.
Whatever you choose, write it down and make the monitoring enforce it. Set an alert on the log volume itself. A firewall that suddenly logs far more, or suspiciously less, is telling you something. Set an alert on collector disk headroom too, so retention shrinks by decision rather than by crash. The full sizing-and-policy discussion — including the awkward legal bits — lives in this guide to log retention.
The triage table
The whole guide condensed into one table — adapt the rows to your product's event names and pin it next to the collector:
| Log event class | Tier | Notes |
|---|---|---|
| HA state change / sync failure | Alert now | Page-worthy at any hour; you are down to one engine |
| Resource limit warning (sessions, memory, disk) | Alert now | The box predicting its own outage — act while it is a prediction |
| Configuration change | Alert now / review | Alert outside change windows at minimum; every commit should match a ticket |
| Burst of management login failures | Alert now | Count-based trigger; unexpected source makes it urgent |
| Single management login failure | Review daily | A typo until it has friends |
| Site-to-site VPN tunnel down/flap | Alert now | Grouped — one incident per tunnel, not per bounce |
| User VPN authentication failures | Review daily | Alert only on unusual volume or pattern |
| Outbound denies from server networks | Review daily, alert on burst | Servers rarely develop new hobbies; bursts suggest malware |
| Inbound internet denies | Keep, do not alert | Background radiation; useful in aggregate only |
| Allowed session logs | Keep, do not alert | Investigation gold; alerting on success is absurd anyway |
| Broadcast and neighbor chatter denies | Consider not logging | Pure noise on most networks; check before muting forever |
A worked example: one night in five minutes
Here is how the sorting feels in practice. Six lines pulled from a night's collection — formats vary by product, but every firewall emits equivalents of these:
01:14 deny tcp 203.0.113.45:52113 -> 198.51.100.7:3389 rule=deny-inbound-default 01:52 system: session table at 92% of capacity 02:03 admin: login failed for "admin" from 203.0.113.9 (attempt 5 in 4 min) 02:41 vpn: tunnel "branch-east" down (peer unreachable) 02:41 ha: state change standby -> active on member fw01b 03:22 config: policy change committed by user jsmith
Line one: a stranger probing a remote-desktop port from the internet, denied — keep-only tier, no action, one of thousands. Line two deserves a real-time alert. Something unusual was filling the session table at two in the morning. The on-call wants to find the source now, not read about the outage at nine. Line three deserves a real-time alert via a count-based trigger. Five failures in four minutes against the management interface is a guessing attempt. The source deserves blocking and a note. Lines four and five together tell a story: a tunnel dropped and the HA pair failed over within the same minute. The alerts fire (grouped, one page, both facts attached), and the correlation hands the on-call their first diagnostic clue for free. Line six is a config commit at three in the morning. It is alert-worthy unless jsmith is on the declared change calendar. In that case it quietly matches the window and nobody is woken. Six lines, three interruptions, zero noise. That ratio — not the volume — is what a healthy firewall logging setup feels like.
Where to go next
Logs are the firewall narrating its own story; the rest of this pillar is about checking the narrator. Pair the shortlist here with the polled metrics in Firewall Health: The Metrics Behind the Little Green Light. Keep the session ceiling honest with Session Tables and Connection Limits. And let Rule-Hit Analysis tell you which rules generate the logs worth reading. The full series lives at Firewall Monitoring Beyond Up/Down.
Put this into practice with SysPeep Monitor
Your syslog collector catches the firewall's story; the numbers need a watcher too. SysPeep Monitor polls firewall health over SNMP — session counts, CPU, memory, interface state. Its thresholds, dependencies, and alert rules are built to interrupt you rarely and mean it every time. It pairs naturally with the log setup in this guide, and the free trial installs on any Windows machine in minutes.
Download Free TrialLearn About SysPeep