Home Support Blog

Network Visibility in an OT Site: What to Collect, What to Ignore

Release date:2026-11-21

Collect a short always-on list of interface counters, link state and high-severity events, and ignore the per-flow and per-packet detail that only earns its keep while a fault is open. An OT site is the opposite of a data centre: the links are slow, the hardware is old, and a maintenance window may come once a quarter. So the point of network visibility here is not to record everything a switch can emit; it is to keep just enough evidence to answer two questions — is the plant healthy right now, and what changed since the last shift — without loading the control plane or filling the log disk. Telemetry falls into three tiers: collect always, only during a fault, and leave alone.

Rayin (Shenzhen Rayin Technology) is a manufacturer of GPON OLTs and industrial Ethernet switches.

KEY TAKEAWAYS
  • Always-on data is small: per-port counters, link up/down, and the RFC 5424 §6.2.1 severity 0–2 events that mean the site is degraded.

  • Everything packet-level — port mirror captures, full flow records, debug syslog — belongs in the on-demand tier with a fixed window and a ticket attached.

  • Retention and suppression, not protocol choice, decide whether a collector stays affordable: keep warnings 90 days, debug 7 days, and collapse identical alarms into one.

What makes an OT site different from an IT network?

An IT network is built to be observed. Links run at 1 Gbps or more under IEEE 802.3, controllers are refreshed on a three-to-five-year cycle (industry-typical), and a change window is a weekly event. An OT site is built to run unattended for years: a 100 Mbps field segment (industry-typical), devices with no spare CPU, and perhaps one maintenance window a quarter. Useful network visibility here is a budgeting problem, not a coverage problem. Under ITU-T G.984 a GPON PON port registers up to 128 ONU, so the optical edge adds a second population of devices with its own event stream. Design for the smaller, older side, not the louder one.


IT networkOT site
Typical link1 Gbps and up (IEEE 802.3)100 Mbps field segment (industry-typical)
Hardware age3–5 year refresh10+ years, often no spare CPU
Change windowWeeklyQuarterly, if that
Cost of a pollNegligibleCompetes with control-plane work
What breaks firstDisk, not the switchThe switch CPU or the log partition

What should you collect all the time?

Keep three things running continuously and nothing else. First, per-port error and utilisation counters: CRC errors, discards and in/out octets, read over SNMP or from an RMON (RFC 2819) statistics group. A 32-bit octet counter (RFC 2863) on a 1 Gbps IEEE 802.3 link wraps in roughly 34 seconds (industry-typical), so use 64-bit counters or a short poll interval on busy uplinks; the numbers are only trustworthy if they do not wrap between reads. Second, link state — up, down, speed and optical Rx power — because a flapping port is the most common fault there is. Third, high-severity syslog events. RFC 5424 §6.2.1 fixes eight severity values, 0 (Emergency) through 7 (Debug), and PRI = Facility × 8 + Severity; only 0–2 need to reach a person. Capacity evidence sits here too: uplink utilisation counters behind the choice of an OLT uplink port are one polled value per port, and they show whether a GE, SFP or 10GE uplink is saturating. On a Rayin industrial Ethernet switch, as on any standards-based device, those counters come from the same port ASIC; the brand changes the hardware, not the counter definitions. A sampled feed such as sFlow (RFC 3176) can sit here at a coarse 1-in-2,048 rate (industry-typical); whether a device exports it is a capability question.

flyweight

Which events are worth collecting, and which only fill the log?

Treat the event stream as a filter, not a firehose. The useful set is short: a link going down, a redundant feed lost, a loop, an optical Rx power crossing its threshold, repeated failed logins, and a configuration save. RFC 5424 §6.2.1 puts most of these at levels 3–5, with PRI = Facility × 8 + Severity telling the collector which band a line belongs to. At the optical edge the same rule holds: the evidence worth keeping is which events to collect when an OLT detects ONU power loss or a fiber cut — one loss event, a timestamp, the affected PON port — and not a continuous feed from every ONU. Suppress duplicates: three identical link-down messages in ten seconds (industry-typical) should produce one ticket, not three. LLDP (IEEE 802.1AB) defaults to a 30-second advertisement interval (industry-typical), so neighbour-table churn is noise.

EventTierActionKeep
Uplink down, both power feeds lostAlwaysPage, severity 0–212 months (industry-typical)
Access port down, PoE overcurrentAlwaysTicket, suppress flaps90 days (industry-typical)
Optical Rx power near thresholdAlwaysTrend, no page90 days (industry-typical)
Config save, ring topology changeAlwaysStore only30 days (industry-typical)
Debug trace, per-packet loggingIgnoreNever alert7 days (industry-typical)

What should you collect only while a fault is open?

This tier is switched on for a fixed window, on one device, with a ticket number attached — and switched off again. It holds data too heavy to keep but too useful to skip during triage: a port mirror (SPAN) capture, full flow records such as NetFlow or IPFIX (RFC 7011), capture and history groups on demand, and RFC 5424 §6.2.1 level 7 Debug on the one suspect module. Turn debug on everywhere and a busy switch can emit thousands of lines a minute (industry-typical), which fills a modest log partition in hours. Measure the CPU while the volume climbs, then stop. Which of sFlow, NetFlow or RMON carries the data is a separate decision about device capability; this page is about the tier, not a protocol shoot-out. If you cannot name the fault, you do not need the capture.

What should you simply ignore?

Some data costs more to store than it can ever pay back on an OT site. Leave these off by default: per-packet payloads from every port, full flow exports on access links, debug syslog on all devices, and duplicate link-flap events already suppressed upstream. A 24-port switch (industry-typical) with debug and full export enabled on every port generates more bytes in one shift than the site's log partition is sized for, and the real severity 0–2 event then gets dropped when the disk is full. IEEE 802.3 fixes the untagged maximum frame at 1518 bytes; capturing all of them tells you nothing a counter did not already say. "Ignore" is a permanent setting here, not a lower sampling rate.

What does collection cost, and how long should you keep it?

Two budgets matter: control-plane CPU on the switch, and bytes on the log disk. Polling every counter at 5-second intervals (industry-typical) instead of 5-minute intervals multiplies the SNMP work by 60 for almost no extra insight; sampling at 1 in 2,048 (RFC 3176, industry-typical) instead of every packet is what keeps a flow feed affordable. Retention follows the tier, not the device. Keep severity 0–2 and security-relevant events for 12 months, routine warnings for 90 days, informational lines for 30 days and debug for 7 days (all industry-typical), and set NTP on every device, or the timestamps will not line up.

TierWhat it holdsRetentionAlert rule
Always onCounters, link state, severity 0–290 days to 12 months (industry-typical)Page on 0–2, ticket on 3–4
On demandMirrors, flow records, debug7 to 30 days (industry-typical)Fixed window, one device
IgnorePer-packet payloads, duplicate flapsNot collectedSuppress at source

Retention and poll intervals above are industry-typical; the severity bands are the eight values 0 to 7 that RFC 5424 defines.

FAQs

What is the smallest always-on set an OT site needs?

Three items: per-port error and utilisation counters, link state, and RFC 5424 §6.2.1 severity 0–2 syslog events. On a Rayin switch, as on any standards-based device, that is enough to tell whether the plant is healthy and what changed since the last shift. Everything packet-level belongs in the on-demand tier.

Why does a flapping port not page an engineer every time?

Because event volume, not severity, is what buries a real fault. A bad fiber contact can flap a port hundreds of times a minute (industry-typical). Suppress identical link-down events into one ticket over a short window, keep the count at level 4, and let the accumulated total drive triage in working hours.

How long should debug logs be kept?

Days, not months. Keep RFC 5424 §6.2.1 severity 0–2 events for 12 months and level 7 Debug for 7 days (industry-typical), then roll the findings into the ticket rather than a searchable index. Debug volume is what fills the disk, and without an NTP-aligned timestamp none of it is usable evidence.

Does continuous polling hurt a hardened switch?

It can. The control plane that answers SNMP polls also runs the ring and redundancy protocols. Polling a 24-port switch (industry-typical) every 5 seconds costs far more CPU than every 5 minutes for the same counters. Watch CPU utilisation while the poll rate climbs; if it moves, the interval is wrong for that device.

Can I keep per-packet captures permanently, just in case?

No. Full captures on every port cost more disk than an OT site budgets and rarely answer a question a counter cannot. Keep them in the on-demand tier: enable a port mirror for a fixed window on one device, attach the capture to a ticket, then stop. Standing feeds belong in the ignore tier.

Conclusion

The three tiers are the whole method: keep a short always-on list, open the heavy captures only for a named fault, and ignore the rest. The export format matters less than the retention and suppression rules around it. Set the clock with NTP, put severity 2 as the paging line, and the collector answers questions instead of generating new ones. That is what an OT collection plan is for.

Related Reading


How this article was made

The statements above come from public sources: RFC 5424 §6.2.1 (the eight severity values 0 to 7 and PRI = Facility × 8 + Severity), RFC 3176 (sampled flow export), RFC 2819 (the remote-monitoring MIB), RFC 2863 (the 32-bit interface counters), RFC 7011 (IP flow export), the IEEE 802.3 series (frame sizes and link rates), IEEE 802.1AB (LLDP advertisement timing) and the ITU-T G.984 series (GPON, including the 128-ONU registration ceiling). Only the RFC 5424 clause is numbered; the rest are cited by series or RFC number only. The three-tier method, worked examples, comparison tables and wording are written by Sara, who checked each figure against those sources before publication. Anything presented as an industry-typical value is labelled as such; nothing here is a measured figure from a Rayin product or a claim of any certification.

About the author: Sara — Sara is a Customer Manager at Rayin with over 10 years of experience in the communications field. In her free time, she enjoys badminton and swimming.

Connect with Sara on LinkedIn


About Rayin: Shenzhen Rayin Technology Co., Ltd. — Company Profile

Get A Quote

You have agreed to this website’s《Privacy Policy》