Firmware on network devices lives at two extremes: either the switch is forgotten for years on the version it was installed with, or after a security advisory it’s updated in a rush at midnight without any pre-checks. Both cause outages. Below we explain the parts of a repeatable policy in the order we apply them in the field: which version is chosen, how security advisories and life cycle dates are tracked, and in what order and with which safety nets the update is done.
Neither postponing updates nor rushing them
Network operating systems such as Cisco IOS XE, NX-OS, Juniper Junos and Aruba AOS-CX run SSH, HTTP(S) servers, SNMP and various automation interfaces on the management plane. A vulnerability in these components means direct compromise of a device whose management interface has been accidentally left reachable. The second cost of staying on an old version is quieter: the memory leak, the link flap with a particular SFP or the stack member dropping that the vendor fixed long ago keeps happening on your device. And you spend hours trying to diagnose a known bug.
Installing every new version immediately isn’t the answer either. The first maintenance releases of a new feature train carry code that hasn’t run enough in the field. The goal isn’t “the newest version”; it’s a supported and tested version that keeps known risk at an acceptable level. For this, three decisions are put in writing: which release train will be used, how often assessment will be done, and in which case an emergency update will be made without waiting for the regular cycle. If it isn’t written down, there is no policy.
Version selection: recommended release, long-lived train, new feature train
Most vendors separate releases by maintenance period. In Cisco IOS XE, some trains are marked “Extended Maintenance” and receive bug fixes for longer; support for “Standard Maintenance” trains ends earlier. The Cisco software download page also has releases marked “Suggested.” Juniper publishes a “JTAC Recommended” list per platform. The naming changes, the logic stays the same.
| Release type | When it’s preferred | Risk |
|---|---|---|
| Recommended / long-maintenance train | Core, distribution, WAN router | Low; limited new features |
| Standard-maintenance train | If a specific feature is needed | Medium; short train life |
| First release of a new feature train | Lab, pilot | High; code with little field use |
| Emergency security fix | If there’s no workaround for a critical vulnerability | Variable; limited testing |
The practical rule in the field: devices of the same role and model are kept on the same version, and a “golden image” list is written per role. For example, “access switches: version X; core: version Y.” The list is updated through a change record, and devices that deviate from the inventory are reported. When troubleshooting a network running three different versions in the same layer, separating which code a behavior difference comes from becomes very hard. Version sprawl prolongs diagnosis.
The sections to read in release notes are clear: the supported upgrade path (some versions are reached via an intermediate version), whether a ROMMON/bootloader or FPGA update is required, removed features and default behavior changes, and entries in the open caveats list that touch features you use. The last is the section you can skip.
Security advisories and EoS/EoL tracking
The source for security tracking is the vendor’s advisory page: for Cisco, PSIRT publications and the Cisco Software Checker tool; for Juniper, security bulletins. Ask the same four questions for every advisory. Is the affected feature enabled on our side? Are we on the affected release train? Is there a workaround (disabling the feature, restricting access with an ACL)? Is the vulnerability being actively exploited? For the last question, CISA’s Known Exploited Vulnerabilities catalog is a good second source; a vulnerability in the catalog that affects your device is grounds for an emergency update without waiting for the regular cycle.
On the life cycle side, vendors publish milestones for the product. According to Cisco’s EoS/EoL policy, the main ones are: End-of-Sale (the product is no longer sold), End of SW Maintenance Releases (no new software with bug fixes is released), End of Vulnerability/Security Support (no security fixes are released) and Last Date of Support (hardware and software support ends completely). In operations, the real date is the day security support ends. After that day, vulnerabilities that appear don’t get patches.
Don’t take dates from memory or third-party lists; open the vendor’s current EoL announcement for each model and record it in the inventory. The inventory should have model, serial number, current version, target version, support contract end and security support end date on the same row. Budget planning, device replacement and contract renewal can then be read from a single table.
Pre-update checks and image integrity
Before the update, the following information is collected and added to the change record. A current configuration backup is non-negotiable; for automatic and versioned backups, see the configuration backup with Oxidized article.
- Current version, boot variable and installation mode:
show version,show boot, and on IOS XEshow install summary - Hardware inventory and modules:
show inventory, and on a stackshow switch - Free space on flash:
dir flash:; if insufficient, clean up old packages - License status and whether the licensing model changes in the new version
- That the running configuration is saved: no difference between running-config and startup-config
- That console or out-of-band access has been tested
- A “before” output of adjacency and port status (to be used in verification)
Image integrity isn’t skipped. The vendor’s download page publishes a digest for each image (MD5 and SHA512 on Cisco). Verify the file first on the server you downloaded it to, then on the device after copying it:
# On the file server
sha512sum cat9k_iosxe.<version>.SPA.bin
# On the Cisco IOS / IOS XE device
copy scp://[email protected]/cat9k_iosxe.<version>.SPA.bin flash:
verify /md5 flash:cat9k_iosxe.<version>.SPA.bin
! With SHA512 if the platform supports it:
verify /sha512 flash:cat9k_iosxe.<version>.SPA.bin
Whichever algorithm you used, compare against the value the vendor published for the same algorithm. If the value doesn’t match exactly, the file isn’t used. A reload issued with an incompletely copied image can leave the device in ROMMON; at a remote location, that means someone going on site with a console cable. Download images only from the vendor’s official portal. An image from a forum or file-sharing site isn’t accepted even if the digest matches, because its source can’t be verified.
Staged updates on stacks, redundant cores and APs
The order determines the outage impact. The rule is simple: start with the least critical device that is easiest to roll back. First a lab or pilot device, then an access switch with few users, then the rest of the access layer, and distribution and core last.
Stacked access switches
A physical stack (Cisco StackWise, etc.) is a single management and control plane; all members must run the same version. On the Catalyst 9000 series, an update in install mode follows this flow:
install add file flash:cat9k_iosxe.<version>.SPA.bin activate commit
! After the device restarts
show version
show install summary
show switch
This command restarts all members of the stack together; every end device connected to the stack is cut off for the duration of the reload. Stack updates are maintenance window work. If the command is issued without commit (install add file … activate), the new version is activated but not made permanent; if install commit isn’t issued before the automatic abort timer expires, the device reverts to the previous version. For remote updates, this behavior is an extra safety net. Verify the exact behavior on your platform from the upgrade guide for the relevant version. To bring maintenance window approval, ownership and records into a single organization-wide process, change and patch management processes can be defined.
Redundant core: two separate control planes
In designs using MLAG, vPC or two independent core switches, devices are updated one at a time: the first device is updated, it’s verified while traffic flows through the second, and then you move on to the second. The prerequisite is that every end device is really connected to both cores. A surprise you may encounter is a server that is dual-homed on paper but whose second leg was removed years ago; when the core being updated restarts, that server is cut off. Check port-channel members one by one before the window. For the differences between a physical stack, MLAG and vPC, see the switch stack, MLAG and vPC comparison article. The version difference between the two devices should stay within the combination and duration the vendor supports.
Wireless access points
In controller-based wireless networks, AP software depends on the controller version; when the controller is upgraded, the APs download the new image. Many controllers support pre-downloading the image to APs and restarting APs in groups so that neighboring APs don’t go down at the same time. In cloud-managed systems, the update time is usually set per site. Check whether the default automatic update time falls within business hours.
Rollback plan and post-update verification
“If there’s a problem, we’ll go back to the old version” isn’t a plan. The plan states: is the old image still on flash, what is the rollback method (rollback in install mode; in bundle mode, pointing the boot variable to the old image and reloading), will the configuration converted by the new version work unchanged on the old version, and who makes the rollback decision based on which criterion. The rollback time is also counted within the maintenance window.
Verification is done by comparing against the outputs taken before the update:
show version | include uptime|Version
show interfaces status | include connected
show etherchannel summary
show spanning-tree root
show ip ospf neighbor
show cdp neighbors
show logging
The number of connected ports, port-channel members, the STP root bridge and routing adjacencies should be the same as before. If the logs show repeating errors, module warnings or license messages, they’re investigated before the window is closed. Define an observation period of at least one business day after the update; memory, CPU and interface error counters are compared with the previous week in the monitoring system. A memory leak doesn’t show in the first hour. For execution discipline, pre-registration and the go/no-go decision, the template in the network change window planning article can be used directly.
Mistakes to avoid
- Skipping the upgrade path: Going straight to the target version in a migration that requires an intermediate version; on some platforms the device may not boot due to bootloader incompatibility.
- Unsaved configuration: Unsaved changes being lost during the reload, and the resulting problem being wrongly attributed to the update.
- The whole network in one night: Updating the entire access layer on the same night without a pilot device; a common bug affects the whole building.
- Not checking SFP and module compatibility: The default behavior for third-party optics may have changed in the new version; the port doesn’t come up.
- Deleting the old image right away: Cleaning up old packages before the observation period ends makes rollback longer.
Annual firmware inventory checklist
This list is run at least once a year, preferably before the budget period, for all devices in the enterprise network infrastructure; security advisories are reviewed every month.
- Is the model, serial number, role, current version and target version of every device up to date in the inventory?
- Is the golden image list per role written down, and is the date of the last review known?
- Are devices deviating from the golden image reported, and is the reason for the deviation recorded?
- Has the security support end date for each model been taken from the vendor’s current announcement?
- Is there a replacement plan and budget request for devices whose security support will end within the next 18 months?
- Do support contracts cover the right to download software, and are their terms consistent with the inventory?
- Are the “not affected / workaround applied / updated” decisions recorded for the security advisories published in the last 12 months?
- Are a pilot device and a test procedure defined?
- Have console and out-of-band access paths been tested for each location?
For teams that want to bring this cycle into periodic maintenance, our network maintenance, monitoring and periodic checks service is the starting point; for commissioning new devices with the right version and a standard configuration from the start, our switch, router and network device installation service is.
Frequently asked questions
How often should switch firmware be updated?
Instead of a fixed calendar, use a two-tier model: security advisories are assessed every month, and regular version updates are planned by role once or twice a year. If there’s an actively exploited vulnerability that affects you, the regular cycle isn’t waited for.
Is the newest version always the most secure?
No. Even though the first release of a new feature train includes the latest fixes, it’s code that has seen little field use. For most production networks, the most recent maintenance release on the vendor’s recommended or long-maintenance train that fixes the relevant vulnerabilities is a more balanced choice.
What is the difference between EoS and EoL?
End-of-Sale is the date the product is no longer sold; the product still receives software and hardware support for a while. When support ends completely, the product reaches the end of its life cycle. The really critical date for operations is the date security fixes stop; it’s read per model from the vendor’s announcement.
Is it possible to update stacked switches without downtime?
A physical stack is a single control plane; a typical install mode update restarts all members together. Some platforms offer non-disruptive upgrades in limited scenarios, but support conditions depend on the platform and version. Plan stack updates as maintenance window work.