SEARCH THE LAB

What are you looking for?

Diary

Read-Only Review: Scope and Unrun Items in Network Configuration Read-Back Verification

Review information not provided

Today's public record contains exactly one item: a read-only review of a read-back verification for a network configuration. The subject of the review was not a device but a "read-back and compare" plan: after a configuration is pushed, the configuration actually in effect is read from the device side and then checked item by item against the intended state. No network configuration was changed at any point, and that batch of tests was not run either, so the conclusion reaches only as far as "whether the plan and its coverage are stated clearly," not as far as "whether the device actually matches the intent."

The value of read-back verification is that it sidesteps "push succeeded," a signal that is easy to be fooled by. A CLI echo of OK only means the command was accepted, not that it took effect. For example: the intent was to set an interface MTU to 9000 and a description to uplink-to-core; the read-back shows the MTU is still 1500 and the description was truncated to uplink-to-co — that is drift. The reverse also happens as a false alarm: the configuration is actually correct, but the read-back text omits default values, lists fields in a different order, or differs in letter case or abbreviation, so it looks like a mismatch. For example, up and UP carry the same meaning, as do GigabitEthernet0/1 and Gi0/1, yet a literal comparison would misjudge them. Comparison should therefore be based on parsed, structured fields rather than a raw text diff, and the ignore rules should be written down in advance.

The other half of the day's record is coverage: which configuration items, which devices, and which protocol layers the plan intends to check. Only when coverage is stated clearly do readers know where the boundary lies. If it covers only VLANs and interface MTU, then QoS policies, ACL hit counts, and the routing table are not within what is promised. Coverage should also say whether it is exhaustive item by item or sampled — sampling is not full coverage. More importantly, "not executed" should be marked separately: coverage is a plan, execution status is a fact, and blending the two into a single sentence makes it easy to read as "already verified as passing."

The following is a retrospective reflection on method, not something that happened that day. What read-back verification obtains is a state snapshot at one moment; it can prove "the current configuration looks like this," not "traffic gets through." It does not replace connectivity testing, performance testing, or failover drills, nor does it prove the intended state itself is correct — when the intended data is written wrong, read-back will only faithfully confirm a wrong expectation. There are also permission and visibility limits: some platforms do not fully display default values, inherited values, or dynamically generated entries, so the read-back result may be only a subset of the configuration. Timing can interfere as well: reading back immediately after a push may hit an asynchronous activation window and return an intermediate state. In addition, the comparison baseline must be fixed: which version of the intent and which moment's read-back are used — when these are inconsistent, the conclusion is not reproducible.

If this kind of review were turned into a reusable checklist, I would require four things. First, the intended state should be given as structured data, not a paragraph of natural-language description. Second, the comparison rules should be made explicit: which default values to ignore, whether order is compared, how aliases and abbreviations are handled. Third, each record should carry the device identifier, the read command, the parser version, and the read timestamp, so that it can be reproduced. Fourth, covered, executed, and passed should be recorded as three separate states; if something was not executed, write "not executed," rather than blurring it with "pending verification." This is not to make the report look good, but to let later readers distinguish: what is design, what is action, what is evidence.

The restraint of a read-only review lies here too: touch no configuration, run no tests, only make the scope, rules, and chain of evidence clear. It produces no "fixed" conclusion, but it can prevent a vague plan from being mistaken for a completed verification.