Predictika · Arista EOS fabric changes

An AI agent's fabric change, checked before it is pushed

An AI or AIOps agent proposes a batch of changes to an Arista EOS leaf-spine data center fabric: configuration changes, EOS software upgrades, route policy changes, optic enablement and link shutdowns. Before anything is pushed, the batch goes to the Predictika validator, which checks it against the fabric graph (DCS-7060X6 leaves, DCS-7808R4 spines, DCS-7280R4 border leaves, their MLAG pairs and the EOS release each runs), Arista's published minimum EOS releases for 800G optics, and the customer's own change policy, and returns every finding in one answer. The validator runs on the graph a fabric management tool exports, with the customer's own policy. The fabric, the tickets and the change policy on this page belong to a fictional customer, Cedar Basin AI Cloud; the switch models, optics and their EOS minimums come from Arista's public documents.

Scope: change batches for leaf-spine data center fabrics with leaves, spines, border leaves and MLAG pairs. See everything this validator covers

This validator checks one kind of request: a batch of proposed changes to a leaf-spine fabric, from an AI agent, an AIOps tool or a person, against the fabric graph and a change policy. It handles any batch of such changes on the devices in the graph, not only the scenarios shown on this page. It does not generate configurations and does not push anything.

  • Covered: redundancy — both leaves (or both border leaves) of an MLAG pair in one batch are blocked; the pair comes from the graph, not from the batch. An EOS upgrade that would leave the two peers more than two minor releases apart is blocked, in either direction.
  • Covered: optics and EOS releases — enabling an 800G optic on a switch that does not yet run the EOS release the optic needs is blocked, and an upgrade in the same batch does not count; linear pluggable optics (LPO) on a switch Arista does not list for LPO are blocked, and where Arista's documents disagree the change goes to a person.
  • Covered: scope, rollback and window — a fabric-wide change with no rollback reference, or outside the approved change window, is blocked; a narrower change outside the window needs approval.
  • Covered: agent actions on live state — an agent action to shut a link whose optic is degraded while the redundant path is already down is blocked, because it would take the endpoint off the fabric; with the redundant path up, the same action is allowed.
  • Covered: approvals and velocity — border leaf changes need a minimum approval rank on the ticket, and a limit on spine and border leaf changes per window sends the batch to a person.
  • Supplied by the customer's own systems, not by the validator — the fabric graph, the ticket approvals, the change policy values, whether each change is inside the approved window, the rollback reference, and the optic and redundant-path state from monitoring.
  • Not covered — configuration syntax, traffic or routing simulation, performance modeling, capacity, and any rule the customer has not written into its policy. Other rules can be added.
The AI agentProposes the batch of fabric changes and says it is ready to push.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

8 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI agent says

Illustrative agent statement. The validator receives the structured request below.

Proposing agent: AI agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      
Waiting for the call.

Where rules-based change approval gets it wrong

Click to expand

Our assessment of what a typical rules-based change approval workflow (ticket fields, approval chains and validation rules) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.

What the validator caughtTypical rules-based change approvalPredictika
An optic that needs a newer EOS than the switch runsOnly with workarounds
Whether an optic works depends on the release the switch runs today, not the one a batch will leave it on. Ticket rules do not know either, so the check is done by an engineer reading release notes, or not at all.
The running release comes from the fabric graph, the optic's minimum from the catalog, and an upgrade in the same batch is correctly not counted.
MLAG peers on EOS releases too far apartOnly with care
An upgrade ticket names one switch. Whether it leaves its MLAG peer too far behind depends on the peer's release and a policy limit, which approval rules rarely hold side by side.
The target release is compared with the peer's running release from the graph, in either direction, against the customer's limit.
Both halves of a redundant pair in one batchOnly with workarounds
A change-approval workflow looks at one change at a time. Seeing that two changes in the same batch hit the two halves of a pair needs a view across the whole batch and the pairing from the topology, which is usually done in a script outside the workflow.
The whole batch is evaluated together against the graph; both changes are reported, each naming its partner.
Fabric-wide scope that depends on a rollback and a windowOnly with care
Approval rules can require a rollback or a window for a change type, but scope has to be set correctly on each ticket, and a rule per combination grows quickly.
Scope, rollback reference and window are checked together for each change in the batch; partial fixes show exactly what is still missing.
An agent action that depends on live stateOnly with workarounds
Whether it is safe to shut a degraded link depends on the state of the redundant path at that moment. Ticket rules do not see telemetry, so the check ends up inside the agent or not at all.
The monitoring state travels with the change as a field, so the same shutdown is blocked with the redundant path down and allowed with it up.
Approval rank by device role, plus a rate limitOnly with care
A rank requirement per device role and a limit on core changes per window are separate thresholds that approval chains handle one at a time.
Both are checked on every batch and reported together, with the numbers behind them.

What this validator catches: findings from its test cases

Click to expand

Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the batch; approval findings need a person to sign off. The wording shown is an example from the test runs; device names and numbers change with the batch. The policy values come from a fictional customer.

1Both halves of a redundant pair in one batch (blocks): change item-02 on leaf-201b and its MLAG peer leaf-201a are both in this batch. Change policy: the two halves of a redundant pair are changed in separate batches.
2Degraded link shut while the redundant path is down (blocks): change item-01 would shut a link on leaf-201a whose optic is degraded, but the redundant path is already down. Shutting it would take the endpoint off the fabric; restore the redundant path first.
3Device not in the fabric graph (blocks): change item-01 names leaf-999z, which is not a device in fabric cedarbasin-ai1.
4Fabric-wide change without a rollback (blocks): change item-03 is fabric-wide and has no rollback reference. Change policy: stage a rollback before a fabric-wide change.
5LPO optic on a switch Arista does not list for LPO (blocks): change item-04 enables LPO-800G-DR8, a linear pluggable optic, on border-leaf-1a, a DCS-7280R4-32PE-F. Arista lists LPO support only on the 7060X6-64PE and the 7800R4 line cards.
6MLAG peers would run EOS releases too far apart (blocks): change item-01 upgrades leaf-202a to EOS 4.36.0, but its MLAG peer leaf-202b runs EOS 4.32.2. Change policy: the two peers of an MLAG pair may not run EOS releases more than two minor versions apart.
7Optic needs a newer EOS release than the device runs (blocks): change item-02 enables OSFP-800G-VSR8 on leaf-202a, which runs EOS 4.32.2; this optic needs EOS 4.35.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release.
8Arista documents disagree on LPO support (needs approval): change item-01 enables LPO-800G-2FR4, a linear pluggable optic, on leaf-203a, a DCS-7060X6-32PE. Arista documents disagree on LPO support for this switch, so a person must confirm it.
9Border leaf change needs higher approval (needs approval): change item-01 is on border leaf border-leaf-1a and ticket chg-8102 carries rank 1 approval; change policy requires rank 3 for border leaf changes.

Scenarios and results

Click to expand

The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

1. Software releases and optics: Upgrade and optics on a leaf pair, a fabric-wide change, an LPO optic on a border leaf

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8103: an upgrade of leaf-202a to EOS 4.36.0 (rollback rb-9101), the enabling of an OSFP-800G-VSR8 optic on leaf-202a (rollback rb-9102), a fabric-wide route policy change through spine-1, and the enabling of an LPO-800G-DR8 optic on border-leaf-1a (rollback rb-9104). All of it is in tonight's approved change window. One spine or border leaf change has already been made in this window. It is ready to push.

Verdict: Blocked.

  • MLAG peers would run EOS releases too far apart. Change item-01 upgrades leaf-202a to EOS 4.36.0, but its MLAG peer leaf-202b runs EOS 4.32.2. Change policy: the two peers of an MLAG pair may not run EOS releases more than two minor versions apart.
  • Optic needs a newer EOS release than the device runs. Change item-02 enables OSFP-800G-VSR8 on leaf-202a, which runs EOS 4.32.2; this optic needs EOS 4.35.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release.
  • Fabric-wide change without a rollback. Change item-03 is fabric-wide and has no rollback reference. Change policy: stage a rollback before a fabric-wide change.
  • Optic needs a newer EOS release than the device runs. Change item-04 enables LPO-800G-DR8 on border-leaf-1a, which runs EOS 4.34.2; this optic needs EOS 4.36.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release.
  • LPO optic on a switch Arista does not list for LPO. Change item-04 enables LPO-800G-DR8, a linear pluggable optic, on border-leaf-1a, a DCS-7280R4-32PE-F. Arista lists LPO support only on the 7060X6-64PE and the 7800R4 line cards.

2. Software releases and optics: Upgrade target lowered and rollback added; two optics still wrong

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8103: an upgrade of leaf-202a to EOS 4.34.2 (rollback rb-9101), the enabling of an OSFP-800G-VSR8 optic on leaf-202a (rollback rb-9102), a fabric-wide route policy change through spine-1 (rollback rb-9103), and the enabling of an LPO-800G-DR8 optic on border-leaf-1a (rollback rb-9104). All of it is in tonight's approved change window. One spine or border leaf change has already been made in this window. It is ready to push.

Verdict: Blocked.

  • Optic needs a newer EOS release than the device runs. Change item-02 enables OSFP-800G-VSR8 on leaf-202a, which runs EOS 4.32.2; this optic needs EOS 4.35.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release.
  • Optic needs a newer EOS release than the device runs. Change item-04 enables LPO-800G-DR8 on border-leaf-1a, which runs EOS 4.34.2; this optic needs EOS 4.36.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release.
  • LPO optic on a switch Arista does not list for LPO. Change item-04 enables LPO-800G-DR8, a linear pluggable optic, on border-leaf-1a, a DCS-7280R4-32PE-F. Arista lists LPO support only on the 7060X6-64PE and the 7800R4 line cards.

3. Software releases and optics: Optics that the running releases support

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8103: an upgrade of leaf-202a to EOS 4.34.2 (rollback rb-9101), the enabling of an OSFP-800G-2FR4 optic on leaf-202a (rollback rb-9102), a fabric-wide route policy change through spine-1 (rollback rb-9103), and the enabling of an OSFP-800G-2FR4 optic on border-leaf-1a (rollback rb-9104). All of it is in tonight's approved change window. One spine or border leaf change has already been made in this window. It is ready to push.

Verdict: Allowed.

4. Redundant pairs: Upgrade of both MLAG peers

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8101: an upgrade of leaf-201a to EOS 4.35.0 (rollback rb-9201) and an upgrade of leaf-201b to EOS 4.35.0 (rollback rb-9202). Both are single-device changes in tonight's approved change window. No spine or border leaf change has been made in this window yet. It is ready to push.

Verdict: Blocked.

  • Both halves of a redundant pair in one batch. Change item-01 on leaf-201a and its MLAG peer leaf-201b are both in this batch. Change policy: the two halves of a redundant pair are changed in separate batches.
  • Both halves of a redundant pair in one batch. Change item-02 on leaf-201b and its MLAG peer leaf-201a are both in this batch. Change policy: the two halves of a redundant pair are changed in separate batches.

5. Redundant pairs: Upgrade one peer, change another pair

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8101: an upgrade of leaf-201a to EOS 4.35.0 (rollback rb-9201) and a configuration change on leaf-203a (rollback rb-9202). Both are single-device changes in tonight's approved change window. No spine or border leaf change has been made in this window yet. It is ready to push.

Verdict: Allowed.

6. AIOps link shutdown: Shut a degraded link, redundant path down

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8102: shut the interface on leaf-201a, whose optic monitoring has flagged as degraded (rollback rb-9301), and make a configuration change on leaf-203a (rollback rb-9302). Both are in tonight's approved change window. One spine or border leaf change has already been made in this window. It is ready to push.

Verdict: Blocked.

  • Degraded link shut while the redundant path is down. Change item-01 would shut a link on leaf-201a whose optic is degraded, but the redundant path is already down. Shutting it would take the endpoint off the fabric; restore the redundant path first.

7. AIOps link shutdown: Redundant path restored

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8102: shut the interface on leaf-201a, whose optic monitoring has flagged as degraded (rollback rb-9301), now that the redundant path for that link has been restored, and make a configuration change on leaf-203a (rollback rb-9302). Both are in tonight's approved change window. One spine or border leaf change has already been made in this window. It is ready to push.

Verdict: Allowed.

  • Redundant path is up. Change item-01 shuts a link on leaf-201a while its redundant path is up; traffic moves to the redundant path.

8. Approvals: LPO optic where Arista documents disagree

The agent's request. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8101: the enabling of an LPO-800G-2FR4 optic on leaf-203a, a DCS-7060X6-32PE (rollback rb-9401), and a configuration change on spine-2 (rollback rb-9402). Both are single-device changes in tonight's approved change window. No spine or border leaf change has been made in this window yet. It is ready to push.

Verdict: Needs approval.

  • Arista documents disagree on LPO support. Change item-01 enables LPO-800G-2FR4, a linear pluggable optic, on leaf-203a, a DCS-7060X6-32PE. Arista documents disagree on LPO support for this switch, so a person must confirm it.

Narration transcript

Click to expand

The text of the narrated walkthrough, step by step.

  1. Welcome to a demo of Predictika's Validator. An AI or AIOps agent proposes a batch of changes to an Arista EOS leaf-spine data center fabric: configuration changes, EOS software upgrades, route policy changes, and more. Before anything is pushed, the batch goes to the Predictika validator, which checks it against the fabric graph (DCS-7060X6 leaves, DCS-7808R4 spines, DCS-7280R4 border leaves, their MLAG pairs and the EOS release each runs), Arista's published minimum EOS releases for 800G optics, and the customer's own change policy, and returns every finding in one answer. The validator runs on the graph a fabric management tool exports, with the customer's own policy. The fabric, the tickets and the change policy on this page belong to a fictional customer, Cedar Basin AI Cloud; the switch models, optics and their EOS minimums come from Arista's public documents.
  2. I will start with the first request by the agent.
  3. The agent says. Please push this batch on the Cedar Basin AI Cloud fabric under ticket chg-8103: an upgrade of leaf-202a to EOS 4.36.0 (rollback rb-9101). All of it is in tonight's approved change window. The rest of the request is on screen.
  4. Before it goes out, the agent calls the validator.
  5. Blocked, with five findings, all in one answer. Change item-01 upgrades leaf-202a to EOS 4.36.0, but its MLAG peer leaf-202b runs EOS 4.32.2. Change policy: the two peers of an MLAG pair may not run EOS releases more than two minor versions apart. Change item-02 enables OSFP-800G-VSR8 on leaf-202a, which runs EOS 4.32.2; this optic needs EOS 4.35.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release. Change item-03 is fabric-wide and has no rollback reference. Change policy: stage a rollback before a fabric-wide change. The other two are on screen.
  6. The validator returned five blocking findings in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the batch are checked against each other.
  7. What this shows: for MLAG peers on EOS releases too far apart, a rules-based change approval is error-prone for this.
  8. Now the agent resubmits.
  9. The corrected request. It changes the item 1 to Software upgrade on leaf-202a, one device, rollback rb-9101, inside the change window; target EOS 4.34.2 and the item 3 to Route policy change on spine-1, fabric-wide, rollback rb-9103, inside the change window.
  10. It calls the validator again.
  11. Still blocked, with three findings, all in one answer. Change item-02 enables OSFP-800G-VSR8 on leaf-202a, which runs EOS 4.32.2; this optic needs EOS 4.35.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release. Change item-04 enables LPO-800G-DR8 on border-leaf-1a, which runs EOS 4.34.2; this optic needs EOS 4.36.0 or later. An upgrade in the same batch does not count: the optic can be enabled only once the device is running the release. Change item-04 enables LPO-800G-DR8, a linear pluggable optic, on border-leaf-1a, a DCS-7280R4-32PE-F. Arista lists LPO support only on the 7060X6-64PE and the 7800R4 line cards.
  12. Two findings cleared. Still blocking: optic needs a newer EOS release than the device runs and LPO optic on a switch Arista does not list for LPO.
  13. The agent resubmits once more.
  14. The revised request. It changes the item 2 to Enable an optic on leaf-202a, one device, rollback rb-9102, inside the change window; optic OSFP-800G-2FR4 and the item 4 to Enable an optic on border-leaf-1a, one device, rollback rb-9104, inside the change window; optic OSFP-800G-2FR4.
  15. It calls the validator once more.
  16. This time the validator clears it.
  17. The validator also returns the line-item quote. It starts with Software upgrade on leaf-202a and Enable an optic on leaf-202a; the other two lines are on screen.
  18. Both earlier findings are cleared.
  19. Five problems caught in one call, fixed in two passes, and a clean batch. That is the value of Predictika's validator: catching mistakes before the change is made. Now you can try the other scenarios yourself.