Summary

  • Revision 05 of the ASH proposal says an implementation sharing one PSNP transmission mechanism should acknowledge received repair LSPs before using that queue for further reconciliation requests.
  • Overlapping reconciliation attempts produced PSNP requests equal to roughly five times the fragment population in one draft experiment. The proposed control is one attempt per peer and IS-IS level, with completion withheld until every resulting transmission and retransmission finishes.

Recovery can starve its own receipts

Imagine two adjacent routers finding a coarse database mismatch. They exchange progressively finer PASH ranges, descend to ordinary SNP requests and start sending the missing LSPs. The receiving router now has two kinds of work for the same PSNP transmission path: acknowledge the repair LSPs that arrived, and ask for the next pieces that are still missing.

If the requests always win, the sender sees no receipt. Its retransmission timer expires. The same LSPs return, consuming the queue that had already failed to acknowledge their first copies. The network is not merely slow at recovery; recovery is producing more recovery traffic.

That is the sharp new point in revision 05 of IS-IS Aggregated SNP Hash Packets, dated 26 September 2026. The Datatracker record identifies the document as an individual Internet-Draft with no RFC stream and no formal IETF standing. Its header targets Experimental status. None of that proves adoption, implementation or operational benefit.

The changed line is about queue authority

ASH compresses database comparison. CASH packets summarize node ranges; PASH packets refine the ranges that disagree; ordinary SNP exchange or flooding repairs the final difference. The mechanism can reduce a very large comparison, but the expensive work returns when a mismatch must be repaired.

The revision 04 text already required one reconciliation attempt at a time for each peer and level. A new periodic CASH must wait while PASHes or PSNPs remain to be sent, or while resulting LSP transmissions and retransmissions remain outstanding. In the draft's roughly 6,700-fragment, 40-node experiment, overlapping attempts generated PSNP requests equal to about five times the total fragment count, and the count was still rising. The authors attribute the amplification to overlapping attempts rather than the topology scale or hash function.

Revision 05 adds the scheduling consequence. When acknowledgment entries and reconciliation entries share PSNP transmission machinery, acknowledgments for received LSPs should go first. Remaining capacity may carry repair requests. Delaying the receipt can make the remote timer expire, create duplicate LSP traffic and extend convergence.

The rule is more consequential than its modest SHOULD suggests. A request says what the receiver still wants. An acknowledgment tells the sender what it can stop sending. Under pressure, the second message can remove work from both systems; the first can add work. Queue order therefore controls whether the feedback loop damps or amplifies.

Three receipts must remain separate

BTW previously covered revision 01 in The Hashes Matched. That Did Not Prove the Network Was Right. That boundary remains intact. A matching ASH range is evidence that a specified comparison found no difference. It does not authenticate physical topology or prove forwarding.

This revision exposes a different boundary. An LSP acknowledgment is not a database-convergence certificate. It records progress for a received protocol object. Reconciliation completes only after the PASHes, PSNPs, LSP transmissions and retransmissions created by that attempt have completed. Even that completion does not prove SPF selected the intended route, a FIB installed it or user traffic crossed the path.

The operational record should therefore have three stages. First, a reconciliation receipt identifies peer, level, attempt, trigger, covered ranges, request backlog, acknowledgment backlog, LSPs outstanding, retransmission count, deferred CASH cycles and the exact completion transition. Second, a control-plane receipt records the resulting LSDB and route computation state. Third, a forwarding receipt tests the relevant traffic path. A green value from one stage must not be copied into the next.

The capability record also stays narrow. ASH uses the separate Hello Capability draft; ASH_RX and ASH_TX are directional and scoped to an adjacency. They say what a node is willing to receive or may transmit, not that a reconciliation completed safely. RFC 1195 supplies the wider integrated IS-IS context. RFC 5304 and RFC 5310 protect PDU integrity under configured keys without turning authenticated state into topology truth. RFC 9681 can accelerate flooding, but speed does not replace completion accounting.

Running-Code Primacy makes the test observable: did acknowledgments leave, did retransmissions stop, did the attempt close, and did forwarding recover? Minimum Initial Specification, Localized Future Decision supports a thin shared invariant—one attempt, receipt priority, explicit completion—while leaving implementation scheduling local. Reality Layers explains why capability, database agreement and packet delivery cannot borrow one another's authority.

Sources