Routing model & coverage¶
This page is the reference for how the node resolves an inbound message, the MAP/CAP operations it recognises, and the runtime features that keep a signalling node honest: availability-driven failover, Service-Indicator- agnostic transit, and the loop guards.
The decision flow¶
Every inbound message runs the same synchronous Rust path, top to bottom, first answer wins:
- MTP3 transfer. If the destination point code is not one of the node's own, the message transits. The route resolver picks the egress and the MSU leaves. No SCCP decode, no Python. This is the cheapest outcome and the common one for an STP.
- Content routing. For a message addressed to us carrying SCCP + TCAP, the
decoded MAP/CAP view is evaluated against the ordered content rules
first, because they decide on the richest layer. A rule routes,
rewrites the called-party GT, screens, or defers to a hook. A
BeginorContinuethat cannot be decoded in full is dropped here instead of being routed unseen (details). - SCCP GTT. Otherwise the called-party global title is translated (after
an E.214 to E.164 pre-step) to a concrete
(dpc, ssn), a group member, or local termination. - Termination. A result of
local(or a called SSN the node owns) hands the message to the dialogue engine.
The resulting decision is one of: forward via an egress destination, route to a
concrete (dpc, ssn), terminate locally, defer to a named Python hook, or drop
with a reason.
The cost ladder¶
Route at the cheapest layer that can answer. Rising cost:
| Decision on | Layer | Cost (single core) |
|---|---|---|
| destination point code | MTP3 route table | ~28 ns |
| GT prefix (+ gti/tt/np/nai) | SCCP GTT | ~40 ns |
| home network (MCC+MNC) | E.214 conversion + GT prefix | SCCP; still a table transform |
| full IMSI / operation / GT-table membership | content routing | ~50 ns (needs the MAP/CAP decode) |
| a live dip the tables can't answer | a Python hook | yours |
PLMN-level steering belongs at SCCP: the roaming subscriber's MCC+MNC is the leading digits of the E.214 called party, so it is a GTT prefix rule, not a content rule. Keep content rules for decisions that genuinely need the decoded IMSI or operation. See the cost ladder in Concepts.
Availability and failover¶
A route is usable only while its destination is available. The resolver picks the lowest-priority available route for a DPC and fails over as state changes. Availability is driven from the live transport:
- M3UA: an AS is a route while it is active, which it becomes when
required_aspsof its ASPs are active, and while it is pending: forrecovery_msafter its last active ASP left, its traffic is queued (RFC 4666 4.3.2). Load-share spreads over the active ASPs by SLS; override sends to the ASP that became active last; broadcast sends to all active. See M3UA management. - M2PA: a linkset is a route while level 3 routes traffic to at least one of its links: a link that is in service, has passed its link test, is not inhibited, and whose adjacent point has said traffic may restart. SLS spreads traffic across those links.
- Route management: a transfer-prohibited from the adjacent point of a linkset takes that one route to the destination out; a transfer-allowed puts it back; a transfer-restricted makes it the last choice among routes of its priority.
- Application servers: a destination-unavailable from an ASP prohibits a DPC, a destination-available allows it.
When the primary route for a DPC goes unavailable the resolver uses the next priority automatically; an alternate linkset behind an AS route (Configuration) is the common failover shape. Routes of the same priority share the traffic by SLS; when one of them goes, only its selection codes move.
Signalling network management¶
A node with linksets is a signalling transfer point, and its neighbours expect
ITU-T Q.704 of it. One task per routing domain runs the procedures below. All
their timers are keys of the mtp3 block. Messages
are the ITU 14-bit formats of Q.704 clause 15; a management message for this
node (service indicator 0) arriving on a link goes to this task, one for
another point code is transferred like any other message.
A link fails (changeover, clause 5). Traffic for the link is buffered. An
extended changeover order goes to the far end on another link (or route) with
the sequence number of the last message accepted on the failed link: M2PA
numbers are 24 bits, so RFC 4165 4.2.3 has XCO and XCA of Q.2210 take the place
of COO and COA. On the acknowledgement, the messages the far end does not have
are sent again on the alternative links, then what was never sent, then the
buffer, in that order: none lost and none twice. An order from the far end is
acknowledged the same way, also for a link that still looked fine here. Without
an acknowledgement in T2, with an emergency acknowledgement, or when nothing
leads to the far end (T1), only what was never sent goes out again: what was
sent and never acknowledged might be a duplicate, so it is not, and is counted
in sigtran_m2pa_discarded_total{reason="tx-lost-on-failure"}. A 7-bit COO is
answered with ECA, since its number cannot name an M2PA message.
A link returns (changeback, clause 6). In a linkset that carries traffic, the link's traffic is buffered, a changeback declaration goes out on every link that carried it meanwhile, and the buffer goes out on the link once each is acknowledged (or after T4 and T5). A linkset that returns as a whole takes its traffic back after T3, or at once when nothing has to move. A declaration from the far end is always acknowledged.
An adjacent point prohibits or allows a route (clauses 7, 8, 13). A transfer-prohibited marks the route and the next message takes the alternative (forced rerouting; the alternative is known the moment the route is marked, so nothing waits). The route is then tested with a signalling route set test every T10 until a transfer-allowed arrives. A transfer-allowed or transfer-restricted that moves traffic between routes holds the destination's messages for T6 first (controlled rerouting), so the ones still on the old route arrive before the first on the new one.
What this node tells its neighbours (clause 13). One rule: an adjacent point is told transfer-prohibited for a destination when the destination is inaccessible from here (the broadcast method, 13.2.2 ii), or when this node routes to it through that very point (13.2.2 i); transfer-allowed otherwise. Whatever changes (a link, a linkset, an application server, a received transfer-prohibited), the difference to what each point was last told is sent. A message for a destination that cannot be reached is also answered with transfer-prohibited toward the point it came from, once per T8 (the response method, 13.2.2 iii), in addition to whatever SCCP returns. A route set test from a neighbour is answered when what it believes differs from what is (13.5.4). Transfer-restricted is acted on and never sent.
Restart (clause 9). When the first link comes into service after the node was isolated (at start, or for longer than T1), it restarts: it collects transfer-prohibited messages and traffic restart allowed from its adjacent points for at most T18, discards traffic from the linksets meanwhile, then sends its own transfer-prohibited messages and a traffic restart allowed to each. When one adjacent point returns, it is sent what it must not route here and a traffic restart allowed, and no traffic goes to it until its own arrives or T21 is over. An unexpected traffic restart allowed is answered once per T19.
Inhibiting (clause 10). A request to inhibit a link is granted (the link leaves traffic after T1, and is tested every T23) unless a destination would become inaccessible, in which case it is denied. Uninhibit, force uninhibit and the inhibit tests are answered as 10.3, 10.4 and 10.6 have it. This node never asks to inhibit a link itself: there is no management interface to ask it to.
Congestion. A transfer-controlled is counted and published; routing does not change, and none is sent.
Not done. Transfer-restricted and transfer-controlled are never
originated; the route set congestion test, the allocation of signalling data
links (clause 12.3 on), uninhibiting on the initiative of routing (10.3.2) and
user part unavailable are not run: such a message for this node is counted in
sigtran_mtp3_discarded_total{reason="management-not-acted-on"}. The formats
are ITU only. Toward application servers nothing of this is translated yet: a
destination that becomes inaccessible is announced to adjacent signalling
points, not to ASPs. The registry publishes every such change for whatever
does that (Registry::destination_changes).
Order. A message is held, sent again or diverted by the task of the link it was handed to, and that task is also what tells routing to lead to the link or to pass it by: it does so after it knows what to do with the traffic, and after it has passed on what it held. So the messages of one link keep their order through a changeover and a changeback, and a changeback declaration travels in the queue of the alternative link, behind the traffic already handed to it. What is left is the instant between a message choosing its link and being handed to it: one that chose a link just as routing stopped leading there is passed on by that link's task and can be overtaken by the next one, and one that chose the alternative link just as a changeback began can follow the declaration. Both need a link to change state inside that instant.
M3UA management¶
Toward the ASPs of its application servers the node is a signalling gateway as
RFC 4666 has one. One task per routing domain keeps the state of every ASP in
every server (4.3.1) and of every server (4.3.2); the keys are in
application_servers and
the m3ua block.
- ASP state. ASP Up and ASP Down are always acknowledged. ASP Active and ASP Inactive are checked: the routing contexts against the servers the association is configured in, the traffic mode against the server's.
- Notify (4.3.4.5). Every ASP of a server that is not down is told when the server becomes inactive, active or pending, after the acknowledgement that caused it; an ASP that comes up is told the state the server is in; an ASP of an override server that loses the traffic is told who took it; the failure of an ASP is told to the others.
- The pending server. See
recovery_ms: T(r), the queue, and what happens to it. - Destinations. What level 3 learns about a destination reaches every active ASP: DUNA when it becomes inaccessible, DAVA when it is accessible again, DRST when it is reached by a restricted route only, SCON for a transfer-controlled (4.5.1). An ASP that becomes active is sent the destinations that are unavailable or restricted at that moment. A server is not told about its own point codes. A DAUD is answered per destination from the routing tables, a range (a mask) for every destination of the table in it. While this node's MTP restart runs every destination behind a linkset is unavailable to the ASPs, also in answer to an audit, and DAVA follows when the restart ends (4.6).
- Between an ASP and a linkset (1.4.3.2). Signalling network management is never carried in DATA. A DUNA or DAVA from an ASP changes the route state and goes to the adjacent points as transfer-prohibited or transfer-allowed. A transfer-controlled or a user part unavailable message that the network addressed to the point code of a server reaches its ASPs as SCON or DUPU. A SCON from an ASP that names a Concerned Destination is sent there as transfer-controlled (3.4.4). Any other management message for the point code of a server, and one an ASP puts in DATA, is discarded and counted.
- Errors (3.8.1). A message that cannot be read or acted on is answered
with the Error that says why, with the first 40 octets of it as Diagnostic
Information, and counted in
sigtran_rx_discarded_total. An Error is never answered.
On a client association the node is the ASP: it sends ASP Up and, for every
server the association is in, ASP Active with the server's traffic mode and
routing context, again every ack_ms until acknowledged. It becomes inactive
when the gateway says an alternate ASP is active, and asks to be active again
when the gateway says the server is pending, inactive, or short of ASPs.
Not done: routing key management (REG REQ and DEREG REQ are answered with Error, "Unsupported Message Class"), the Network Appearance parameter, the IPSP models, a check that ASP Identifiers are unique, and a refusal of DATA on stream 0 (it is counted and accepted). SUA associations keep the plain handshake they had: none of the above applies to them.
Service-Indicator-agnostic transit¶
MTP3 transfer routes by point code for any Service Indicator. SCCP (SI 3) is decapsulated up to GTT / TCAP / termination when the node is the destination; a non-SCCP MSU (call control and other SIs) that is not addressed to the node transits by DPC alone, with no codec for its upper layer needed. An STP relays those for free.
Loop guards¶
An STP that loops is a signalling storm, so transit carries runtime guards. Each
drops the offending MSU and counts it in
sigtran_loops_detected_total{kind=...}:
- own-opc: the originating point code of a transit MSU equals the node's own point code, so a message the node originated has come back. Dropped.
- route-reflect: the only available route would send the MSU back out the linkset it arrived on. Dropped rather than reflected.
- hop-counter: an SCCP XUDT/LUDT carries a Q.713 hop counter, decremented at each GTT relay; at zero the message is a translation loop, dropped and returned as an XUDTS/LUDTS "hop counter violation" when it asked to be returned on error. The translation that leads to a subsystem of this node counts as one (Q.714 2.3.1): an XUDT addressed by global title that arrives with one hop left is dropped the same way and not terminated. One addressed by subsystem number is not translated, and its hop counter is not looked at.
The first two catch route-config loops that no upper-layer mechanism sees (call-control SIs have no application-level hop counter); the hop counter is the SCCP-layer complement on the GTT relay path. Because they run on the transit path, they cost nothing on the normal case and show up on a graph when a route table is wrong.
Admission of new dialogues¶
A composing binary can set an admission policy on the router: a Rust function that sees each new TCAP dialogue the node relays or terminates and admits or refuses it. It runs on the per-message path in Rust, never in Python. With no policy set, everything is admitted.
use std::sync::Arc;
use siphon_sigtran::admission::{Admission, DialogueStart, RefuseCause};
let router = Router::new(&cfg).with_admission_policy(Arc::new(
|start: &DialogueStart<'_>| {
if accepting() {
Admission::Admit
} else {
Admission::Refuse(RefuseCause::new("closed"))
}
},
));
(siphon_sigtran::python::task_with_admission(cfg, policy) is the same for the
addon's runtime task.) The policy is shown the TCAP message type, the calling
and called party (global title, SSN, point code), the MTP3 OPC/DPC, the
application context name when the dialogue portion carries one, the routing
domain, and whether the dialogue terminates at this node.
Only a TCAP Begin or Unidirectional is put to the policy. Continue,
End and Abort belong to a dialogue that was already admitted and always
pass, so a relaying node needs no dialogue state. Non-TCAP SCCP traffic (SCCP
management among it) and every other Service Indicator always pass.
What a refusal does:
- Relayed (MTP3 transfer of an SCCP message, or GTT to another point
code): the SCCP message return procedure (ITU-T Q.714 4.2). With the "return
message on error" option set, the caller gets a UDTS / XUDTS / LUDTS with
return cause "unqualified" (Q.713 3.12,
0x07);RefuseCause::with_return_causepicks another. Without the option the message is discarded. - Terminated here: a
Beginis answered with an Abort, P-Abort cause "resource limitation" (Q.773, Q.772 2.3.5), as for aBeginover themax_dialoguesceiling. AUnidirectionalhas no transaction to abort and is discarded.
Every refusal is counted in sigtran_admission_refused_total{reason,action},
action being returned, discarded or aborted, and logged at warn level.
MAP/CAP operations¶
The operation names a content rule can match (operation:), and the
subset with a termination decorator. Names are kebab-case; an unknown name is a
config load error.
| Operation | Match name | Terminate with |
|---|---|---|
| SendRoutingInfoForSM | sri-sm |
@gsm_map.on_operation("sri-sm") |
| MO-ForwardSM | mo-forward-sm |
@gsm_map.on_operation("mo-forward-sm") |
| MT-ForwardSM | mt-forward-sm |
@gsm_map.on_operation("mt-forward-sm") |
| updateLocation | update-location |
@gsm_map.on_operation("update-location") |
| cancelLocation | cancel-location |
@gsm_map.on_operation("cancel-location") |
| sendAuthenticationInfo | send-auth-info |
@gsm_map.on_operation("send-auth-info") |
| insertSubscriberData | insert-subscriber-data |
(staged as an invoke leg) |
| provideSubscriberInfo | provide-subscriber-info |
@gsm_map.on_operation("provide-subscriber-info") |
| initialDP (CAMEL/INAP) | initial-dp |
@gsm_cap.on_operation("initial-dp") / @inap.on_operation("initial-dp") |
| connect (CAMEL) | connect |
(staged as an invoke via gsm_cap.connect) |
The match names above are the content-rule vocabulary (operation:). The
termination vocabulary is a per-namespace superset (it adds ready-for-sm,
purge-ms, report-sm-delivery-status, event-report-bcsm, and the INAP
operations); the full list each on_operation accepts is in the
Script API. One handler can take several operations
pipe-separated (@gsm_map.on_operation("mo-forward-sm|mt-forward-sm")), the same
on_<message>("<name>") shape as @smpp.on_pdu and the SIP proxy's
@proxy.on_request.
Operations staged only as an invoke (insert-subscriber-data, connect) are not
terminated; you send them inside an open dialogue. Anything you don't terminate is
still first-class for routing: match and route or screen it at the content
layer. The full MAP/CAP argument surface lives in the published gsm_map /
gsm_cap codecs.
Termination shapes¶
When routing says local, the TCAP dialogue engine dispatches the operation to your handler. Three shapes cover MAP and CAMEL practice:
- single request/response: a
Begin(AARQ, Invoke)arrives, the handler replies, the reply is anEnd(AARE, ReturnResultLast)echoing the peer's transaction id. SRI-SM and initialDP are this shape. - held-open, multi-leg: the handler answers a
Beginwith aContinuethat keeps the dialogue open (an updateLocation answered with an insertSubscriberData invoke), and the peer's follow-up re-enters the handler to finish with anEnd. See Building an HLR.
Invoke and dialogue timers, and the dialogue ceiling, come from the
tcap: block and are enforced by the engine.
Metrics¶
The node maintains a Prometheus family set in Rust and renders it with
ss7.metrics() (or node.metrics()); it is never a per-message Python
call. The families:
| Family | Kind | Labels |
|---|---|---|
sigtran_association_state |
gauge | assoc, adaptation |
sigtran_association_refused_total |
counter | reason |
sigtran_association_dial_attempts_total |
counter | assoc |
sigtran_association_dial_failures_total |
counter | assoc, reason |
sigtran_association_dial_backoff_seconds |
gauge | assoc |
sigtran_asp_state |
gauge | asp, as |
sigtran_linkset_available |
gauge | linkset |
sigtran_linkset_active_links |
gauge | linkset |
sigtran_m2pa_link_state |
gauge | link |
sigtran_route_available |
gauge | dpc, linkset |
sigtran_mtp3mg_events_total |
counter | dpc, type |
sigtran_msu_total |
counter | linkset, dir, si |
sigtran_gtt_translations_total |
counter | selector, result |
sigtran_gtt_errors_total |
counter | reason |
sigtran_content_rule_hits_total |
counter | rule, action |
sigtran_content_unscreened_total |
counter | reason, action |
sigtran_active_dialogues |
gauge | |
sigtran_dialogue_timeouts_total |
counter | |
sigtran_invoke_timeouts_total |
counter | operation |
sigtran_abort_total |
counter | source |
sigtran_tcap_abort_total |
counter | dir, kind, cause |
sigtran_tcap_reject_total |
counter | dir, problem |
sigtran_tcap_discarded_total |
counter | reason |
sigtran_tcap_send_errors_total |
counter | |
sigtran_rx_discarded_total |
counter | reason |
sigtran_m2pa_discarded_total |
counter | reason |
sigtran_m2pa_link_failures_total |
counter | reason |
sigtran_link_tests_total |
counter | result |
sigtran_mtp3_link_traffic |
gauge | link |
sigtran_mtp3_restarting |
gauge | |
sigtran_mtp3_management_total |
counter | dir, type |
sigtran_mtp3_procedures_total |
counter | procedure |
sigtran_mtp3_retrieved_total |
counter | fate |
sigtran_mtp3_discarded_total |
counter | reason |
sigtran_as_state |
gauge | as |
sigtran_m3ua_management_total |
counter | dir, type |
sigtran_m3ua_error_total |
counter | dir, code |
sigtran_m3ua_notify_total |
counter | dir, status |
sigtran_m3ua_events_total |
counter | event |
sigtran_m3ua_discarded_total |
counter | reason |
sigtran_loops_detected_total |
counter | kind |
sigtran_isup_screened_total |
counter | reason |
sigtran_admission_refused_total |
counter | reason, action |
sigtran_rx_discarded_total counts what an association received and dropped
before routing: ppid (not the adaptation layer's payload protocol identifier),
undecodable, out-of-state (traffic from an ASP that is not active),
no-payload (a data message without its user part), not-unitdata (addressed
to a local subsystem and not a message the dialogue engine reads: a segment of
a segmented XUDT, an LUDT, a service message), link-test (a signalling link
test message for another point code, for a link code the linkset does not have,
or an acknowledgement of a test that is not running), and for M3UA
routing-context (a routing context the association is not configured in),
unexpected (a message the peer's role or the ASP's state does not allow),
parameter (a parameter missing, of a wrong length or with an invalid value)
and management (signalling network management in a DATA message). Each is
also a log line at warn level naming the association. Alert on any of them moving.
The dialogue engine reads a UDT and an XUDT that holds a whole message (no
segmentation parameter, or one that says first segment, none remaining), and
answers in the form the dialogue was opened in: XUDT for XUDT, with a hop
counter of 15. It does not reassemble: the segments of a segmented XUDT are
counted under not-unitdata and the message is not answered.
sigtran_association_refused_total counts the SCTP associations a listening
endpoint aborted (who a listening association accepts):
unknown-peer (no association on the endpoint lists any address of the peer),
unlisted-address (the peer presented a listed address and one that is not),
peer-port (the addresses of an association with another peer_port),
duplicate (the association is established already and stayed so) and
closed (the peer was gone before its addresses were read). Each is a log line at
warn level with the addresses, at most one a second per endpoint. unknown-peer
moving is someone who is not configured knocking; duplicate moving is a peer
that dials while its association is up, or one that came back from another
port before the old association had failed.
The dial families are about the client associations
(the dial block).
sigtran_association_dial_attempts_total{assoc} counts every attempt to
establish one, sigtran_association_dial_failures_total{assoc,reason} those
that failed: refused (the far end answered the INIT with an ABORT: nothing
listens there), timeout (no answer within connect_timeout_ms),
unreachable (no route, or the network or host is down), local (the local
addresses or port could not be bound) and other. The difference of the two
is how often the association was established, so a peer that accepts and
drops shows as attempts without failures.
sigtran_association_dial_backoff_seconds{assoc} is the wait that is running
before the next attempt, and 0 while the association is being dialled or is
established: together with sigtran_association_state at 0 it says "down,
next attempt in so long". Alert on the state, not on a single failure.
The M3UA families (M3UA management):
sigtran_as_state{as}is0down,1inactive,2active,3pending.sigtran_asp_state{asp,as}is1while the ASP is active in that server.sigtran_m3ua_management_total{dir,type}counts every M3UA message that is not DATA by its abbreviation (ASPUP,ASPAC_ACK,NTFY,ERR,DUNA,DAVA,DAUD,SCON,DUPU,DRST,BEAT, ...).sigtran_m3ua_error_total{dir,code}counts Error messages by the name of the code in RFC 4666 3.8.1 (invalid-routing-context,unsupported-traffic-mode-type,unexpected-message, ...).dir="rx"is a peer refusing what this node sent: alert on it.sigtran_m3ua_notify_total{dir,status}:as-inactive,as-active,as-pending,insufficient-asp-resources,alternate-asp-active,asp-failure.sigtran_m3ua_events_total{event}:recovery-expired(T(r) ran out with no ASP active),takeover(an override server changed hands),ack-timeout(an ASP Up or ASP Active of this node was sent again),beat-timeout(a silent peer was dropped),beat-mismatch(a Heartbeat Ack that does not return what was sent),changes-lagged(destination changes came faster than they were read; the ASPs were told the state as it is) anddata-on-stream-0.sigtran_m3ua_discarded_total{reason}counts what was not sent to a server:pending-fullandpending-expired(the queue of a pending server),management(a management message for the point code of a server that has no M3UA message) andnot-sent(a management message the association did not take within 250 ms).
An M2PA link has three families of its own. sigtran_m2pa_link_state is 0
while the link is out of service or aligning, 1 once level 2 is in service
(aligned, the link test still to pass, or the peer in processor outage) and 2
when the link is available for traffic. sigtran_m2pa_link_failures_total
counts every time a link leaves service or fails to align, by reason: t1,
t2, t3 (alignment timers), t6 (the peer stayed busy), t7 (a message was
not acknowledged), peer-status (the peer sent Alignment, Proving or Out of
Service), abnormal-bsn, link-test, association (SCTP lost) and stopped.
sigtran_m2pa_discarded_total counts messages a link did not carry:
rx-out-of-sequence, rx-not-in-service, rx-abnormal-bsn,
tx-not-in-service, tx-congestion (the transmit queue was full),
tx-lost-on-failure (sent or queued, not acknowledged, and the link failed),
tx-flushed-by-peer (discarded by a peer recovering from processor outage) and
encode. sigtran_link_tests_total counts signalling link test messages
sent, passed and failed.
The level-3 families (Signalling network management):
sigtran_mtp3_link_traffic{link}is1while level 3 routes traffic to the link,0while the link is down, changed over, inhibited, or waiting for a restart to end.sigtran_m2pa_link_statesays what level 2 thinks of the same link.sigtran_mtp3_restartingis1while this node's restart runs.sigtran_mtp3_management_total{dir,type}counts management messages by their abbreviation in Table 1/Q.704 (TFP,TFA,XCO,TRA, ...).sigtran_mtp3mg_events_total{dpc,type}counts per destination:tfpandtfafor a route marked by an adjacent point,pauseandresumefor the destination becoming inaccessible and accessible,congestionfor a transfer-controlled.sigtran_mtp3_procedures_total{procedure}counts what ran and how it ended:changeover-acknowledged,changeover-emergency(acknowledged without a usable sequence number),changeover-unacknowledged(T2),changeover-time-controlled(T1),changeback-acknowledged,changeback-unacknowledged(T5),changeback-time-controlled(T3),forced-rerouting,controlled-rerouting,restart-begun,restart-ended,adjacent-restart-begun,adjacent-restart-ended,adjacent-restart-expired(T21, the peer never said traffic restart allowed),link-inhibited,link-inhibit-denied,link-uninhibitedanduninhibit-abandoned.sigtran_mtp3_retrieved_total{fate}counts what a failed link still held at changeover:resenton an alternative link,confirmed(the far end had it),link-test(a link test message, which goes nowhere else). What was sent, never acknowledged and not sent again is insigtran_m2pa_discarded_total{reason="tx-lost-on-failure"}: that counter moving is traffic that may be lost, andchangeover-unacknowledgedorchangeover-time-controlledmoving with it says why.sigtran_mtp3_discarded_total{reason}counts what level 3 did not carry:no-route(a message for a point code nothing leads to, answered with transfer-prohibited),no-egress(a route was chosen and nothing on it carries),restarting(received from a linkset during this node's restart),no-alternative(held at changeover, and nothing else leads to its destination),buffer-full(a changeover, changeback or rerouting buffer overflowed),management-not-acted-onandmanagement-no-path(a management message of this node had no link or route to leave on).
Good first panels: linkset / ASP state and route availability (is the node
wired up?), MSU rate by SI, GTT translations vs errors, active dialogues, and
sigtran_loops_detected_total (a route-table mistake shows up here instead of
as an outage).