From eb66e6dbc2600e3b23090f55c7351a7f06f2180e Mon Sep 17 00:00:00 2001
From: Moses Narrow <36607567+0pcom@users.noreply.github.com>
Date: Tue, 25 Aug 2026 11:48:48 -0500
Subject: [PATCH] Revert "feat(cxo): consolidate tp-list onto the telemetry
 feed (#4171)"
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Reverts #4171. On a busy hub (my dev visor, ~750-1271 live transports)
TPD stopped agreeing after the visor moved to the combined feed: its
transport count churned at ~15-30% of live and TPD showed no aggregator
activity for that visor's PK — the combined-feed subscribe/fetch did not
land for the large Root, while the dedicated tiny tp-list Root filled
completely (1271=1271) before. The consolidation works for normal visors
(the port-50 targeted fetch landed for ~143) but regresses large hubs,
and TPD's global total dropped as the fleet adopted it, so restore the
proven two-feed model until the busy-hub combined-feed path is fixed and
verified. The connection-halving win will return once that's understood.
---
 pkg/visor/init_stats.go     | 31 ++++++++++++++-----------------
 pkg/visor/visorconfig/v1.go | 16 ----------------
 2 files changed, 14 insertions(+), 33 deletions(-)

diff --git a/pkg/visor/init_stats.go b/pkg/visor/init_stats.go
index 60aa5deeba..c5e2366a77 100644
--- a/pkg/visor/init_stats.go
+++ b/pkg/visor/init_stats.go
@@ -102,23 +102,20 @@ func initStats(_ context.Context, v *Visor, log *logging.Logger) error {
 		}
 		tracker.SetSink(sink)
 
-		// Transport-list discovery feed. CONSOLIDATED default: the tp-list
-		// snapshot leaf rides THIS telemetry feed (`pub`) as a top-level
-		// leaf, so a visor holds ONE CXO node/conn to TPD, not two. TPD's
-		// aggregator extracts the small tp-list leaf via its targeted
-		// discovery-leaf fetch — reliable even when the telemetry Root's
-		// deep fill breaks on a busy hub — so the combined feed keeps the
-		// full transport count without a dedicated second feed. This halves
-		// the visor↔TPD connection count (and the per-session dmsg
-		// handshakes) versus the old two-feed model.
-		//
-		// Opt back into the dedicated port-69 feed with
-		// Stats.DedicatedTPListFeed=true (a SECOND CXO node under the same
-		// visor PK on DmsgVisorTPListCXOPort) — kept as an escape hatch in
-		// case a busy hub is ever seen under-filling on the combined feed.
-		if conf != nil && conf.DedicatedTPListFeed {
-			tplistPub = buildTPListPublisher(v, log)
-		}
+		// Dedicated tp-list discovery feed: a SECOND CXO node under the
+		// SAME visor identity PK, on DmsgVisorTPListCXOPort, carrying ONLY
+		// the compact transport-list snapshot leaf (publishTPDList). Its
+		// Root is a handful of objects, so TPD's second aggregator fills it
+		// COMPLETELY in ~1 round-trip — the durable cure for the ~10%
+		// transport under-report on busy hubs, whose combined telemetry
+		// Root on `pub` (DmsgCXOPort) can't finish its whole-Root fill in
+		// the announce conn's window. Same PK keeps TPD's reporter=feed-PK
+		// edge-auth intact. The telemetry leaves (transports/<id>/current)
+		// stay on `pub`. If the dedicated publisher can't start, fall back
+		// to publishing the tp-list on the telemetry feed (legacy combined
+		// behavior) so discovery still works — just without the
+		// fill-completeness win.
+		tplistPub = buildTPListPublisher(v, log)
 		leafPub := pub
 		if tplistPub != nil {
 			leafPub = tplistPub
diff --git a/pkg/visor/visorconfig/v1.go b/pkg/visor/visorconfig/v1.go
index f4166ba527..ae9111c6e7 100644
--- a/pkg/visor/visorconfig/v1.go
+++ b/pkg/visor/visorconfig/v1.go
@@ -255,22 +255,6 @@ type Stats struct {
 	// Disabled, when true, skips the entire telemetry store and
 	// associated /stats/* endpoints + CXO publisher.
 	Disabled bool `json:"disabled,omitempty"`
-	// DedicatedTPListFeed, when true, publishes the transport-list
-	// discovery snapshot on its OWN CXO node/port
-	// (skyenv.DmsgVisorTPListCXOPort) instead of on the telemetry feed.
-	//
-	// Default (false) is the CONSOLIDATED single-feed model: the tp-list
-	// leaf rides the telemetry Root as a top-level leaf, so TPD holds ONE
-	// CXO connection per visor instead of two. TPD's aggregator extracts
-	// the tp-list via its targeted discovery-leaf fetch, which lands the
-	// small leaf reliably even when the telemetry Root's deep fill breaks
-	// on a busy hub. Halving the visor↔TPD connections cuts TPD's
-	// per-connection goroutines and the per-session dmsg handshake load on
-	// the dmsg servers.
-	//
-	// Set true only to restore the separate dedicated feed — e.g. if a
-	// busy hub's tp-list is observed to under-fill on the combined feed.
-	DedicatedTPListFeed bool `json:"dedicated_tplist_feed,omitempty"`
 }
 
 // LogServer configures the dmsghttp log server's optional localhost endpoint.