Development Update — April 25

Skywire: Production Performance Push

A run of profile-driven optimizations targeting the deployment services. Every change came from a specific pprof or production-redis observation.

5e9044141 Setup-node porter watchdog — standalone Node.Serve was missing the SweepStalePorterEntries loop the embedded RSN already runs. Production saw ~13 ports/min leaking, full ephemeral exhaustion (16,387 ports reserved) after ~20h of uptime, and ~1 GB of stranded yamux.Stream objects parented by the porter map. Adds the same 60s sweep that drops entries older than 5 minutes.

e3074d7f3 SD: replace SCAN-for-existence with index lookup — isServiceOnline was using SCAN cursor=0 pattern="service:*:<pk>" COUNT=10 to test whether any service entry existed for a visor. That’s O(N) over the whole 150K-key keyspace when the entry is absent (offline visor). refreshUptimesCache runs every 5 minutes and called this once per visor seen today, twice (v1+v2 cache) — production redis spent ~33% of CPU on SCAN, ~3,316 calls/sec sustained. Switches to the per-visor service-type index set already maintained by UpdateService / DeleteService: one SMEMBERS plus a multi-key EXISTS. Constant Redis round-trips regardless of keyspace size.

2baa4da8a TPD: per-edge entry cache — mirrorEdges fires GetTransportsByEdge once per touched edge after every register/deregister. Hub edges (those that participate in many transports) were re-fetched on every mutation even when the list hadn’t changed. pprof showed mirrorEdges → GetTransportsByEdge dominating TPD CPU at ~35% cumulative even after the recent DHT mirror PRs. Adds a 4096-slot, 5-second-TTL cache keyed by edge pubkey, with explicit invalidation on register/deregister/batch.

3a433560d TPD: replace tp:* SCAN with SMEMBERS — getAllTransports used SCAN MATCH tp:* to list registered transports. Now maintains a tp:ids set updated transactionally on every RegisterTransport / DeregisterTransport, and uses SMEMBERS tp:ids + pipelined MGET. Constant-time membership operations replace cursor iteration over a growing keyspace.

99f8c377d AR: replace GetAll SCAN with SMEMBERS on per-netType index — same shape, applied to address-resolver. Bind/del-bind paths maintain addr-ids:stcpr / addr-ids:sudph index sets; GetAll reads them directly.

6b8a3cb8a / de8116036 / f29473ca4 Short-TTL caches for hot list endpoints — GetAllTransports (TPD), GetAllTransportsWithQoS (TPD), and GetAll (AR) each gain a small in-process TTL cache keyed by query shape. Soaks up the synchronized bursts that visor sync=true re-registration cycles produce on the /transports endpoints.

0a2a592f5 Cipher: disable DebugLevel1 — secp256k1.SignDeterministic runs an internal verify-after-sign pass when DebugLevel1 is set. The flag was on by default, doubling the CPU cost of every signature on TPD/SD/AR, the visor’s DHT publisher, and the dmsg client’s entry update path. Off in release builds; tests still set it explicitly.

49b211304 dmsghttp: per-destination stream pool — every HTTP request through dmsg opened a brand-new stream and paid a full noise handshake. With ~50 visor heartbeats/sec network-wide this drove dmsg-server CPU into the ground (production pprof: handshakeResponder + secp256k1 sign/verify at ~50% cumulative wall). HTTPTransport now keeps a small per-destination pool of dmsg.Stream objects keyed by dmsg.Addr. RoundTrip tries one pooled stream, falls through to a fresh dial on I/O error, and returns the stream to the pool after body close (or discards on error). One retry only — pooled streams might already be torn down by the remote.

Skywire: DHT Value Size Bug

79c398daa Raise MaxValueSize 16K → 64K, log size-induced publish drops — production observation 2026-04-25: hub edges with hundreds of transports silently failed to publish to the DHT because the marshalled list of compact entries (~80 bytes each) crossed the 16 KB ceiling at ~200 entries. The DHT state for those subjects froze at whatever fit on the last successful publish. Concrete examples (cli dht get vs TPD GetTransportsByEdge): edge 027087fe40… had 883 transports in TPD vs 107 in DHT (12%); 03d1d78e7323… 916 vs 149 (16%); 02859d5aacdf… 149 vs 16 (11%). Two changes: raise the per-MutableItem ceiling to 64K, and surface size-induced drops as a WARN instead of swallowing them.

Skywire: Visor Survey IP Requirement

81c8d6a14 Require IP in survey, retry dmsg LookupIP indefinitely — survey was written even when dmsgC.LookupIP() failed all 8 retries (no delay between attempts), producing surveys with an empty ip_address field. The reward calc drops them but reports the empty IP as the reason, so operators saw blank rejection reasons in production logs. Now retries indefinitely with a 30s per-attempt timeout and 15s spacing between attempts; the survey is never generated without a public IP.

Skywire: Hypervisor TUI + Hypervisor-of-Hypervisors (PR #2337)

The big rollup PR that lands in v1.3.47:

• Hypervisor terminal UI (cli visor hv tui) — interactive tview-based table listing all visors connected to this hypervisor, with detail view per visor. Pulls data via the new HVListVisors / HVVisorSummary RPCs that proxy through the hypervisor’s existing DMSG connections.
• Transitive HV management — cli visor hv ls lists remote visors connected to this hypervisor; --detail queries each via transport RPC for version, uptime, and transports.
• tp-rpc accepts arguments — third positional arg is JSON args, enabling calling any RPC method with parameters.
• Top-level CLI shortcuts in feat/next — preview of the verb shortcuts that land on April 28 (cli pk, cli status, etc).
• Route-finder LabelSetup filter — setup-node transports (LabelSetup) excluded from regular route discovery so user routes don’t accidentally traverse them.
• DMSG-only-as-last-hop — DMSG hops are now permitted only as the terminal hop in a route, not as intermediate hops (DMSG over DMSG was a routing dead end).
• No-mux on DMSG routes — mux is a transport-level multiplexing for direct legs; DMSG sessions are already multiplexed at the protocol layer.
• Configurable transport-type preference — visors can declare ordering for STCPR/SUDPH/DMSG when calculating routes.
• RouteGroupInfo hop list — reports the ordered hop sequence so route groups output shows the actual path each group took, not just endpoints.

Skywire: v1.3.47 Released

1f927d5db release(v1.3.47) — bundles PRs #2322–#2349 (the DHT consistency work, perf optimizations, transport-RPC, hypervisor TUI, and the feat/next rollup). Mainnet cutoff bumped from v1.3.43 (was 2026-04-23) to v1.3.47 effective 2026-05-09 — the standard 14-day announcement window.