Files
meshcore-analyzer/cmd/server/load_graph_invariant_test.go
efiten 938153dd92 fix(nodes): rebuild relay-hop history on startup from path_json (#1643)
## Problem

A relay node's **activity timeline** — and its per-node `packetsToday` /
observer counts — collapses to *"only the hour the server restarted"*
after every restart. Before the restart the timeline shows only the
node's own adverts (~1–2/hr); all of its relay activity piles into the
single post-restart hour.

## Root cause

All DB cold-load paths (`Load`, `loadChunk`, `scanAndMergeChunk`) index
relay-hop attribution into `byNode` **only** from
`observations.resolved_path`. But since #1287 the ingestor persists
relay data as aggregate `neighbor_edges` and **never writes
`resolved_path`** — it is `NULL` on every deployment (verified on a live
DB: 0 of ~440k rows populated). So relay attribution is never
reconstructed on startup; it only re-accumulates from live traffic
(`IngestNew*`, which re-resolves from `path_json` + the neighbor graph),
piling a relay node's whole history into the post-restart window.

## Fix

Server read-side only — **no schema / ingestor / migration change**.
When `resolved_path` is empty, re-resolve relay hops from the
already-persisted `path_json` using the in-memory prefix map + neighbor
graph (the same `resolvePathForObs` compute the live ingest path already
runs). `main.go` now loads the persisted neighbor graph *before* the
packet load so resolution has the graph available.

Two correctness details worth a close look:

1. **Fetch the prefix-map/graph snapshot BEFORE opening each load
cursor.** `getCachedNodesAndPM` issues its own DB query; doing so while
a load cursor is open deadlocks on a single-connection SQLite pool (the
test harness uses one).
2. **Index into `byNode` ONLY** — not the `resolved_path` / path-hop
indexes. Those are cross-checked by `handleNodePaths` against the
persisted `resolved_path` column (NULL here); populating them from an
in-memory re-resolution would make that SQL confirmation fail and
wrongly drop the tx from paths-through (#1352).

## Tests

New coverage asserts a relay pubkey reachable *only* via `path_json`
lands in `byNode` after a restart-style load, for both the hot-window
(`LoadChunked`) and background-window (`loadChunk`) paths. Existing
#1558 (`resolved_path`) and #1352 (paths-through) tests still pass. Full
`cd cmd/server && go test ./...` is green under `-race`.

## Perf

The fallback runs `resolvePathForObs` per observation with a non-empty
`path_json` during cold load — the same per-packet compute the live
ingest path already performs, so no new asymptotic cost. The prefix map
+ graph are snapshotted **once per load** (not per row);
`getCachedNodesAndPM` is 30s-cached. In `loadChunk` the resolution runs
in the existing lock-free scan and is accumulated locally, matching that
function's "build local, merge under lock" design.

## Note on a pre-existing flaky test

`TestDistanceConcurrentRequestsDuringBuildReturn202` is timing-fragile
(fails ~1/15 on `master` without this change). It relies on the lazy
distance build being slow because it's the first caller of
`getCachedNodesAndPM` (cold cache). This PR pre-warms that cache during
`Load`, narrowing the build window, so the test fails more often in
**non-race** local runs. It passes reliably under `-race` (CI mode),
where the build stays slow. Flagging in case you want to harden the test
separately.

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: openclaw-bot <openclaw-bot@users.noreply.github.com>
Co-authored-by: openclaw-bot <bot@openclaw>
2026-06-11 11:36:49 -07:00

91 lines
3.1 KiB
Go

package main
import (
"database/sql"
"path/filepath"
"testing"
"time"
_ "modernc.org/sqlite"
)
// TestLoad_PanicsWhenGraphNotLoadedAndEdgesExist pins the startup-ordering
// invariant (munger R1 #2). Graph-load-before-packet-load is the entire
// premise of PR #1643's fix: without an in-memory neighbor graph, the
// path_json relay-hop fallback cannot resolve hops, so relay-node analytics
// history collapses. main.go currently does the right thing — but nothing
// asserts the ordering, so a future refactor could silently regress.
//
// Load() must panic when neighbor_edges has rows but s.graph.Load() returns
// nil. Fast-fail at startup beats silently-wrong attribution.
func TestLoad_PanicsWhenGraphNotLoadedAndEdgesExist(t *testing.T) {
dir := t.TempDir()
dbPath := filepath.Join(dir, "test.db")
rw, err := sql.Open("sqlite", "file:"+dbPath+"?_journal_mode=WAL")
if err != nil {
t.Fatal(err)
}
defer rw.Close()
exec := func(s string, args ...interface{}) {
if _, err := rw.Exec(s, args...); err != nil {
t.Fatalf("setup exec failed: %v\nSQL: %s", err, s)
}
}
// Minimal CoreScope schema. PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE transmissions (
id INTEGER PRIMARY KEY,
raw_hex TEXT, hash TEXT, first_seen TEXT,
route_type INTEGER, payload_type INTEGER, payload_version INTEGER,
decoded_json TEXT
)`)
// PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE observations (
id INTEGER PRIMARY KEY, transmission_id INTEGER,
observer_id TEXT, observer_name TEXT,
direction TEXT, snr REAL, rssi REAL, score INTEGER,
path_json TEXT, timestamp TEXT, raw_hex TEXT, resolved_path TEXT
)`)
// PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE observers (rowid INTEGER PRIMARY KEY, id TEXT, name TEXT, iata TEXT)`)
// PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE nodes (
public_key TEXT PRIMARY KEY, name TEXT, role TEXT, lat REAL, lon REAL,
last_seen TEXT, first_seen TEXT, advert_count INTEGER DEFAULT 0
)`)
// PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE schema_version (version INTEGER)`)
exec(`INSERT INTO schema_version (version) VALUES (1)`)
// PREFLIGHT: async=true reason="test fixture, in-memory tmpdir DB"
exec(`CREATE TABLE neighbor_edges (
node_a TEXT NOT NULL,
node_b TEXT NOT NULL,
count INTEGER DEFAULT 1,
last_seen TEXT,
PRIMARY KEY (node_a, node_b)
)`)
now := time.Now().UTC().Format(time.RFC3339)
exec(`INSERT INTO neighbor_edges (node_a, node_b, count, last_seen) VALUES (?, ?, ?, ?)`,
"aaa", "bbb", 5, now)
d, err := OpenDB(dbPath)
if err != nil {
t.Fatalf("OpenDB: %v", err)
}
defer d.conn.Close()
// Deliberately DO NOT call store.graph.Store(...). s.graph.Load() returns
// nil → the bug condition the invariant guard must catch.
store := NewPacketStore(d, &PacketStoreConfig{RetentionHours: 72})
defer func() {
r := recover()
if r == nil {
t.Fatalf("Load() must panic when neighbor_edges has rows but graph is nil; got no panic")
}
}()
_ = store.Load()
}