-- 节点健康 — per-minute rollup of satellite host vitals. -- -- Granularity is the whole design. Satellites heartbeat every 10s; writing a row -- per heartbeat would be ~60k rows/day for 7 nodes and grow forever. Rolling up -- to one row per node per minute makes it ~10k/day, and the retention sweep -- (node_health.ts::pruneNodeMetrics, 30 days) holds the table at a CONSTANT -- ~300k rows. A table that doesn't grow is one nobody has to deal with later — -- that is worth more here than the resolution we give up. -- -- avg AND max because they answer different questions: avg is what a trend line -- plots, max is what a threshold was crossed by. A disk filling up over ten -- seconds is a spike a one-minute average swallows whole. -- -- The CURRENT value of each metric is deliberately NOT here. It lives in control -- plane memory (same call as nodes.pendingDeploy): the node owns that truth and -- re-asserts it every ~10s, so a restart re-learns it within one heartbeat, and -- a last-known CPU reading is worthless after a restart anyway. CREATE TABLE IF NOT EXISTS "node_metrics" ( "node_id" text NOT NULL, -- Start of the minute bucket, epoch ms. "at" bigint NOT NULL, "cpu_avg" double precision, "cpu_max" double precision, "mem_avg" double precision, "mem_max" double precision, "disk_avg" double precision, "disk_max" double precision, "load_max" double precision, "swap_max_bytes" bigint, -- Samples folded into this bucket; a value below the expected count is how a -- reader spots a gap rather than misreading it as a flat line. "samples" bigint DEFAULT 0 NOT NULL, CONSTRAINT "node_metrics_node_id_at_pk" PRIMARY KEY("node_id","at") ); -- Every metric column is nullable on purpose: a probe that failed must record -- "unknown", never 0. A zeroed disk reading would mean "plenty of room" on -- precisely the node we can no longer measure, and the alarm would stay silent. -- The primary key already indexes (node_id, at), which is the panel's range -- scan. This second index serves the retention sweep, which deletes by age -- across all nodes and would otherwise seq-scan the whole table every hour. CREATE INDEX IF NOT EXISTS "node_metrics_at_idx" ON "node_metrics" ("at");