History: server load, memory and disks are also kept as hourly (four months) and daily (over two years) averages with lowest and highest; endpoint answers as hourly (four months) beside the daily figures. Charts ask for 24 hours to a year and are drawn from whichever is fine enough. A disk that two weeks call flat is looked at again over ninety days for the forecast. Built once from what is kept. One chart: TimeChart replaces the sparkline and the multi-line chart: hover and arrow-key readout, a lowest/highest band, gaps shown as gaps, the same ranges everywhere and the choice remembered. Used for server trends, comparing servers, a check's availability and answer times (Endpoints > History) and a server's availability. Read API: /api/v1 (servers, metrics, alerts, incidents, inventory, network, endpoints and their history, maintenance) with personal read-only tokens made under the user menu: shown once, hashed, 30/90/365 days or none, ten at most, removed with their owner. GET only. Documented in API.md. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
365 lines
18 KiB
TypeScript
365 lines
18 KiB
TypeScript
/**
|
|
* A history of what each agent-reported server's load, memory and disks looked like, so a page can show a trend
|
|
* (and say how soon a filling disk will run out) instead of only the latest figure.
|
|
*
|
|
* One sample is kept per server every ten minutes at most, however often its agent reports, for 35 days. Everything here is a
|
|
* percentage: CPU load (as the agent reports it), memory in use, and each disk's used share.
|
|
*
|
|
* Beyond those 35 days the same figures are kept coarser: each sample also goes into its hour (kept four months) and its day (kept over two years), as a
|
|
* sum, a count, a lowest and a highest. A chart asks for a number of days and is answered from whichever is fine enough and still there: the samples up to
|
|
* 35 days, the hours up to 120, the days beyond. A year of a disk filling is then one point a day.
|
|
*/
|
|
import { and, desc, eq, gte, lt, sql } from "drizzle-orm";
|
|
import { db } from "../db/client.js";
|
|
import { serverMetrics, serverMetricsRollup } from "../db/schema.js";
|
|
import { getInternalFlag, setInternalFlag } from "./settingsStore.js";
|
|
import type { IncomingSystemInfo } from "./taskSync.js";
|
|
|
|
/** Reports closer together than this add no sample. */
|
|
export const MIN_GAP_MS = 10 * 60_000;
|
|
export const RETENTION_DAYS = 35;
|
|
/** How long the hourly and daily figures are kept. */
|
|
export const HOURLY_KEEP_DAYS = 120;
|
|
export const DAILY_KEEP_DAYS = 800;
|
|
/** The longest stretch a chart may ask for. */
|
|
export const MAX_RANGE_DAYS = 365;
|
|
const DAY_MS = 24 * 3600_000;
|
|
const HOUR_MS = 3600_000;
|
|
|
|
export interface Sample {
|
|
metric: string;
|
|
value: number;
|
|
}
|
|
|
|
const round1 = (n: number) => Math.round(n * 10) / 10;
|
|
const clampPercent = (n: number) => Math.min(100, Math.max(0, n));
|
|
|
|
/** The percentages in one agent report. Anything missing or nonsensical (no size, not a number) is left out. */
|
|
export function samplesFromSystem(system: IncomingSystemInfo | null | undefined): Sample[] {
|
|
if (!system) return [];
|
|
const out: Sample[] = [];
|
|
const load = system.cpu?.load_percent;
|
|
if (typeof load === "number" && Number.isFinite(load)) out.push({ metric: "cpu", value: round1(clampPercent(load)) });
|
|
const total = system.memory?.total_bytes;
|
|
const used = system.memory?.used_bytes;
|
|
if (typeof total === "number" && typeof used === "number" && total > 0) out.push({ metric: "mem", value: round1(clampPercent((used / total) * 100)) });
|
|
for (const d of system.disks ?? []) {
|
|
if (typeof d.mount !== "string" || !d.mount || !(d.size_bytes > 0) || typeof d.used_bytes !== "number") continue;
|
|
out.push({ metric: `disk:${d.mount}`, value: round1(clampPercent((d.used_bytes / d.size_bytes) * 100)) });
|
|
}
|
|
return out;
|
|
}
|
|
|
|
/** The start of the hour or day a moment falls in, as ISO. */
|
|
export const hourStart = (t: number | string) => new Date(Math.floor(new Date(t).getTime() / HOUR_MS) * HOUR_MS).toISOString();
|
|
export const dayStart = (t: number | string) => new Date(Math.floor(new Date(t).getTime() / DAY_MS) * DAY_MS).toISOString();
|
|
|
|
/** Adds samples to their hour and day. Never allowed to break recording the sample itself. */
|
|
async function addToRollups(serverId: number, samples: Sample[], at: Date): Promise<void> {
|
|
try {
|
|
for (const grain of ["h", "d"] as const) {
|
|
const bucket = grain === "h" ? hourStart(at.getTime()) : dayStart(at.getTime());
|
|
for (const s of samples) {
|
|
await db
|
|
.insert(serverMetricsRollup)
|
|
.values({ serverId, metric: s.metric, grain, at: bucket, sum: s.value, n: 1, min: s.value, max: s.value })
|
|
.onConflictDoUpdate({
|
|
target: [serverMetricsRollup.serverId, serverMetricsRollup.metric, serverMetricsRollup.grain, serverMetricsRollup.at],
|
|
set: {
|
|
sum: sql`${serverMetricsRollup.sum} + ${s.value}`,
|
|
n: sql`${serverMetricsRollup.n} + 1`,
|
|
min: sql`min(${serverMetricsRollup.min}, ${s.value})`,
|
|
max: sql`max(${serverMetricsRollup.max}, ${s.value})`,
|
|
},
|
|
});
|
|
}
|
|
}
|
|
} catch (err) {
|
|
console.error("[metrics] couldn't add to the hourly and daily figures:", err);
|
|
}
|
|
}
|
|
|
|
/** Records a report's figures, unless this server was sampled a moment ago. Returns whether anything was stored. */
|
|
export async function recordServerMetrics(serverId: number, system: IncomingSystemInfo | null | undefined, now: Date = new Date()): Promise<boolean> {
|
|
const samples = samplesFromSystem(system);
|
|
if (samples.length === 0) return false;
|
|
|
|
const [latest] = await db
|
|
.select({ at: serverMetrics.recordedAt })
|
|
.from(serverMetrics)
|
|
.where(eq(serverMetrics.serverId, serverId))
|
|
.orderBy(desc(serverMetrics.recordedAt))
|
|
.limit(1);
|
|
if (latest && now.getTime() - Date.parse(latest.at) < MIN_GAP_MS) return false;
|
|
|
|
const at = now.toISOString();
|
|
await db.insert(serverMetrics).values(samples.map((s) => ({ serverId, metric: s.metric, value: s.value, recordedAt: at })));
|
|
await addToRollups(serverId, samples, now);
|
|
await db.delete(serverMetrics).where(and(eq(serverMetrics.serverId, serverId), lt(serverMetrics.recordedAt, new Date(now.getTime() - RETENTION_DAYS * DAY_MS).toISOString())));
|
|
await db.delete(serverMetricsRollup).where(and(eq(serverMetricsRollup.serverId, serverId), eq(serverMetricsRollup.grain, "h"), lt(serverMetricsRollup.at, new Date(now.getTime() - HOURLY_KEEP_DAYS * DAY_MS).toISOString())));
|
|
await db.delete(serverMetricsRollup).where(and(eq(serverMetricsRollup.serverId, serverId), eq(serverMetricsRollup.grain, "d"), lt(serverMetricsRollup.at, new Date(now.getTime() - DAILY_KEEP_DAYS * DAY_MS).toISOString())));
|
|
return true;
|
|
}
|
|
|
|
const BACKFILL_FLAG = "metricsRollupBackfilled";
|
|
|
|
/**
|
|
* The first time after the hourly and daily figures were added, builds them from the samples already kept (up to 35 days), so the longer ranges are not empty
|
|
* for weeks. A bucket is written from what the samples say, replacing what is there, so it is right however far it got before. Done once.
|
|
*/
|
|
export async function backfillMetricRollups(): Promise<boolean> {
|
|
if (await getInternalFlag(BACKFILL_FLAG)) return false;
|
|
for (const [grain, len] of [["h", 13], ["d", 10]] as const) {
|
|
const suffix = grain === "h" ? ":00:00.000Z" : "T00:00:00.000Z";
|
|
await db.run(sql`
|
|
INSERT INTO server_metrics_rollup (server_id, metric, grain, at, sum, n, min, max)
|
|
SELECT server_id, metric, ${grain}, substr(recorded_at, 1, ${len}) || ${suffix}, sum(value), count(*), min(value), max(value)
|
|
FROM server_metrics
|
|
GROUP BY server_id, metric, substr(recorded_at, 1, ${len})
|
|
ON CONFLICT (server_id, metric, grain, at) DO UPDATE SET sum = excluded.sum, n = excluded.n, min = excluded.min, max = excluded.max
|
|
`);
|
|
}
|
|
await setInternalFlag(BACKFILL_FLAG, new Date().toISOString());
|
|
return true;
|
|
}
|
|
|
|
export interface Point {
|
|
/** ISO time. */
|
|
t: string;
|
|
v: number;
|
|
/** The lowest and highest figure in what this point stands for, when it stands for several (hourly and daily figures). */
|
|
lo?: number;
|
|
hi?: number;
|
|
}
|
|
|
|
/** Averages points into at most `maxPoints` equal slices of [from, to], so a month of samples draws as a short line. */
|
|
export function bucketPoints(points: { t: number; v: number }[], from: number, to: number, maxPoints: number): Point[] {
|
|
if (points.length === 0 || to <= from) return [];
|
|
const size = Math.max((to - from) / maxPoints, 1);
|
|
const buckets = new Map<number, { sum: number; n: number; tSum: number }>();
|
|
for (const p of points) {
|
|
if (p.t < from || p.t > to) continue;
|
|
const key = Math.min(maxPoints - 1, Math.floor((p.t - from) / size));
|
|
const b = buckets.get(key) ?? { sum: 0, n: 0, tSum: 0 };
|
|
b.sum += p.v;
|
|
b.n += 1;
|
|
b.tSum += p.t;
|
|
buckets.set(key, b);
|
|
}
|
|
return [...buckets.entries()]
|
|
.sort((a, b) => a[0] - b[0])
|
|
.map(([, b]) => ({ t: new Date(b.tSum / b.n).toISOString(), v: round1(b.sum / b.n) }));
|
|
}
|
|
|
|
/** A row of the hourly or daily figures, with the moment it starts at as epoch ms. */
|
|
export interface RollupRow {
|
|
t: number;
|
|
sum: number;
|
|
n: number;
|
|
min: number;
|
|
max: number;
|
|
}
|
|
|
|
/** The average of what each row stands for. */
|
|
export function rollupAverages(rows: RollupRow[]): { t: number; v: number }[] {
|
|
return rows.filter((r) => r.n > 0).map((r) => ({ t: r.t, v: r.sum / r.n }));
|
|
}
|
|
|
|
/** Rows into at most `maxPoints` slices of [from, to]: the average of the samples in each (not of the rows), with the lowest and highest. */
|
|
export function bucketRollups(rows: RollupRow[], from: number, to: number, maxPoints: number): Point[] {
|
|
if (rows.length === 0 || to <= from) return [];
|
|
const size = Math.max((to - from) / maxPoints, 1);
|
|
const buckets = new Map<number, { sum: number; n: number; lo: number; hi: number; tSum: number; rows: number }>();
|
|
for (const r of rows) {
|
|
if (r.t < from || r.t > to || r.n <= 0) continue;
|
|
const key = Math.min(maxPoints - 1, Math.floor((r.t - from) / size));
|
|
const b = buckets.get(key) ?? { sum: 0, n: 0, lo: Infinity, hi: -Infinity, tSum: 0, rows: 0 };
|
|
b.sum += r.sum;
|
|
b.n += r.n;
|
|
b.lo = Math.min(b.lo, r.min);
|
|
b.hi = Math.max(b.hi, r.max);
|
|
b.tSum += r.t;
|
|
b.rows += 1;
|
|
buckets.set(key, b);
|
|
}
|
|
return [...buckets.entries()]
|
|
.sort((a, b) => a[0] - b[0])
|
|
.map(([, b]) => ({ t: new Date(b.tSum / b.rows).toISOString(), v: round1(b.sum / b.n), lo: round1(b.lo), hi: round1(b.hi) }));
|
|
}
|
|
|
|
export interface Trend {
|
|
/** Change in the percentage per day, from a straight line through the points. */
|
|
perDay: number;
|
|
/** How well a straight line describes the points, 0 to 1 (R²). A steady climb is near 1; a single jump or a wobble is low. */
|
|
fit: number;
|
|
/** How much time the points cover, in days. */
|
|
spanDays: number;
|
|
/** For something that fills up: days until it reaches 100% at that rate. Null when it isn't rising or the rate is negligible. */
|
|
daysToFull: number | null;
|
|
}
|
|
|
|
/**
|
|
* A straight-line fit over the points. Needs at least six points covering at least a day; with less there is no honest trend to
|
|
* give. A slope under 0.05 points a day counts as flat, so a disk that wobbles by a fraction isn't forecast to fill in a century.
|
|
*/
|
|
export function trendOf(points: { t: number; v: number }[]): Trend | null {
|
|
if (points.length < 6) return null;
|
|
const span = points[points.length - 1].t - points[0].t;
|
|
if (span < DAY_MS) return null;
|
|
const t0 = points[0].t;
|
|
const xs = points.map((p) => (p.t - t0) / DAY_MS);
|
|
const ys = points.map((p) => p.v);
|
|
const n = points.length;
|
|
const meanX = xs.reduce((a, b) => a + b, 0) / n;
|
|
const meanY = ys.reduce((a, b) => a + b, 0) / n;
|
|
let num = 0;
|
|
let den = 0;
|
|
let totalVariance = 0;
|
|
for (let i = 0; i < n; i++) {
|
|
num += (xs[i] - meanX) * (ys[i] - meanY);
|
|
den += (xs[i] - meanX) ** 2;
|
|
totalVariance += (ys[i] - meanY) ** 2;
|
|
}
|
|
if (den === 0) return null;
|
|
const slope = num / den;
|
|
// R²: the share of the movement the straight line explains. A flat series has nothing to explain, so it is no "fit" at all.
|
|
const fit = totalVariance === 0 ? 0 : Math.round(((num * num) / (den * totalVariance)) * 100) / 100;
|
|
const perDay = Math.round(slope * 100) / 100;
|
|
const last = ys[n - 1];
|
|
const daysToFull = perDay >= 0.05 && last < 100 ? Math.ceil((100 - last) / perDay) : null;
|
|
return { perDay, fit, spanDays: Math.round((span / DAY_MS) * 10) / 10, daysToFull };
|
|
}
|
|
|
|
export interface DiskForecast {
|
|
mount: string;
|
|
latest: number;
|
|
perDay: number;
|
|
daysToFull: number;
|
|
}
|
|
|
|
/** What a forecast needs before it is worth raising an alert about: enough history, and growth that really is steady. */
|
|
export const FORECAST_MIN_SPAN_DAYS = 3;
|
|
export const FORECAST_MIN_FIT = 0.9;
|
|
|
|
/**
|
|
* The disks that are heading for full within `withinDays`. A disk already at or above `alreadyAlertedAt` percent is left out — the
|
|
* ordinary "disk nearly full" alert covers it — and so is one whose growth is erratic (a big copy, a log that rotates), where a
|
|
* straight-line forecast would just be a guess. `alreadyAlertedAt` may be a function of the disk, for servers with a limit of their own
|
|
* for one of them; returning null leaves that disk out altogether (it has been told not to alert).
|
|
*/
|
|
export function disksFillingSoon(series: MetricSeries[], withinDays: number, alreadyAlertedAt: number | ((mount: string) => number | null)): DiskForecast[] {
|
|
const out: DiskForecast[] = [];
|
|
for (const s of series) {
|
|
if (s.kind !== "disk" || !s.trend || s.latest === null) continue;
|
|
const { perDay, fit, spanDays, daysToFull } = s.trend;
|
|
if (daysToFull === null || daysToFull > withinDays) continue;
|
|
if (spanDays < FORECAST_MIN_SPAN_DAYS || fit < FORECAST_MIN_FIT) continue;
|
|
const mount = s.metric.slice("disk:".length);
|
|
const limit = typeof alreadyAlertedAt === "function" ? alreadyAlertedAt(mount) : alreadyAlertedAt;
|
|
if (limit === null || s.latest >= limit) continue;
|
|
out.push({ mount, latest: s.latest, perDay, daysToFull });
|
|
}
|
|
return out.sort((a, b) => a.daysToFull - b.daysToFull);
|
|
}
|
|
|
|
export interface MetricSeries {
|
|
metric: string;
|
|
kind: "cpu" | "mem" | "disk";
|
|
label: string;
|
|
points: Point[];
|
|
latest: number | null;
|
|
trend: Trend | null;
|
|
}
|
|
|
|
function describe(metric: string): { kind: MetricSeries["kind"]; label: string } {
|
|
if (metric === "cpu") return { kind: "cpu", label: "CPU load" };
|
|
if (metric === "mem") return { kind: "mem", label: "Memory" };
|
|
return { kind: "disk", label: `Disk ${metric.slice("disk:".length)}` };
|
|
}
|
|
|
|
const ORDER = { cpu: 0, mem: 1, disk: 2 } as const;
|
|
|
|
/** Which of the kept figures answers a stretch of `days`: the samples while they reach back that far, then the hours, then the days. */
|
|
export function sourceFor(days: number): "samples" | "hourly" | "daily" {
|
|
return days <= RETENTION_DAYS ? "samples" : days <= HOURLY_KEEP_DAYS ? "hourly" : "daily";
|
|
}
|
|
|
|
/** How many points a chart over `days` is drawn from at most. */
|
|
const pointsFor = (days: number) => (days <= RETENTION_DAYS ? 96 : days <= HOURLY_KEEP_DAYS ? 200 : MAX_RANGE_DAYS + 5);
|
|
|
|
/** The latest sample of each figure of a server, whatever range is asked for. */
|
|
async function latestSamples(serverId: number): Promise<Map<string, number>> {
|
|
const rows = await db.all<{ metric: string; value: number }>(
|
|
sql`SELECT metric, value FROM server_metrics WHERE id IN (SELECT max(id) FROM server_metrics WHERE server_id = ${serverId} GROUP BY metric)`,
|
|
);
|
|
return new Map(rows.map((r) => [r.metric, r.value]));
|
|
}
|
|
|
|
/** Every series recorded for a server over the last `days` (up to a year), ready to draw. */
|
|
export async function getServerMetrics(serverId: number, days: number, now: Date = new Date()): Promise<MetricSeries[]> {
|
|
days = Math.min(MAX_RANGE_DAYS, Math.max(1, days));
|
|
const to = now.getTime();
|
|
const from = to - days * DAY_MS;
|
|
const source = sourceFor(days);
|
|
|
|
// The figures, per metric, as { t, v } for a trend and as points for the chart.
|
|
const raw = new Map<string, { t: number; v: number }[]>();
|
|
const points = new Map<string, Point[]>();
|
|
if (source === "samples") {
|
|
const rows = await db
|
|
.select({ metric: serverMetrics.metric, value: serverMetrics.value, at: serverMetrics.recordedAt })
|
|
.from(serverMetrics)
|
|
.where(and(eq(serverMetrics.serverId, serverId), gte(serverMetrics.recordedAt, new Date(from).toISOString())))
|
|
.orderBy(serverMetrics.recordedAt);
|
|
for (const r of rows) raw.set(r.metric, [...(raw.get(r.metric) ?? []), { t: Date.parse(r.at), v: r.value }]);
|
|
for (const [metric, list] of raw) points.set(metric, bucketPoints(list, from, to, pointsFor(days)));
|
|
} else {
|
|
const rows = await db
|
|
.select()
|
|
.from(serverMetricsRollup)
|
|
.where(and(eq(serverMetricsRollup.serverId, serverId), eq(serverMetricsRollup.grain, source === "hourly" ? "h" : "d"), gte(serverMetricsRollup.at, new Date(from).toISOString())))
|
|
.orderBy(serverMetricsRollup.at);
|
|
const byMetric = new Map<string, RollupRow[]>();
|
|
for (const r of rows) byMetric.set(r.metric, [...(byMetric.get(r.metric) ?? []), { t: Date.parse(r.at), sum: r.sum, n: r.n, min: r.min, max: r.max }]);
|
|
for (const [metric, list] of byMetric) {
|
|
raw.set(metric, rollupAverages(list));
|
|
points.set(metric, bucketRollups(list, from, to, pointsFor(days)));
|
|
}
|
|
}
|
|
const latest = source === "samples" ? null : await latestSamples(serverId);
|
|
|
|
return [...raw.entries()]
|
|
.map(([metric, list]) => {
|
|
const { kind, label } = describe(metric);
|
|
const newest = latest?.get(metric) ?? list[list.length - 1]?.v;
|
|
return {
|
|
metric,
|
|
kind,
|
|
label,
|
|
points: points.get(metric) ?? [],
|
|
latest: newest === undefined ? null : source === "samples" ? newest : round1(newest),
|
|
// A disk filling is the thing worth forecasting; load and memory go up and down by the hour.
|
|
trend: kind === "disk" ? trendOf(list) : null,
|
|
};
|
|
})
|
|
.sort((a, b) => ORDER[a.kind] - ORDER[b.kind] || a.label.localeCompare(b.label));
|
|
}
|
|
|
|
/** How far back the longer look at a disk goes when the last two weeks say it is flat. */
|
|
export const LONG_FORECAST_DAYS = 90;
|
|
|
|
/**
|
|
* The series to forecast from: the last two weeks, as before — except that a disk those two weeks call flat is looked at again over the last
|
|
* ninety days, so a slow, steady climb (a point a week) is seen too. The longer trend replaces the short one only if it is really rising.
|
|
*/
|
|
export async function getForecastSeries(serverId: number, now: Date = new Date()): Promise<MetricSeries[]> {
|
|
const short = await getServerMetrics(serverId, 14, now);
|
|
if (!short.some((s) => s.kind === "disk" && (!s.trend || s.trend.daysToFull === null))) return short;
|
|
const long = await getServerMetrics(serverId, LONG_FORECAST_DAYS, now);
|
|
return short.map((s) => {
|
|
if (s.kind !== "disk" || (s.trend && s.trend.daysToFull !== null)) return s;
|
|
const longer = long.find((l) => l.metric === s.metric);
|
|
return longer?.trend && longer.trend.daysToFull !== null ? { ...s, trend: longer.trend } : s;
|
|
});
|
|
}
|