@mason-data/cli 0.7.0 → 0.9.0

package/dist/src/duck/token.js

@@ -14,11 +14,14 @@ export function mintToken(apiKey, seconds, now = Date.now()) {
     const expiry = Math.floor(now / 1000) + seconds;
     return { token: `${PREFIX}_${expiry}_${mac(apiKey, expiry)}`, expiresAt: new Date(expiry * 1000).toISOString() };
 }
+const TOKEN_RE = /^mason_(\d{1,12})_([A-Za-z0-9_-]{43})$/;
+/** Whether a password is a token by its shape, good or not. */
+export const isToken = (password) => TOKEN_RE.test(password);
 /** null when the token is good; else why not, in a sentence a person reads after "password authentication failed" */
 export function tokenProblem(apiKey, token, now = Date.now()) {
     if (token === apiKey)
         return "the API key is not a password here: mason user add <name> makes a user whose password does not expire";
-    const m = /^mason_(\d{1,12})_([A-Za-z0-9_-]{43})$/.exec(token);
+    const m = TOKEN_RE.exec(token);
     if (m === null)
         return "not a Mason user's password or token: mason user add <name> makes a user whose password does not expire";
     const expiry = Number(m[1]);

package/dist/src/duck/types.js

@@ -11,8 +11,11 @@
  * UNION as its text. Each column that diverges carries the reason, and the runner logs it once per statement.
  */
 import { Buffer } from "node:buffer";
-import { DuckDBTypeId, DuckDBDateValue, DuckDBTimestampValue, DuckDBTimestampTZValue, DuckDBDecimalValue, DuckDBUUIDValue, DuckDBListValue, DuckDBArrayValue, DuckDBIntervalValue, DuckDBBlobValue, DuckDBTimeValue, DuckDBListType, DuckDBArrayType, DuckDBTimestampSecondsValue, DuckDBTimestampMillisecondsValue, DuckDBTimestampNanosecondsValue, } from "@duckdb/node-api";
+import { DuckDBTypeId, DuckDBDateValue, DuckDBTimestampValue, DuckDBTimestampTZValue, DuckDBDecimalValue, DuckDBUUIDValue, DuckDBListValue, DuckDBArrayValue, DuckDBIntervalValue, DuckDBBlobValue, DuckDBTimeValue, DuckDBListType, DuckDBArrayType, DuckDBTimeTZValue, DuckDBTimestampSecondsValue, DuckDBTimestampMillisecondsValue, DuckDBTimestampNanosecondsValue, } from "@duckdb/node-api";
 import { jsonbText } from "./jsonb.js";
+import { gapMark } from "./gaps.js";
+import { heldType, ZONED_HELD } from "./held-types.js";
+import { pgEvalSync } from "./pg-eval.js";
 import { EncodeError, encodeValue, OID, TIMESTAMP } from "../api/postgres/encode.js";
 const MICROS_PER_DAY = 86400000000n;
 const DATE_INFINITY = 2_147_483_647, DATE_NEG_INFINITY = -2_147_483_647;
@@ -124,7 +127,7 @@ function timestampColumn(name, tz) {
                 }
                 // a zone's offset before year 1 is its local mean time, which Mason does not compute
                 if (tz && s.timeZone.toUpperCase() !== "UTC")
-                    throw new EncodeError(`Mason writes a timestamptz before year 1 only in UTC, not in ${s.timeZone} (Mason gap bc-timestamptz-zone)`, "0A000");
+                    throw new EncodeError(`Mason writes a timestamptz before year 1 only in UTC, not in ${s.timeZone} (${gapMark("bc-timestamptz-zone")})`, "0A000", "bc-timestamptz-zone");
                 return Buffer.from(bcTimestampText(m, tz), "ascii");
             }
             return base.encode(v, format, s);
@@ -136,6 +139,12 @@ function timeText(micros) {
     const secs = micros / 1000000n, frac = micros % 1000000n;
     return `${pad(secs / 3600n)}:${pad((secs % 3600n) / 60n)}:${pad(secs % 60n)}${frac === 0n ? "" : `.${pad(frac, 6).replace(/0+$/, "")}`}`;
 }
+/** A timetz as timetz_out writes it: the time, then its offset as `+HH`, `+HH:MM` or `+HH:MM:SS` (DuckDB's offset, as
+ *  Postgres writes it, is east of UTC; its stored zone is west) */
+export function timetzText(v) {
+    const abs = Math.abs(v.offset), s = abs % 60, m = Math.trunc(abs / 60) % 60;
+    return `${timeText(v.micros)}${v.offset < 0 ? "-" : "+"}${pad(Math.trunc(abs / 3600))}${m !== 0 || s !== 0 ? `:${pad(m)}` : ""}${s !== 0 ? `:${pad(s)}` : ""}`;
+}
 /** An interval as Postgres's `interval_out` writes it under IntervalStyle postgres: `1 year 2 mons 3 days 04:05:06.7` */
 export function intervalText(months, days, micros) {
     const parts = [];
@@ -212,7 +221,7 @@ function arrayElement(text) {
 /** Postgres's array OIDs of the element OIDs this file sends */
 const ARRAY_OID = {
     16: 1000, 17: 1001, 20: 1016, 21: 1005, 23: 1007, 25: 1009, 700: 1021, 701: 1022, 1082: 1182, 1083: 1183,
-    1114: 1115, 1184: 1185, 1186: 1187, 1700: 1231, 2950: 2951, 3802: 3807,
+    1114: 1115, 1184: 1185, 1186: 1187, 1266: 1270, 1700: 1231, 2950: 2951, 3802: 3807,
 };
 function custom(name, oid, size, text, binary, divergence = null) {
     return {
@@ -267,7 +276,7 @@ export function wireColumn(name, t) {
         case DuckDBTypeId.UHUGEINT:
         case DuckDBTypeId.UBIGINT:
         case DuckDBTypeId.BIGNUM:
-            return viaApi(name, "numeric", bigintText, `DuckDB's ${t.toString()} (a sum of integers is one), sent as numeric: Postgres says bigint for a sum of int4 and numeric for a sum of int8`);
+            return viaApi(name, "numeric", bigintText, `the engine's ${t.toString()} (a sum of integers is one), sent as numeric: Postgres says bigint for a sum of int4 and numeric for a sum of int8`);
         case DuckDBTypeId.FLOAT: return viaApi(name, "float4", (v) => v);
         case DuckDBTypeId.DOUBLE: return viaApi(name, "float8", (v) => v);
         case DuckDBTypeId.DECIMAL: return viaApi(name, "numeric", (v) => v.toString());
@@ -284,6 +293,8 @@ export function wireColumn(name, t) {
         case DuckDBTypeId.TIMESTAMP_TZ: return timestampColumn(name, true);
         case DuckDBTypeId.TIME:
             return custom(name, 1083, 8, (v) => timeText(v.micros), (v) => { const b = Buffer.alloc(8); b.writeBigInt64BE(v.micros); return b; });
+        case DuckDBTypeId.TIME_TZ:
+            return custom(name, 1266, 12, (v) => timetzText(v), (v) => { const t = v; const b = Buffer.alloc(12); b.writeBigInt64BE(t.micros, 0); b.writeInt32BE(-t.offset, 8); return b; });
         case DuckDBTypeId.INTERVAL:
             return custom(name, 1186, 16, (v) => { const i = v; return intervalText(i.months, i.days, i.micros); }, (v) => { const i = v; const b = Buffer.alloc(16); b.writeBigInt64BE(i.micros, 0); b.writeInt32BE(i.days, 8); b.writeInt32BE(i.months, 12); return b; });
         case DuckDBTypeId.BLOB:
@@ -303,7 +314,7 @@ export function wireColumn(name, t) {
         }
         default: break;
     }
-    return custom(name, 25, -1, (v) => String(v), null, `DuckDB's ${t.toString()} has no Postgres type here: sent as its text`);
+    return custom(name, 25, -1, (v) => String(v), null, `the engine's ${t.toString()} has no Postgres type here: sent as its text`);
 }
 // ---------------------------------------------------------------------------------------------------------------------
 // A COLUMN OF POSTGRES'S TYPE (docs/DESIGN.md "Postgres decides, DuckDB computes"): the catalog Postgres's
@@ -312,7 +323,7 @@ export function wireColumn(name, t) {
 // int4, its DECIMAL as a numeric, its JSON as jsonb's canonical text. Where DuckDB's value cannot be written as
 // Postgres's type (a timestamp where Postgres says date), DuckDB's own encoding stands and the column says why.
 /** pg_type.typlen of the OIDs a result column may have */
-const TYPLEN = { 16: 1, 18: 1, 19: 64, 20: 8, 21: 2, 23: 4, 26: 4, 700: 4, 701: 8, 1082: 4, 1083: 8, 1114: 8, 1184: 8, 1186: 16, 2950: 16 };
+const TYPLEN = { 16: 1, 18: 1, 19: 64, 20: 8, 21: 2, 23: 4, 26: 4, 700: 4, 701: 8, 1082: 4, 1083: 8, 1114: 8, 1184: 8, 1186: 16, 1266: 12, 2950: 16 };
 const PG_TEXT = new Set([25, 1043, 1042, 19, 18]);
 const PG_INT = { 21: "int2", 23: "int4", 20: "int8" };
 const DUCK_EXACT = new Set([DuckDBTypeId.TINYINT, DuckDBTypeId.SMALLINT, DuckDBTypeId.INTEGER, DuckDBTypeId.BIGINT, DuckDBTypeId.HUGEINT,
@@ -348,6 +359,45 @@ export function floatAsNumeric(x, digits = 15) {
     }
     return s === "-0" ? "0" : s;
 }
+/** the binary forms of the text-held types a driver asks for by default (pgjdbc: point, box): each a float8's bytes */
+const HELD_BINARY = {
+    600: floats, 603: floats,
+};
+function floats(text) {
+    const parts = text.match(/-?(?:Infinity|NaN|[\d.]+(?:e[-+]?\d+)?)/g) ?? [];
+    const b = Buffer.alloc(parts.length * 8);
+    parts.forEach((p, i) => b.writeDoubleBE(Number(p), i * 8));
+    return b;
+}
+/** a time range's text in the session's zone: Postgres's own output, a value at a time, the last few remembered */
+const zonedTexts = new Map();
+function inZone(type, text, zone) {
+    const key = `${zone}\u0000${type}\u0000${text}`;
+    let written = zonedTexts.get(key);
+    if (written === undefined) {
+        written = pgEvalSync(`SELECT $1::${type}::text`, [text], zone)[0]?.[0] ?? text;
+        if (zonedTexts.size >= 1024)
+            zonedTexts.clear();
+        zonedTexts.set(key, written);
+    }
+    return written;
+}
+/** a column of a type held as its text. A value's text only: Mason writes the binary form of few of them */
+function heldColumn(name, oid, own, size) {
+    const binary = HELD_BINARY[oid], zoned = ZONED_HELD.get(oid);
+    return {
+        name, oid, size, divergence: null,
+        encode: (v, format, s) => {
+            if (v === null)
+                return null;
+            if (format === 0)
+                return zoned !== undefined ? Buffer.from(inZone(zoned, str(v), s.timeZone), "utf8") : own.encode(v, 0, s);
+            if (binary !== undefined)
+                return binary(str(v));
+            throw new EncodeError(`column ${JSON.stringify(name)} (type OID ${oid}) is held as its text and cannot be sent in binary by Mason: ask for text (${gapMark("binary-held-type")})`, "0A000", "binary-held-type");
+        },
+    };
+}
 /** The wire column of a result column whose Postgres type the analysis gave, holding DuckDB values of type `t`. */
 export function pgTypedColumn(name, oid, t, typmod = -1) {
     const own = wireColumn(name, t);
@@ -359,6 +409,16 @@ export function pgTypedColumn(name, oid, t, typmod = -1) {
         return retyped(viaApi(name, "numeric", (v) => padDecimals(exactText(v), (typmod - 4) & 0xffff)));
     if (own.oid === oid)
         return retyped(own);
+    // a type the copy holds as its text, or an array of one (src/duck/held-types.ts): the text, under Postgres's OID
+    const held = heldType(oid);
+    if (held !== undefined && (own.oid === 25 || own.oid === 1009))
+        return heldColumn(name, oid, own, held.oid === oid ? held.size : -1);
+    // an oid element of an oid[], held as its digits
+    if (oid === 26 && own.oid === 25)
+        return custom(name, 26, 4, str, (v) => { const b = Buffer.alloc(4); b.writeUInt32BE(Number(str(v))); return b; });
+    // an oid: DuckDB's UINTEGER
+    if (oid === 26 && DUCK_EXACT.has(t.typeId))
+        return custom(name, 26, 4, (v) => exactText(v), (v) => { const b = Buffer.alloc(4); b.writeUInt32BE(Number(exactText(v))); return b; });
     // an inet: stored as Postgres's text with its mask (`10.3.6.45/32`), sent as inet_out writes it (a full mask dropped)
     if (oid === 869 && own.oid === 25)
         return custom(name, 869, -1, (v) => inetOut(str(v)), (v) => inetBytes(str(v)));
@@ -399,16 +459,13 @@ export function pgTypedColumn(name, oid, t, typmod = -1) {
     if (oid === 1700 && DUCK_FLOAT.has(t.typeId)) {
         // a double that overflowed is Postgres's numeric overflow (its exact arithmetic raises it)
         return { ...retyped(viaApi(name, "numeric", (v) => { if (v === Infinity || v === -Infinity)
-                throw new EncodeError("value overflows numeric format", "22003"); return floatAsNumeric(v); })), divergence: `Postgres computes this numeric exactly; DuckDB computed it as a double (${t.toString()}), sent with 15 significant digits` };
+                throw new EncodeError("value overflows numeric format", "22003"); return floatAsNumeric(v); })), divergence: `Postgres computes this numeric exactly; the engine computed it as a double (${t.toString()}), sent with 15 significant digits` };
     }
-    return { ...own, divergence: `Postgres says type OID ${oid}; DuckDB computed ${t.toString()}, sent as its own type` };
+    return { ...own, divergence: `Postgres says type OID ${oid}; the engine computed ${t.toString()}, sent as its own type` };
 }
-/** How a cell of an exact numeric (src/duck/rewrite.ts ExactColumn) reads as a decimal text, by DuckDB's type: an exact
- *  number's digits, a float as Postgres's cast to numeric writes it — or null where the cell needs digits DuckDB did not
- *  keep (a double where the program wants exact ones). */
 export function cellText(t, conv) {
     // a numeric's text as Postgres's numeric_out wrote it (a text read by numeric_in, src/duck/pg-eval.ts)
-    if (conv === "text")
+    if (conv === "text" || conv === "parsed")
         return t.typeId === DuckDBTypeId.VARCHAR ? (v) => String(v) : null;
     if (DUCK_EXACT.has(t.typeId))
         return exactText;

package/dist/src/duck/usage.js

@@ -0,0 +1,326 @@
+import { MASON_SCHEMA } from "./store.js";
+import { History, Streak, cutoff } from "./history.js";
+import { outcomeOf } from "../api/request-log.js";
+export const USAGE_TABLE = `${MASON_SCHEMA}.usage`;
+export const SIZES_TABLE = `${MASON_SCHEMA}.copy_sizes`;
+/**
+ * HOW OFTEN A SAMPLE IS TAKEN: a minute. A chart of a day is 1,440 points and of the retention's 7 days 10,080 (a few
+ * hundred KB of rows); a statement's p95 over a minute of a BI tool's traffic is a number, over 5 s mostly one statement;
+ * and a sample is a read of the state the commands already ask every few seconds. No setting: nothing an operator
+ * would tune it for has been met.
+ */
+export const SAMPLE_SECONDS = 60;
+/** how often each copy's own bytes are measured (`copySizes`): its cost grows with the copy, see above */
+export const SIZE_EVERY_MS = 10 * 60_000;
+const COLUMNS = ["sampled_at", "seconds", "copy_bytes", "copy_rows", "tables_added", "tables_copied", "tables_failed", "oldest_sync_seconds",
+    "statements", "statements_refused", "statements_failed", "statements_cancelled", "rows_sent", "bytes_sent", "duration_p50_ms", "duration_p95_ms",
+    "stream_lag_bytes", "stream_lag_seconds", "rss_bytes", "heap_used_bytes", "copy_memory_bytes", "copy_memory_limit_bytes", "slot_wal_held_bytes", "disk_free_bytes", "sessions_open"];
+/** the samples' DDL: one row a sample; a level is as of `sampled_at`, a count is of the `seconds` before it */
+export const USAGE_DDL = `CREATE TABLE IF NOT EXISTS ${USAGE_TABLE} (sampled_at TIMESTAMPTZ, seconds DOUBLE, copy_bytes BIGINT, copy_rows BIGINT, ` +
+    `tables_added INTEGER, tables_copied INTEGER, tables_failed INTEGER, oldest_sync_seconds DOUBLE, statements BIGINT, statements_refused BIGINT, ` +
+    `statements_failed BIGINT, statements_cancelled BIGINT, rows_sent BIGINT, bytes_sent BIGINT, duration_p50_ms INTEGER, duration_p95_ms INTEGER, stream_lag_bytes BIGINT, ` +
+    `stream_lag_seconds DOUBLE, rss_bytes BIGINT, heap_used_bytes BIGINT, copy_memory_bytes BIGINT, copy_memory_limit_bytes BIGINT)`;
+/** the columns added since the first samples (a file a run before them made): each added where it is missing */
+export const USAGE_ADDED = ["slot_wal_held_bytes BIGINT", "disk_free_bytes BIGINT", "sessions_open INTEGER"];
+/** each copy's own bytes, every SIZE_EVERY_MS */
+export const SIZES_DDL = `CREATE TABLE IF NOT EXISTS ${SIZES_TABLE} (sampled_at TIMESTAMPTZ, table_name VARCHAR, bytes BIGINT)`;
+const noInterval = () => ({ statements: 0, outcomes: { refused: 0, failed: 0, cancelled: 0 }, rowsSent: 0, bytesSent: 0, durations: [], completed: 0 });
+/**
+ * THE COMPLETED STATEMENTS' DURATIONS ONE INTERVAL KEEPS, at most: an interval is a minute, but one whose sample cannot
+ * be taken is carried into the next (`sample`), so while samples keep failing it would otherwise keep every statement's
+ * duration for as long as they fail. Up to this many the p50 / p95 are exact; past it the interval keeps a uniform
+ * random sample of this many (a reservoir: the n-th duration replaces a kept one with probability KEPT / n), whose
+ * quantiles estimate the interval's — 100,000 numbers are ~800 KB, ~1,700 statements a second for a minute.
+ */
+export const DURATIONS_KEPT = 100_000;
+/**
+ * The q-quantile of sorted values as DuckDB's `quantile_cont` gives it (linear between the two nearest ranks), rounded
+ * as the request log's summary rounds it, or null for none: the summary's own definition of a p50 and a p95
+ * (tests/duck-usage.test.ts holds the two equal).
+ */
+export function quantileCont(sorted, q) {
+    if (sorted.length === 0)
+        return null;
+    const at = (sorted.length - 1) * q, lo = Math.floor(at), hi = Math.ceil(at);
+    return Math.round(sorted[lo] + (sorted[hi] - sorted[lo]) * (at - lo));
+}
+/**
+ * `quantileCont` over a HISTOGRAM: `[value, count]` pairs sorted by value (each count at least 0), as though every
+ * value were listed its count of times — the same ranks and the same interpolation, rounded the same way, without
+ * listing them (the request log's summary reads a week of durations this way: src/duck/requests.ts `summary`). Null
+ * for none; tests/duck-usage.test.ts holds it equal to DuckDB's `quantile_cont`.
+ */
+export function quantileContOf(histogram, q) {
+    const n = histogram.reduce((s, [, c]) => s + c, 0);
+    if (n === 0)
+        return null;
+    const at = (n - 1) * q, lo = Math.floor(at), hi = Math.ceil(at);
+    // the value at a 0-based rank: the first whose running count passes it
+    const valueAt = (rank) => {
+        let seen = 0;
+        for (const [v, c] of histogram) {
+            seen += c;
+            if (rank < seen)
+                return v;
+        }
+        return histogram[histogram.length - 1][0];
+    };
+    const a = valueAt(lo), b = hi === lo ? a : valueAt(hi);
+    return Math.round(a + (b - a) * (at - lo));
+}
+export class UsageSampler {
+    interval = noInterval();
+    totals = { completed: 0, refused: 0, failed: 0, cancelled: 0, completedMs: 0 };
+    latest = null;
+    last;
+    timer = null;
+    sampling = Promise.resolve();
+    take = null;
+    lastSized = -Infinity;
+    sizedOnce = false;
+    /** a streak of failures, said once: what failed (a sample taken, written, sized), and how many */
+    lost = new Map();
+    closed = false;
+    history;
+    log;
+    options;
+    days;
+    memoryLimitBytes;
+    sampleSeconds;
+    constructor(history, log, memoryLimitBytes, options) {
+        this.history = history;
+        this.log = log;
+        this.options = options;
+        this.days = options.days;
+        this.memoryLimitBytes = memoryLimitBytes;
+        this.sampleSeconds = (options.sampleMs ?? SAMPLE_SECONDS * 1000) / 1000;
+        this.last = this.now();
+    }
+    /** The samples of a store: their tables made, the rows past the retention deleted, and a trim every hour. Sampling
+     *  begins with `start`, once the run's facts can be read. */
+    static async open(store, log, options) {
+        const history = await History.open(store, [USAGE_DDL, ...USAGE_ADDED.map((c) => `ALTER TABLE ${USAGE_TABLE} ADD COLUMN IF NOT EXISTS ${c}`), SIZES_DDL]);
+        const u = new UsageSampler(history, log, store.memoryLimitBytes, options);
+        await history.keepTrimmed(() => u.trim(), (e) => log(`usage: deleting the samples older than ${u.days} days failed: ${e.message}`));
+        return u;
+    }
+    now() { return this.options.now?.() ?? Date.now(); }
+    /** One row of the listener's, as the request log records it: counted in the interval it ends, by the summary's rule
+     *  (a refused login is no statement; `outcomeOf`, the one outcome rule). Never throws. */
+    count(row) {
+        if (row.method === "CONNECT")
+            return;
+        const i = this.interval;
+        i.statements++;
+        const outcome = outcomeOf(row.sqlstate);
+        if (outcome === "ok") {
+            this.totals.completed++;
+            this.totals.completedMs += row.durationMs;
+        }
+        else
+            this.totals[outcome]++;
+        if (outcome === "ok") {
+            i.completed++;
+            const kept = this.options.durationsKept ?? DURATIONS_KEPT;
+            if (i.durations.length < kept)
+                i.durations.push(row.durationMs);
+            else {
+                const j = Math.floor((this.options.random?.() ?? Math.random()) * i.completed);
+                if (j < kept)
+                    i.durations[j] = row.durationMs;
+            }
+        }
+        else
+            i.outcomes[outcome]++;
+        i.rowsSent += row.rows ?? 0;
+        i.bytesSent += row.bytesSent ?? 0;
+    }
+    /** The statements since the process started, by outcome (a copy: the counter itself is never handed out). */
+    statementTotals() { return { ...this.totals }; }
+    /** The last sample's p50 and p95 of its completed statements; null before the first sample. */
+    lastDurations() { return this.latest; }
+    /** Takes a sample every interval from `take` (the run's facts), and each copy's bytes every SIZE_EVERY_MS. */
+    start(take) {
+        this.take = take;
+        this.schedule();
+    }
+    schedule() {
+        if (this.closed)
+            return;
+        // one after the other, never two at once: a sample slower than its interval delays the next
+        this.timer = setTimeout(() => { this.timer = null; this.sampling = this.sample().then(() => this.schedule()); }, this.options.sampleMs ?? SAMPLE_SECONDS * 1000);
+        this.timer.unref();
+    }
+    /** a failure of `what`, said once per streak; `ok` ends the streak, said with what it cost */
+    failure(what, e) {
+        const streak = this.lost.get(what) ?? new Streak();
+        this.lost.set(what, streak);
+        streak.failed(1, () => this.log(`usage: ${what} failed: ${e.message}; said once until one succeeds`));
+    }
+    ok(what) {
+        this.lost.get(what)?.succeeded((n) => this.log(`usage: ${what} again; ${n} time(s) it was not`));
+    }
+    /** One sample now: the run's facts, the interval's requests, the memory; written. Resolves whatever happens (a failure is said). */
+    async sample() {
+        if (this.take === null)
+            return;
+        let facts;
+        let copyMemory;
+        try {
+            facts = await this.take();
+            copyMemory = await this.copyMemory();
+        }
+        catch (e) {
+            // Tolerated, narrowly: a sample is history, and the run goes on without it. The interval's counts are KEPT for
+            // the next sample, whose `seconds` covers both, so no request goes uncounted.
+            this.failure("taking a sample", e);
+            return;
+        }
+        this.ok("taking a sample");
+        const now = this.now();
+        const i = this.interval;
+        this.interval = noInterval();
+        const seconds = (now - this.last) / 1000;
+        this.last = now;
+        const sorted = i.durations.sort((a, b) => a - b);
+        const mem = this.options.memory?.() ?? process.memoryUsage();
+        // the p50 and p95 written are the ones GET /metrics says (lastDurations): one computation
+        this.latest = { at: new Date(now).toISOString(), p50Ms: quantileCont(sorted, 0.5), p95Ms: quantileCont(sorted, 0.95) };
+        const values = [this.latest.at, seconds, facts.copyBytes, facts.totals.rows, facts.totals.added, facts.totals.copied, facts.totals.failed,
+            facts.oldestSyncSeconds, i.statements, i.outcomes.refused, i.outcomes.failed, i.outcomes.cancelled, i.rowsSent, i.bytesSent, this.latest.p50Ms, this.latest.p95Ms,
+            facts.streamLag?.bytes ?? null, facts.streamLag?.seconds ?? null, mem.rss, mem.heapUsed, copyMemory, this.memoryLimitBytes, facts.slotWalHeldBytes, facts.diskFreeBytes, facts.sessionsOpen];
+        const sql = `INSERT INTO ${USAGE_TABLE} (${COLUMNS.join(", ")}) VALUES (${COLUMNS.map((c, j) => `$${j + 1}${c === "sampled_at" ? "::TIMESTAMPTZ" : ""}`).join(", ")})`;
+        try {
+            if (this.options.insert !== undefined)
+                await this.options.insert(sql, values);
+            else
+                await this.history.writer.run(sql, values);
+            this.ok(`writing a sample to ${USAGE_TABLE}`);
+        }
+        catch (e) {
+            // Tolerated, narrowly, as the request log's: a sample that cannot be written must not stop the run. SAID.
+            this.failure(`writing a sample to ${USAGE_TABLE}`, e);
+        }
+        if (now - this.lastSized >= (this.options.sizeEveryMs ?? SIZE_EVERY_MS)) {
+            this.lastSized = now;
+            await this.size(facts.copied, now);
+        }
+    }
+    /** the copy's memory in use: DuckDB's buffer manager's counters, summed */
+    async copyMemory() {
+        const v = (await this.history.reader.runAndReadAll("SELECT sum(memory_usage_bytes) FROM duckdb_memory()")).getRowsJS()[0]?.[0];
+        return v === null || v === undefined ? null : Number(v);
+    }
+    /** Each served copy's own bytes, written as one batch (see the module's comment for how, and what it costs). */
+    async size(copied, now) {
+        const t0 = performance.now();
+        const rows = [];
+        try {
+            const want = new Set(copied);
+            const tables = (await this.history.reader.runAndReadAll(`SELECT schema_name, table_name FROM duckdb_tables() WHERE NOT internal AND database_name = current_database() AND schema_name <> '${MASON_SCHEMA}'`))
+                .getRowsJS();
+            // THE COPY'S OWN database: pragma_database_size() has a row for every attached one, and the in-memory catalog the
+            // visible names live in (src/duck/kept.ts) has block_size 0 — read first, it made every copy "0 B"
+            const block = Number((await this.history.reader.runAndReadAll("SELECT block_size FROM pragma_database_size() WHERE database_name = current_database()")).getRowsJS()[0][0]);
+            for (const [schema, table] of tables) {
+                const name = `${schema}.${table}`;
+                if (!want.has(name))
+                    continue;
+                const q = `${quote(schema)}.${quote(table)}`;
+                let blocks;
+                try {
+                    blocks = Number((await this.history.reader.runAndReadAll(`WITH s AS (SELECT block_id, additional_block_ids FROM pragma_storage_info($1) WHERE persistent), ` +
+                        `b AS (SELECT block_id FROM s WHERE block_id >= 0 UNION SELECT unnest(additional_block_ids) FROM s) SELECT count(*) FROM b`, [q])).getRowsJS()[0][0]);
+                }
+                catch (e) {
+                    // a copy dropped since the list was read (mason remove; a copy swapped in again): not sized this time, by design.
+                    // Anything else is the sizing's failure, said
+                    if (/does not exist/i.test(e.message))
+                        continue;
+                    throw e;
+                }
+                // NO BLOCK of its own yet is not 0 bytes: its rows are still in the write-ahead log (a small copy is written to
+                // the file only at a checkpoint), so it is not measured this time — a page said "0.0 KiB" for a 5,000-row copy
+                if (blocks === 0)
+                    continue;
+                rows.push([name, blocks * block]);
+            }
+            if (rows.length > 0) {
+                const sql = `INSERT INTO ${SIZES_TABLE} (sampled_at, table_name, bytes) VALUES ` + rows.map((_, i) => `($1::TIMESTAMPTZ, $${i * 2 + 2}, $${i * 2 + 3})`).join(", ");
+                const values = [new Date(now).toISOString(), ...rows.flat()];
+                if (this.options.insert !== undefined)
+                    await this.options.insert(sql, values);
+                else
+                    await this.history.writer.run(sql, values);
+            }
+        }
+        catch (e) {
+            this.failure("measuring each synced table's size", e);
+            return;
+        }
+        this.ok("measuring each synced table's size");
+        // the first measure of a run is said with its time: what it costs on this machine's copies, read from the log
+        if (!this.sizedOnce) {
+            this.sizedOnce = true;
+            this.log(`usage: measured ${rows.length} synced tables' own size in ${Math.round(performance.now() - t0)} ms (every ${Math.round((this.options.sizeEveryMs ?? SIZE_EVERY_MS) / 60_000)} min)`);
+        }
+    }
+    /** Deletes the samples and sizes older than the retention, counted back from the module's ONE clock (`now`, which
+     *  stamps each sample), never the engine's `now()` (src/duck/events.ts `trim` says why). */
+    async trim() {
+        const before = cutoff(this.now(), this.days);
+        await this.history.writer.run(`DELETE FROM ${USAGE_TABLE} WHERE sampled_at < $1::TIMESTAMPTZ`, [before]);
+        await this.history.writer.run(`DELETE FROM ${SIZES_TABLE} WHERE sampled_at < $1::TIMESTAMPTZ`, [before]);
+    }
+    /**
+     * The samples from `since` until `until`, oldest first; with `stepSeconds`, re-bucketed in ONE query: buckets of that
+     * many seconds from `since`, each count summed and each level at its highest (the disk free at its lowest). Committed rows only.
+     */
+    async samples(o) {
+        const values = [o.since, o.until];
+        const range = `sampled_at >= $1::TIMESTAMPTZ AND sampled_at < $2::TIMESTAMPTZ`;
+        let sql;
+        if (o.stepSeconds === undefined) {
+            sql = `SELECT epoch_ms(sampled_at) AS at, ${COLUMNS.slice(1).join(", ")} FROM ${USAGE_TABLE} WHERE ${range} ORDER BY sampled_at`;
+        }
+        else {
+            values.push(o.stepSeconds);
+            const sum = new Set(["seconds", "statements", "statements_refused", "statements_failed", "statements_cancelled", "rows_sent", "bytes_sent"]);
+            sql = `SELECT epoch_ms(time_bucket(to_seconds($3::BIGINT), sampled_at, $1::TIMESTAMPTZ)) AS at, ` +
+                COLUMNS.slice(1).map((c) => `${sum.has(c) ? "sum" : c === "disk_free_bytes" ? "min" : "max"}(${c}) AS ${c}`).join(", ") +
+                ` FROM ${USAGE_TABLE} WHERE ${range} GROUP BY 1 ORDER BY 1`;
+        }
+        const rows = (await this.history.reader.runAndReadAll(sql, values)).getRowObjectsJson();
+        return rows.map(sampleOf);
+    }
+    /** Each copy's own bytes as last measured within the retention, by table. */
+    async copySizes() {
+        const rows = (await this.history.reader.runAndReadAll(`SELECT table_name, arg_max(bytes, sampled_at) AS bytes, epoch_ms(max(sampled_at)) AS at FROM ${SIZES_TABLE} WHERE bytes > 0 GROUP BY table_name ORDER BY table_name`))
+            .getRowsJS();
+        return rows.map(([table, bytes, at]) => ({ table, bytes: Number(bytes), at: new Date(Number(at)).toISOString() }));
+    }
+    /** Stops sampling, waits for a sample in flight, and closes the samples' connections (mason stop). */
+    async close() {
+        this.closed = true;
+        if (this.timer !== null) {
+            clearTimeout(this.timer);
+            this.timer = null;
+        }
+        this.history.stopTrimming();
+        await this.sampling;
+        this.history.closeConnections();
+    }
+}
+const quote = (name) => `"${name.replaceAll('"', '""')}"`;
+const num = (v) => (v === null || v === undefined ? null : Number(v));
+function sampleOf(r) {
+    return {
+        at: new Date(Number(r.at)).toISOString(), seconds: Number(r.seconds), copyBytes: num(r.copy_bytes), copyRows: Number(r.copy_rows ?? 0),
+        tablesAdded: Number(r.tables_added ?? 0), tablesCopied: Number(r.tables_copied ?? 0), tablesFailed: Number(r.tables_failed ?? 0), oldestSyncSeconds: num(r.oldest_sync_seconds),
+        statements: Number(r.statements ?? 0), statementsRefused: Number(r.statements_refused ?? 0), statementsFailed: Number(r.statements_failed ?? 0),
+        statementsCancelled: Number(r.statements_cancelled ?? 0), rowsSent: Number(r.rows_sent ?? 0), bytesSent: Number(r.bytes_sent ?? 0), durationP50Ms: num(r.duration_p50_ms), durationP95Ms: num(r.duration_p95_ms),
+        streamLagBytes: num(r.stream_lag_bytes), streamLagSeconds: num(r.stream_lag_seconds),
+        rssBytes: Number(r.rss_bytes), heapUsedBytes: Number(r.heap_used_bytes), copyMemoryBytes: num(r.copy_memory_bytes), copyMemoryLimitBytes: Number(r.copy_memory_limit_bytes),
+        slotWalHeldBytes: num(r.slot_wal_held_bytes), diskFreeBytes: num(r.disk_free_bytes), sessionsOpen: num(r.sessions_open),
+    };
+}