@ixo/oracle-runtime-workers 0.33.2 → 0.34.1

package/src/plugins/ixo-chain/lookup.ts

@@ -42,6 +42,8 @@ import type {
   KindRestOptions,
   KindResult,
 } from './kinds/types';
+import { dropLastName, pruneNames } from './names/collect';
+import type { NameMap } from './names/types';
 import { firstOf, type Subject } from './queries';
 import { restJson, type Absent } from './rest';
 import { isRecord } from './summaries';
@@ -395,6 +397,60 @@ export function shrink(
   }
 }
 
+/**
+ * One row cut: the last row of the largest list inside `within` left out
+ * and `holder` marked `truncated`; false when no list has more than one
+ * row left.
+ */
+export function cutRow(
+  holder: { truncated?: boolean },
+  within: unknown,
+): boolean {
+  const found = largestArray(within);
+  if (!found) return false;
+  found.list.pop();
+  holder.truncated = true;
+  return true;
+}
+
+/**
+ * One row cut across `results`: the last row of the largest list in any
+ * `ok` result left out and that result marked `truncated`; false when no
+ * result has a list of more than one row left.
+ */
+export function cutRowAcross(results: LookupResult[]): boolean {
+  let largest: { result: LookupResult; size: number } | undefined;
+  for (const result of results) {
+    if (!result.ok) continue;
+    const found = largestArray(result.data);
+    if (found && (!largest || found.size > largest.size))
+      largest = { result, size: found.size };
+  }
+  return largest?.result.ok
+    ? cutRow(largest.result, largest.result.data)
+    : false;
+}
+
+/**
+ * Fits `answer`, which carries `names`, into {@link MAX_RESULT_CHARS}.
+ * Rows go first, one `cutRows` at a time (one row each: {@link cutRow},
+ * {@link cutRowAcross}), each cut followed by leaving out the names of
+ * identifiers no longer in the answer; with no row left to cut, names go
+ * from the end of the map (the least placed identifiers). An answer is
+ * never left out whole to make room for names. An empty map is removed.
+ */
+export function fitNames(
+  answer: { names?: NameMap },
+  cutRows: () => boolean,
+): void {
+  while (size(answer) > MAX_RESULT_CHARS) {
+    if (!cutRows()) break;
+    pruneNames(answer);
+  }
+  while (size(answer) > MAX_RESULT_CHARS) if (!dropLastName(answer)) break;
+  if (answer.names && !Object.keys(answer.names).length) delete answer.names;
+}
+
 /**
  * Fits `parts` (the results inside `whole`) into what
  * {@link MAX_RESULT_CHARS} leaves after the envelope: each part gets an
@@ -446,18 +502,22 @@ export function fitResults(whole: unknown, results: LookupResult[]): void {
   }
 }
 
+export interface LookupAnswer {
+  ok: true;
+  network: ChainCall['network'];
+  fetchedAt: string;
+  results: LookupResult[];
+  /** What the identifiers in the results are called (added by the tool). */
+  names?: NameMap;
+}
+
 /** `chain_lookup`'s answer: every request's result in order, cut to size. */
 export async function chainLookup(
   call: ChainCall,
   identity: UserIdentity,
   requests: readonly KindRequest[],
   options: LookupOptions,
-): Promise<{
-  ok: true;
-  network: ChainCall['network'];
-  fetchedAt: string;
-  results: LookupResult[];
-}> {
+): Promise<LookupAnswer> {
   const results = await runRequests(call, identity, requests, options);
   const whole = {
     ok: true as const,

package/src/plugins/ixo-chain/names/client.ts

@@ -0,0 +1,467 @@
+/**
+ * The IXO Names worker client: names for the identifiers of a tool result,
+ * asked of the shared worker (`POST <base>/v1/resolve`).
+ *
+ * Per isolate, three things sit in front of the worker:
+ * - a memo of its answers (at most {@link MEMO_ENTRIES}, each kept until
+ *   the worker's `freshUntil`, at most {@link MEMO_MAX_MS} and at least
+ *   {@link MEMO_MIN_MS}); partial answers are never kept;
+ * - a batcher per worker and client id: the ids every call misses are
+ *   gathered for {@link BATCH_WINDOW_MS} after the first one (or until
+ *   {@link MAX_BATCH_IDS} are waiting) and asked in one POST; an id queued
+ *   or in flight is joined, never asked twice, until it is too old to
+ *   trust;
+ * - a pause: after a 429 or 5xx no POST goes to that worker for its
+ *   `Retry-After` ({@link PAUSE_DEFAULT_MS} without one, at most
+ *   {@link PAUSE_MAX_MS}).
+ *
+ * A POST belongs to no call: it runs under its own {@link POST_TIMEOUT_MS}
+ * through the plugin's bounded exchange (no redirects, body cut off at
+ * {@link MAX_RESPONSE_BYTES}) and asks the worker to answer by
+ * {@link RESPONSE_BUDGET_MS} with what is ready. Each call waits for its own ids only, until
+ * its own deadline or its turn's abort; its answer then lands in the memo
+ * for the next call. Names never fail a tool: any failure is one log line
+ * (a status or an error class, never an identifier) and no names.
+ *
+ * A batcher is shared by every Durable Object of the isolate, while timers
+ * and fetches belong to the actor that started them: if it dies they end
+ * unsettled and nothing that follows them runs. A promise it made wakes no
+ * one either, even when a live actor settles it. So no shared state waits
+ * on one request finishing: age recovers it, noticed by the next caller.
+ * That caller asks an open batch past {@link BATCH_STALE_MS} afresh and
+ * sends it, and an id outstanding past {@link OUTSTANDING_MAX_MS} is
+ * dropped, to be asked afresh by whoever needs it.
+ */
+import { z } from 'zod';
+import { LruMap } from '../../../core/domain-context/lru';
+import {
+  READ_GRACE_MS,
+  exchange,
+  untilDeadline,
+  type ChainCall,
+} from '../http';
+import type { NameInfo, NameMap } from './types';
+
+/** The worker per IXO network; `IXO_NAMES_URL` overrides. */
+export const NAMES_URLS = {
+  mainnet: 'https://names.ixo.earth',
+  testnet: 'https://testnet.names.ixo.earth',
+  devnet: 'https://devnet.names.ixo.earth',
+} as const;
+
+/** The most ids the worker takes in one request. */
+export const MAX_BATCH_IDS = 40;
+/** How long a batch gathers ids after its first one before it is sent. */
+export const BATCH_WINDOW_MS = 10;
+/** An open batch this old lost its window timer: the next caller sends it. */
+export const BATCH_STALE_MS = 100;
+/** One POST's whole time, the body read included; no caller shortens it. */
+export const POST_TIMEOUT_MS = 6_000;
+/**
+ * The worker answers by this time with what is ready, the rest as `partial`
+ * ids, and finishes them in the background for the next call. Below
+ * {@link POST_TIMEOUT_MS} by enough for network time and the worker's D1
+ * reads, which come on top: an answer past the timeout would lose every id.
+ */
+export const RESPONSE_BUDGET_MS = 5_000;
+/**
+ * An id outstanding this long lost its POST: it is never joined again and
+ * no longer counts toward {@link MAX_OUTSTANDING_IDS}.
+ */
+export const OUTSTANDING_MAX_MS = POST_TIMEOUT_MS + READ_GRACE_MS;
+/**
+ * Ids queued or in flight per batcher; a miss past this is not asked (its
+ * call goes without that name). Five full POSTs, as many as the batcher's
+ * permits let run at once.
+ */
+export const MAX_OUTSTANDING_IDS = 200;
+/** The largest answer body read. */
+export const MAX_RESPONSE_BYTES = 256 * 1024;
+export const MEMO_ENTRIES = 5_000;
+export const MEMO_MAX_MS = 10 * 60_000;
+export const MEMO_MIN_MS = 30_000;
+export const PAUSE_DEFAULT_MS = 60_000;
+export const PAUSE_MAX_MS = 300_000;
+/** The worker cleans names to at most this many characters. */
+const NAME_MAX_CHARS = 120;
+
+/** The client's one log line: a status or an error class, never an identifier. */
+function logFailure(reason: string): void {
+  // eslint-disable-next-line no-console -- console is the logger on Workers
+  console.warn(`[ixo-chain] names: ${reason}`);
+}
+
+type Network = ChainCall['network'];
+
+/** The worker's answer for one id; extra fields are ignored. */
+const resultSchema = z.object({
+  id: z.string(),
+  kind: z.string(),
+  status: z.string(),
+  complete: z.boolean(),
+  name: z.string().min(1).max(NAME_MAX_CHARS).nullable(),
+  source: z.string().nullable(),
+  selfChosen: z.boolean(),
+  cache: z.object({ freshUntil: z.number().nullable() }).nullish(),
+});
+const responseSchema = z.object({
+  network: z.string(),
+  results: z.record(z.string(), resultSchema),
+});
+type WorkerResult = z.infer<typeof resultSchema>;
+
+/** One id's answer: its name, or null for none (unnamed, unanswered or failed). */
+type Answer = NameInfo | null;
+
+/** Where a batch goes: the worker, the network it must answer for, and who asks. */
+interface Target {
+  network: Network;
+  baseUrl: string;
+  clientId: string;
+}
+
+/* ── memo ─────────────────────────────────────────────────────────────── */
+
+const newMemo = () =>
+  new LruMap<string, { answer: Answer; until: number }>({
+    maxEntries: MEMO_ENTRIES,
+  });
+let memo = newMemo();
+
+const memoKey = (baseUrl: string, id: string) => `${baseUrl}|${id}`;
+
+/** The memoised answer for `id`; undefined when none is kept or it expired. */
+function memoised(baseUrl: string, id: string): { answer: Answer } | undefined {
+  const key = memoKey(baseUrl, id);
+  const kept = memo.get(key);
+  if (!kept) return undefined;
+  if (kept.until > Date.now()) return kept;
+  memo.delete(key);
+  return undefined;
+}
+
+/**
+ * The answer a result gives: a name only when the worker resolved the id
+ * completely (a partial answer is the worker running out of time).
+ */
+function answerOf(result: WorkerResult): Answer {
+  if (
+    result.status === 'partial' ||
+    !result.complete ||
+    result.kind === 'unknown' ||
+    result.name === null ||
+    result.source === null
+  )
+    return null;
+  return {
+    name: result.name,
+    source: result.source,
+    selfChosen: result.selfChosen,
+  };
+}
+
+/** Keeps a complete answer until `freshUntil`, clamped to {@link MEMO_MIN_MS}–{@link MEMO_MAX_MS}. */
+function remember(baseUrl: string, result: WorkerResult, answer: Answer): void {
+  if (result.status === 'partial' || !result.complete) return;
+  const now = Date.now();
+  const fresh = result.cache?.freshUntil;
+  const until =
+    typeof fresh === 'number'
+      ? Math.min(Math.max(fresh * 1000, now + MEMO_MIN_MS), now + MEMO_MAX_MS)
+      : now + MEMO_MAX_MS;
+  memo.set(memoKey(baseUrl, result.id), { answer, until });
+}
+
+/* ── pause ────────────────────────────────────────────────────────────── */
+
+/** Per worker base URL: the epoch ms until which no POST is sent to it. */
+const pausedUntil = new Map<string, number>();
+
+function paused(baseUrl: string): boolean {
+  const until = pausedUntil.get(baseUrl);
+  if (until === undefined) return false;
+  if (until > Date.now()) return true;
+  pausedUntil.delete(baseUrl);
+  return false;
+}
+
+function pause(baseUrl: string, retryAfterMs: number | undefined): number {
+  const ms = Math.min(retryAfterMs ?? PAUSE_DEFAULT_MS, PAUSE_MAX_MS);
+  pausedUntil.set(baseUrl, Date.now() + ms);
+  return ms;
+}
+
+/* ── one POST ─────────────────────────────────────────────────────────── */
+
+/** A POST's own signal: no caller's abort reaches it. */
+const NEVER_ABORTED = new AbortController().signal;
+
+/**
+ * Asks the worker for `ids` (at most {@link MAX_BATCH_IDS}); the answers it
+ * gave, memoised when complete. Never throws: a failure is logged once
+ * and answers nothing; a 429 or 5xx also pauses the worker.
+ */
+async function post(
+  target: Target,
+  ids: readonly string[],
+  fetcher: typeof fetch,
+): Promise<Map<string, Answer>> {
+  const answers = new Map<string, Answer>();
+  // The POST's own "call": its permits are the batcher's, not a user's.
+  const call: ChainCall = {
+    network: target.network,
+    restUrl: '',
+    blocksyncUrl: '',
+    user: `names|${target.baseUrl}|${target.clientId}`,
+    deadline: Date.now() + POST_TIMEOUT_MS,
+    signal: NEVER_ABORTED,
+    fetch: fetcher,
+  };
+  try {
+    const attempt = await exchange(
+      call,
+      'names',
+      `${target.baseUrl}/v1/resolve`,
+      {
+        method: 'POST',
+        headers: {
+          'content-type': 'application/json',
+          'x-client-id': target.clientId,
+        },
+        body: JSON.stringify({
+          ids,
+          options: { include: ['name', 'cache'], budgetMs: RESPONSE_BUDGET_MS },
+        }),
+      },
+      { timeoutMs: POST_TIMEOUT_MS, maxBytes: MAX_RESPONSE_BYTES },
+    );
+    if (!attempt.ok) {
+      const { status } = attempt;
+      if (status !== undefined && (status === 429 || status >= 500)) {
+        const ms = pause(target.baseUrl, attempt.retryAfterMs);
+        logFailure(`${attempt.note}; paused for ${Math.ceil(ms / 1000)} s`);
+      } else logFailure(attempt.note);
+      return answers;
+    }
+    let body: unknown;
+    try {
+      body = JSON.parse(attempt.value);
+    } catch {
+      logFailure('the answer is not JSON');
+      return answers;
+    }
+    const parsed = responseSchema.safeParse(body);
+    if (!parsed.success) {
+      logFailure('the answer has an unexpected shape');
+      return answers;
+    }
+    if (parsed.data.network !== target.network) {
+      logFailure('the worker answered for another network');
+      return answers;
+    }
+    const asked = new Set(ids);
+    for (const result of Object.values(parsed.data.results)) {
+      if (!asked.has(result.id)) continue;
+      const answer = answerOf(result);
+      remember(target.baseUrl, result, answer);
+      answers.set(result.id, answer);
+    }
+  } catch (error) {
+    // Nothing above throws by design; a bug here still names nothing.
+    logFailure(
+      `the request threw (${error instanceof Error ? error.name : 'non-error value'})`,
+    );
+  }
+  return answers;
+}
+
+/* ── batcher ──────────────────────────────────────────────────────────── */
+
+/** An id queued or in flight: the promise callers join, and since when it is outstanding. */
+interface Outstanding {
+  answer: Promise<Answer>;
+  settle: (answer: Answer | Promise<Answer>) => void;
+  since: number;
+}
+
+/** The batch gathering ids: its ids, the fetch it is sent with, its start and timer. */
+interface OpenBatch {
+  ids: Map<string, Outstanding>;
+  fetch: typeof fetch;
+  since: number;
+  timer: ReturnType<typeof setTimeout>;
+}
+
+/**
+ * Gathers one worker's misses into POSTs of at most {@link MAX_BATCH_IDS}.
+ * Each id queued or in flight has one promise every caller joins; it
+ * settles with the id's answer (null when the POST failed or skipped it)
+ * and never rejects, unless the actor behind it died (see the file header).
+ */
+class Batcher {
+  private open: OpenBatch | null = null;
+  /** In the order the ids were asked, so the oldest come first. */
+  private readonly outstanding = new Map<string, Outstanding>();
+
+  constructor(private readonly target: Target) {}
+
+  /**
+   * The promise of `id`'s answer: joined when it is already queued or in
+   * flight, else queued in the open batch (opened with `fetcher`); undefined
+   * when {@link MAX_OUTSTANDING_IDS} are outstanding.
+   */
+  ask(id: string, fetcher: typeof fetch): Promise<Answer> | undefined {
+    this.recover(fetcher);
+    const joined = this.outstanding.get(id);
+    if (joined) return joined.answer;
+    if (this.outstanding.size >= MAX_OUTSTANDING_IDS) return undefined;
+    return this.queue(id, fetcher);
+  }
+
+  /** Puts `id` in the open batch, opened with `fetcher` if none is. */
+  private queue(id: string, fetcher: typeof fetch): Promise<Answer> {
+    let settle: Outstanding['settle'] = () => undefined;
+    const answer = new Promise<Answer>((resolve) => {
+      settle = resolve;
+    });
+    const entry = { answer, settle, since: Date.now() };
+    this.outstanding.set(id, entry);
+    this.open ??= {
+      ids: new Map(),
+      fetch: fetcher,
+      since: entry.since,
+      timer: setTimeout(() => this.flush(), BATCH_WINDOW_MS),
+    };
+    this.open.ids.set(id, entry);
+    if (this.open.ids.size >= MAX_BATCH_IDS) this.flush();
+    return answer;
+  }
+
+  /**
+   * Drops the ids too old to trust, then sends an open batch whose timer is
+   * presumed dead, in this caller's name: with its `fetcher` (the batch's
+   * own may belong to the dead actor) and under new entries (the old
+   * promises may too). The old entries follow the new ones, so a live call
+   * that joined the batch still gets its answer.
+   */
+  private recover(fetcher: typeof fetch): void {
+    const now = Date.now();
+    for (const [id, entry] of this.outstanding) {
+      if (now - entry.since <= OUTSTANDING_MAX_MS) break;
+      this.outstanding.delete(id);
+    }
+    const stale = this.open;
+    if (!stale || now - stale.since <= BATCH_STALE_MS) return;
+    this.open = null;
+    clearTimeout(stale.timer);
+    for (const [id, entry] of stale.ids) {
+      if (this.outstanding.get(id) !== entry) continue;
+      this.outstanding.delete(id);
+      entry.settle(this.queue(id, fetcher));
+    }
+    this.flush();
+  }
+
+  /** Sends the open batch now (its timer cleared). */
+  private flush(): void {
+    const batch = this.open;
+    if (!batch) return;
+    this.open = null;
+    clearTimeout(batch.timer);
+    void this.send(batch);
+  }
+
+  private async send(batch: OpenBatch): Promise<void> {
+    // A pause that began while the batch gathered: nothing is asked.
+    const answers = paused(this.target.baseUrl)
+      ? new Map<string, Answer>()
+      : await post(this.target, [...batch.ids.keys()], batch.fetch);
+    for (const [id, entry] of batch.ids) {
+      entry.settle(answers.get(id) ?? null);
+      // An id asked afresh meanwhile keeps its new entry.
+      if (this.outstanding.get(id) === entry) this.outstanding.delete(id);
+    }
+  }
+
+  /** Tests only: ids queued or in flight, and whether a window timer runs. */
+  state(): { outstanding: number; timer: boolean } {
+    return { outstanding: this.outstanding.size, timer: this.open !== null };
+  }
+
+  /** Tests only: drops the open batch and its timer (its callers' waits end at their deadlines). */
+  drop(): void {
+    if (this.open) clearTimeout(this.open.timer);
+    this.open = null;
+    this.outstanding.clear();
+  }
+}
+
+/** The isolate's batchers, one per worker base URL and client id. */
+const batchers = new Map<string, Batcher>();
+
+function batcherFor(target: Target): Batcher {
+  const key = `${target.baseUrl}|${target.clientId}`;
+  let batcher = batchers.get(key);
+  if (!batcher) {
+    batcher = new Batcher(target);
+    batchers.set(key, batcher);
+  }
+  return batcher;
+}
+
+/** Tests only: the batcher of `baseUrl` and `clientId`, if one was made. */
+export function batcherState(
+  baseUrl: string,
+  clientId: string,
+): { outstanding: number; timer: boolean } | undefined {
+  return batchers.get(`${baseUrl}|${clientId}`)?.state();
+}
+
+/** Tests only: forget the memo, the pauses and every batch. */
+export function resetNamesClient(): void {
+  memo = newMemo();
+  pausedUntil.clear();
+  for (const batcher of batchers.values()) batcher.drop();
+  batchers.clear();
+}
+
+/* ── one call ─────────────────────────────────────────────────────────── */
+
+export interface NamesRequest extends Target {
+  fetch: typeof fetch;
+  /** Epoch ms after which this call stops waiting (the POST runs on). */
+  deadline: number;
+  /** The turn's abort: thrown as soon as it fires. */
+  signal: AbortSignal;
+}
+
+/**
+ * The names of `ids` known by `request.deadline`, in the order of `ids`
+ * (best placed first); identifiers without a name are absent. Memoised
+ * answers come straight from the memo; the rest join or open a batch,
+ * unless the worker is paused. Only the turn's abort is thrown.
+ */
+export async function resolveNames(
+  ids: readonly string[],
+  request: NamesRequest,
+): Promise<NameMap> {
+  const answers = new Map<string, Answer>();
+  const waits: Promise<void>[] = [];
+  const batcher = paused(request.baseUrl) ? null : batcherFor(request);
+  for (const id of new Set(ids)) {
+    const kept = memoised(request.baseUrl, id);
+    if (kept) {
+      answers.set(id, kept.answer);
+      continue;
+    }
+    const asked = batcher?.ask(id, request.fetch);
+    if (asked) waits.push(asked.then((answer) => void answers.set(id, answer)));
+  }
+  if (waits.length) await untilDeadline(Promise.all(waits), request);
+  const names: NameMap = {};
+  for (const id of ids) {
+    const answer = answers.get(id);
+    if (answer) names[id] = { ...answer };
+  }
+  return names;
+}