@ixo/oracle-runtime-workers 0.29.0 → 0.30.0

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

@@ -0,0 +1,470 @@
+/**
+ * The `chain_lookup` executor. Every request is planned first; one its kind
+ * refuses answers `invalid_input` on its own and is never read. The
+ * Blocksync plans of a call become ONE GraphQL document (aliases `r0…r7`
+ * by request position); the chain plans run their REST reads and denom
+ * lookups at once. Every read takes one of the user's permits
+ * (`USER_READ_PERMITS` in `http.ts`, shared by all of the user's calls), so
+ * a call never has more in flight; all of it runs inside the call's 20 s
+ * budget and the turn's abort. Each request answers on its own, in request
+ * order; the answer is cut to {@link MAX_RESULT_CHARS}.
+ */
+import type { z } from 'zod';
+import type { UserIdentity } from '../../core/user-identity';
+import {
+  blocksyncLookup,
+  STATEMENT_TIMEOUT,
+  TRANSIENT_ERROR,
+  type AliasedField,
+  type BlocksyncAnswer,
+  type GraphqlError,
+} from './blocksync';
+import {
+  denomDisplay,
+  knownDisplay,
+  unreadDisplay,
+  type IbcDenomTable,
+  type TokenDisplay,
+} from './denoms';
+import {
+  throwIfAborted,
+  tooLarge,
+  unavailable,
+  type ChainCall,
+  type Failure,
+  type Outcome,
+  type Source,
+} from './http';
+import type {
+  KindContext,
+  KindDef,
+  KindPlan,
+  KindRestOptions,
+  KindResult,
+} from './kinds/types';
+import { firstOf, type Subject } from './queries';
+import { restJson, type Absent } from './rest';
+import { isRecord } from './summaries';
+
+/** Largest answer handed to the model, in JSON characters. */
+export const MAX_RESULT_CHARS = 12_000;
+/**
+ * Distinct denoms one request looks up to show its bank coins; any further
+ * denom is shown unread (base units, `metadataUnavailable`).
+ */
+export const MAX_DISPLAY_DENOMS = 20;
+
+/** One request: its kind and the arguments the kind validates. */
+export interface KindRequest {
+  kind: KindDef;
+  args?: unknown;
+  /**
+   * The rows a list kind reads (1–20), used as given instead of the
+   * request's `args.first` scaled for the call.
+   */
+  first?: number;
+}
+
+/** What a call's requests are read with besides the call itself. */
+export interface LookupOptions {
+  /**
+   * The IBC table bank coins are shown with: the network's built-in
+   * entries with the operator's `IXO_CHAIN_IBC_DENOMS` merged over them.
+   */
+  ibcDenoms: IbcDenomTable;
+}
+
+export type LookupResult =
+  | {
+      kind: string;
+      source: Source;
+      ok: true;
+      subject?: Subject;
+      first?: number;
+      data: unknown;
+      truncated?: boolean;
+    }
+  | ({ kind: string; source: Source; ok: false } & Failure);
+
+const LEFT_OUT = tooLarge(
+  'Left out to keep the answer small; ask for it on its own.',
+);
+const NO_ANSWER = unavailable('blocksync', 'no answer for this request', true);
+const NO_BLOCKSYNC = unavailable(
+  'blocksync',
+  'no valid BLOCKSYNC_GRAPHQL_URL is configured',
+  false,
+);
+const KIND_FAILED: Failure = {
+  reason: 'unavailable',
+  detail: 'This request failed inside the oracle; asking again will not help.',
+  retryable: false,
+};
+
+/* ── contexts ─────────────────────────────────────────────────────────── */
+
+/**
+ * The kind contexts of one call. A REST read is `restJson` over the call
+ * (each attempt under one of the user's permits); a denom that needs a
+ * lookup is looked up once per call (and through `denoms.ts`'s per-isolate
+ * cache), and each request looks up at most {@link MAX_DISPLAY_DENOMS}
+ * distinct denoms. Returns each request's context for the rows it reads.
+ */
+export function callContexts(
+  call: ChainCall,
+  identity: UserIdentity,
+  options: LookupOptions,
+): (first: number) => KindContext {
+  function rest<S extends z.ZodType>(
+    path: string,
+    schema: S,
+    opts: KindRestOptions & { noneOn?: undefined },
+  ): Promise<Outcome<z.infer<S>>>;
+  function rest<S extends z.ZodType>(
+    path: string,
+    schema: S,
+    opts: KindRestOptions & { noneOn: readonly number[] },
+  ): Promise<Outcome<z.infer<S>> | Absent>;
+  function rest<S extends z.ZodType>(
+    path: string,
+    schema: S,
+    opts: KindRestOptions,
+  ): Promise<Outcome<z.infer<S>> | Absent> {
+    return restJson(call, path, schema, opts);
+  }
+  const lookups = new Map<string, Promise<TokenDisplay>>();
+  const lookUp = (denom: string): Promise<TokenDisplay> => {
+    const running = lookups.get(denom);
+    if (running) return running;
+    const pending = denomDisplay(call, denom, options.ibcDenoms);
+    lookups.set(denom, pending);
+    return pending;
+  };
+  return (first) => {
+    const asked = new Set<string>();
+    return {
+      call,
+      identity,
+      first,
+      rest,
+      display(denom) {
+        const known = knownDisplay(denom, options.ibcDenoms);
+        if (known) return Promise.resolve(known);
+        if (!asked.has(denom)) {
+          if (asked.size >= MAX_DISPLAY_DENOMS)
+            return Promise.resolve(unreadDisplay(denom));
+          asked.add(denom);
+        }
+        return lookUp(denom);
+      },
+    };
+  };
+}
+
+/** The request's own `first` when it is a whole number of at least 1 (its kind's schema checks the rest). */
+function requestedFirst(args: unknown): number | undefined {
+  const value = isRecord(args) ? args.first : undefined;
+  return typeof value === 'number' && Number.isInteger(value) && value >= 1
+    ? value
+    : undefined;
+}
+
+/* ── one request ──────────────────────────────────────────────────────── */
+
+/**
+ * A kind that throws for anything but the turn's abort has a bug: its
+ * request fails on its own (logged with the kind and the error's class
+ * only: a message may carry request data) and the others still answer.
+ */
+function kindFailed(call: ChainCall, kind: string, error: unknown): KindResult {
+  throwIfAborted(call.signal);
+  // eslint-disable-next-line no-console -- console IS the logger on Workers
+  console.error(
+    `[ixo-chain] kind ${kind} threw: ${error instanceof Error ? error.name : 'non-error value'}`,
+  );
+  return { ok: false, ...KIND_FAILED };
+}
+
+type BlocksyncPlan = Extract<KindPlan, { source: 'blocksync' }>;
+type ChainPlan = Extract<KindPlan, { source: 'chain' }>;
+
+/** A request as its plan (and, for Blocksync, its document field), or its refusal. */
+type Planned =
+  | { plan: ChainPlan }
+  | { plan: BlocksyncPlan; field: AliasedField }
+  | { refused: KindResult };
+
+function planOf(
+  call: ChainCall,
+  request: KindRequest,
+  ctx: KindContext,
+  alias: string,
+): Planned {
+  try {
+    const plan = request.kind.plan(request.args, ctx);
+    if ('invalid' in plan)
+      return {
+        refused: {
+          ok: false,
+          reason: 'invalid_input',
+          detail: plan.invalid,
+          retryable: false,
+        },
+      };
+    if (plan.source === 'chain') return { plan };
+    const field = plan.field(alias);
+    // A field under another alias would break every request of the document.
+    if (!field.startsWith(`${alias}: `))
+      throw new Error(`its field is not aliased ${alias}`);
+    return { plan, field: { alias, field } };
+  } catch (error) {
+    return { refused: kindFailed(call, request.kind.name, error) };
+  }
+}
+
+async function settle(
+  call: ChainCall,
+  kind: string,
+  answer: () => KindResult | Promise<KindResult>,
+): Promise<KindResult> {
+  try {
+    return await answer();
+  } catch (error) {
+    return kindFailed(call, kind, error);
+  }
+}
+
+/** The failure the document's errors give one alias, if any. */
+function aliasFailure(alias: string, errors: GraphqlError[]): Failure | null {
+  const mine = errors.filter(
+    (error) => error.path.length === 0 || error.path[0] === alias,
+  );
+  if (!mine.length) return null;
+  if (mine.some((error) => STATEMENT_TIMEOUT.test(error.message)))
+    return {
+      reason: 'index_timeout',
+      detail:
+        'The index could not answer in time; ask for fewer rows or a narrower filter.',
+      retryable: false,
+    };
+  if (mine.some((error) => TRANSIENT_ERROR.test(error.message)))
+    return unavailable('blocksync', 'a transient database error', true);
+  const message = mine[0]?.message ?? '';
+  return {
+    reason: 'unavailable',
+    detail: `Blocksync could not answer this request: ${message.length > 160 ? `${message.slice(0, 159)}…` : message}`,
+    retryable: false,
+  };
+}
+
+/** One alias's answer from the document, or why it has none. */
+function aliasAnswer(
+  read: Outcome<BlocksyncAnswer>,
+  alias: string,
+  plan: BlocksyncPlan,
+): KindResult | Promise<KindResult> {
+  if (!read.ok) return { ok: false, ...read.failure };
+  const failure = aliasFailure(alias, read.value.errors);
+  if (failure) return { ok: false, ...failure };
+  // An alias the answer does not carry was not answered, which says
+  // nothing about whether the record exists.
+  const { data } = read.value;
+  if (!data || !(alias in data)) return { ok: false, ...NO_ANSWER };
+  return plan.shape(data[alias]);
+}
+
+function resultOf(
+  kind: string,
+  source: Source,
+  result: KindResult,
+): LookupResult {
+  if (!result.ok)
+    return {
+      kind,
+      source,
+      ok: false,
+      reason: result.reason,
+      detail: result.detail,
+      retryable: result.retryable,
+    };
+  return {
+    kind,
+    source,
+    ok: true,
+    ...(result.subject ? { subject: result.subject } : {}),
+    ...(result.first !== undefined ? { first: result.first } : {}),
+    data: result.data,
+  };
+}
+
+/** The alias of the request at `index`. */
+const aliasOf = (index: number) => `r${index}`;
+
+/**
+ * Every request's answer, in request order. Only the turn's abort is
+ * thrown.
+ */
+export async function runRequests(
+  call: ChainCall,
+  identity: UserIdentity,
+  requests: readonly KindRequest[],
+  options: LookupOptions,
+): Promise<LookupResult[]> {
+  const contextFor = callContexts(call, identity, options);
+  const planned = requests.map((request, index) => ({
+    kind: request.kind,
+    planned: planOf(
+      call,
+      request,
+      contextFor(
+        request.first ?? firstOf(requestedFirst(request.args), requests.length),
+      ),
+      aliasOf(index),
+    ),
+  }));
+  const fields = planned.flatMap(({ planned: entry }) =>
+    'field' in entry ? [entry.field] : [],
+  );
+  const document: Promise<Outcome<BlocksyncAnswer>> = !fields.length
+    ? Promise.resolve({ ok: true, value: { data: null, errors: [] } })
+    : call.blocksyncUrl
+      ? blocksyncLookup(call, fields)
+      : Promise.resolve({ ok: false, failure: NO_BLOCKSYNC });
+  return Promise.all(
+    planned.map(async ({ kind, planned: entry }) => {
+      if ('refused' in entry)
+        return resultOf(kind.name, kind.source, entry.refused);
+      if (!('field' in entry)) {
+        const { plan } = entry;
+        return resultOf(
+          kind.name,
+          'chain',
+          await settle(call, kind.name, () => plan.run()),
+        );
+      }
+      const { plan, field } = entry;
+      const read = await document;
+      return resultOf(
+        kind.name,
+        'blocksync',
+        await settle(call, kind.name, () =>
+          aliasAnswer(read, field.alias, plan),
+        ),
+      );
+    }),
+  );
+}
+
+/* ── size ─────────────────────────────────────────────────────────────── */
+
+const size = (value: unknown) => JSON.stringify(value).length;
+
+/** The largest array (by JSON size) with more than one element inside `value`. */
+function largestArray(
+  value: unknown,
+): { list: unknown[]; size: number } | null {
+  let best: { list: unknown[]; size: number } | null = null;
+  const visit = (node: unknown) => {
+    if (Array.isArray(node)) {
+      if (node.length > 1) {
+        const bytes = size(node);
+        if (!best || bytes > best.size) best = { list: node, size: bytes };
+      }
+      node.forEach(visit);
+    } else if (isRecord(node)) Object.values(node).forEach(visit);
+  };
+  visit(value);
+  return best;
+}
+
+/**
+ * Halves the largest list inside `within` (keeping its first rows) until
+ * `holder` serialises within `budget` or no list is left to halve; marks
+ * `holder` `truncated` when anything was cut.
+ */
+export function shrink(
+  holder: { truncated?: boolean },
+  within: unknown,
+  budget: number,
+): void {
+  while (size(holder) > budget) {
+    const found = largestArray(within);
+    if (!found) return;
+    found.list.length = Math.ceil(found.list.length / 2);
+    holder.truncated = true;
+  }
+}
+
+/**
+ * Fits `parts` (the results inside `whole`) into what
+ * {@link MAX_RESULT_CHARS} leaves after the envelope: each part gets an
+ * equal share, smallest first, and a share a part does not use passes to
+ * the larger parts after it. A cut re-measures only its own part.
+ */
+function shareOut(
+  whole: unknown,
+  parts: { data: unknown; truncated?: boolean }[],
+): void {
+  const sizes = new Map(parts.map((part) => [part, size(part)]));
+  let left =
+    MAX_RESULT_CHARS -
+    (size(whole) - [...sizes.values()].reduce((sum, n) => sum + n, 0));
+  const bySize = [...parts].sort(
+    (a, b) => (sizes.get(a) ?? 0) - (sizes.get(b) ?? 0),
+  );
+  bySize.forEach((part, index) => {
+    shrink(part, part.data, Math.floor(left / (bySize.length - index)));
+    left -= size(part);
+  });
+}
+
+/**
+ * Cuts `results` (inside `whole`) until `whole` fits
+ * {@link MAX_RESULT_CHARS}: lists are halved within each result's share;
+ * a result that cannot be cut to its share leaves the largest answers out,
+ * one at a time, as `too_large`.
+ */
+export function fitResults(whole: unknown, results: LookupResult[]): void {
+  shareOut(
+    whole,
+    results.flatMap((result) => (result.ok ? [result] : [])),
+  );
+  while (size(whole) > MAX_RESULT_CHARS) {
+    let largest = -1;
+    results.forEach((result, index) => {
+      if (result.ok && (largest < 0 || size(result) > size(results[largest])))
+        largest = index;
+    });
+    const evicted = results[largest];
+    if (!evicted) break;
+    results[largest] = {
+      kind: evicted.kind,
+      source: evicted.source,
+      ok: false,
+      ...LEFT_OUT,
+    };
+  }
+}
+
+/** `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[];
+}> {
+  const results = await runRequests(call, identity, requests, options);
+  const whole = {
+    ok: true as const,
+    network: call.network,
+    fetchedAt: new Date().toISOString(),
+    results,
+  };
+  fitResults(whole, results);
+  return whole;
+}