AztecProtocol
diff --git a/‎yarn-project/pxe/src/tagging/constants.ts‎
Lines changed: 10 additions & 0 deletions b/‎yarn-project/pxe/src/tagging/constants.ts‎
Lines changed: 10 additions & 0 deletions
diff --git a/‎yarn-project/pxe/src/tagging/index.ts‎
Lines changed: 1 addition & 1 deletion b/‎yarn-project/pxe/src/tagging/index.ts‎
Lines changed: 1 addition & 1 deletion
diff --git a/‎yarn-project/pxe/src/tagging/recipient_sync/sync_tagged_private_logs.bench.test.ts‎
Lines changed: 267 additions & 0 deletions b/‎yarn-project/pxe/src/tagging/recipient_sync/sync_tagged_private_logs.bench.test.ts‎
Lines changed: 267 additions & 0 deletions
@@ -18,3 +18,13 @@ import { MAX_PRIVATE_LOGS_PER_TX } from '@aztec/constants';
 // reserved at log emission time, before squashing is decided, and the kernel's reset/squash loop is not bounded by
 // MAX_PRIVATE_LOGS_PER_TX. No fixed window value closes that gap.
 export const UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN = MAX_PRIVATE_LOGS_PER_TX + 20;
+
+// The number of tags probed per constrained secret in the first round.
+//
+// The probe doubles each round (2, 4, 8, ..., capped at UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN) while every probed
+// index is a hit, stopping at the first missing tag. Constrained delivery is gapless, so a single missing tag proves
+// the stream has ended: at steady state this turns a full WINDOW_LEN probe into two tags. A secret K logs behind
+// catches up in ~log2(K) round-trips while the probe is still doubling (2, 4, 8, 16, 32), but once it saturates the cap
+// and advances WINDOW_LEN tags per round, deeper catch-up is linear at ~K/WINDOW_LEN rounds. Either way it beats both
+// the full window every round and one round per log.
+export const INITIAL_CONSTRAINED_PROBE_LEN = 2;
@@ -12,7 +12,7 @@
 export { syncTaggedPrivateLogs } from './recipient_sync/sync_tagged_private_logs.js';
 export { syncSenderTaggingIndexes } from './sender_sync/sync_sender_tagging_indexes.js';
 export { persistSenderTaggingIndexRangesForTx } from './persist_sender_tagging_index_ranges.js';
-export { UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN } from './constants.js';
+export { UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN, INITIAL_CONSTRAINED_PROBE_LEN } from './constants.js';
 export { getAllPrivateLogsByTags, getAllPublicLogsByTagsFromContract } from './get_all_logs_by_tags.js';
 
 // Re-export tagging-related types from stdlib
 
@@ -0,0 +1,267 @@
+import { MAX_TX_LIFETIME } from '@aztec/constants';
+import { BlockNumber } from '@aztec/foundation/branded-types';
+import { sum, times } from '@aztec/foundation/collection';
+import { createLogger } from '@aztec/foundation/log';
+import { sleep } from '@aztec/foundation/sleep';
+import { openTmpStore } from '@aztec/kv-store/lmdb-v2';
+import type { AztecNode } from '@aztec/stdlib/interfaces/server';
+import { AppTaggingSecretKind, type PrivateLogsQuery, randomLogResult } from '@aztec/stdlib/logs';
+import { randomAppTaggingSecret } from '@aztec/stdlib/testing';
+import { BlockHeader } from '@aztec/stdlib/tx';
+
+import { mkdir, writeFile } from 'fs/promises';
+import { type MockProxy, mock } from 'jest-mock-extended';
+import path from 'path';
+
+import { BenchmarkedNodeFactory } from '../../contract_function_simulator/benchmarked_node.js';
+import { RecipientTaggingStore } from '../../storage/tagging_store/recipient_tagging_store.js';
+import { UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN, syncTaggedPrivateLogs } from '../index.js';
+import { computeSiloedTagForIndex, extractTags } from '../testing/tag_query_test_utils.js';
+
+/**
+ * Benchmark for constrained recipient tag-sync.
+ *
+ * Measures the per-sync cost of `syncTaggedPrivateLogs` for constrained secrets. Constrained streams are gapless, so
+ * the scan uses a doubling first-miss probe (see `INITIAL_CONSTRAINED_PROBE_LEN` in `../constants.ts` for the probe
+ * schedule and complexity analysis), stopping at the first missing tag instead of fetching the full window. This only
+ * reports costs; the scan behavior itself (probe schedules, round counts) is pinned by the unit tests in
+ * `sync_tagged_private_logs.test.ts`, which run in CI while this bench is opt-in.
+ *
+ * Manual run:
+ * ```bash
+ * RUN_TAG_SYNC_BENCH=1 JEST_MAX_WORKERS=1 BENCH_OUTPUT=/tmp/tag-sync-bench-current.json \
+ *   yarn workspace @aztec/pxe test src/tagging/recipient_sync/sync_tagged_private_logs.bench.test.ts
+ * ```
+ *
+ * Metrics, per scenario:
+ * - `tag-queries`: total tags queried, the throughput win.
+ * - `rpc-round-trips`: sequential blocking waits on the node, via `BenchmarkedNodeFactory`. The latency axis: this
+ *    grows with K per the complexity analysis on `INITIAL_CONSTRAINED_PROBE_LEN`, and depends only on K, not on secret
+ *    count. A round's tags are chunked at MAX_RPC_LEN (=100) into parallel calls internally, but those overlap, so a
+ *    wide round is still one round-trip; that is why round-trips, not raw call count, is the latency axis.
+ * - `rpc-blocking-time`: measured wall-clock the caller blocks on the node, under a modeled
+ *   `MODELED_NODE_RPC_LATENCY_MS` per call plus a little per-round overhead. Parallel calls within a round overlap, so
+ *   it tracks round-trips (a 1000-secret round is many parallel chunks but ~one round-trip of blocking time).
+ *
+ * Scenario labels: `steady-state` is no new logs (K = 0); `catch-up-K` is K new contiguous logs per secret since the
+ * last sync; `secrets=N` is N secrets synced together in one batched pass. Because round-trips depend only on K (not
+ * N), the light catch-up scenarios run at both 100 and 1000 secrets to show tag-queries scale with N while round-trips
+ * do not. The `mixed` row is the realistic active sync (999 idle secrets + 1 deep straggler at K = 100): it isolates
+ * that tag-queries stay dominated by the idle majority while a single straggler alone sets the round-trip count.
+ */
+
+const logger = createLogger('pxe:tagging:bench');
+
+const FINALIZED_BLOCK_NUMBER = BlockNumber(10);
+const ANCHOR_BLOCK_NUMBER = BlockNumber(100);
+const CURRENT_TIMESTAMP = BigInt(Math.floor(Date.now() / 1000));
+const ANCHOR_BLOCK_HEADER = BlockHeader.random({ blockNumber: ANCHOR_BLOCK_NUMBER, timestamp: CURRENT_TIMESTAMP });
+const AGED_TIMESTAMP = CURRENT_TIMESTAMP - BigInt(MAX_TX_LIFETIME) - 1000n;
+const JOB_ID = 'bench-job';
+
+// Every scenario starts warm: index 0 is already persisted, so the scan resumes at index 1 rather than cold-starting.
+const PRIOR_FINALIZED_INDEX = 0;
+
+// Models per-call node RPC latency so round-trip blocking time is meaningful against an otherwise-instant mock node.
+// The round-trip *count* is independent of this value; only `rpc-blocking-time` scales with it.
+const MODELED_NODE_RPC_LATENCY_MS = 5;
+
+const describeBench = process.env.RUN_TAG_SYNC_BENCH ? describe : describe.skip;
+
+/** One benchmark measurement in the benchmark JSON shape. */
+type BenchResult = { name: string; value: number; unit: string };
+
+type Scenario = {
+  label: string;
+  kind: AppTaggingSecretKind;
+  /** Number of secrets the recipient holds for this directional app. */
+  secretCount: number;
+  /** New contiguous finalized logs available per secret since the last sync (0 = steady state). */
+  newLogs: number;
+  /**
+   * Optional heterogeneous load: the first `count` secrets get `newLogs` new logs each, the remaining
+   * `secretCount - count` are idle (K = 0). When set, the scenario-level `newLogs` is ignored. Models a realistic
+   * sync where most senders are quiet and a few are deep in catch-up.
+   */
+  deepCohort?: { count: number; newLogs: number };
+};
+
+/**
+ * The per-secret new-log distribution for a scenario: `deepCohort.count` secrets at `deepCohort.newLogs`, the rest
+ * idle, or a uniform `newLogs` for every secret when no cohort is set. Single source for both log seeding and the
+ * seeding sanity check.
+ */
+function newLogsPerSecret(scenario: Scenario): number[] {
+  return Array.from({ length: scenario.secretCount }, (_, i) =>
+    scenario.deepCohort && i < scenario.deepCohort.count ? scenario.deepCohort.newLogs : scenario.newLogs,
+  );
+}
+
+const SCENARIOS: Scenario[] = [
+  // steady-state (K = 0): no new logs since the last sync, across recipient secret counts. The dominant case and where
+  // first-miss wins by roughly the window size.
+  ...[1, 10, 100, 1000].map(secretCount => ({
+    label: `constrained/steady-state/secrets=${secretCount}`,
+    kind: AppTaggingSecretKind.CONSTRAINED,
+    secretCount,
+    newLogs: 0,
+  })),
+  // Light catch-up (K new logs per secret) at 100 and 1000 secrets. Round-trips depend only on K and P, not on N, so
+  // the two secret counts share a round-trip count and differ only in tag-queries (10x) and blocking time.
+  ...[100, 1000].flatMap(secretCount =>
+    [1, 3].map(newLogs => ({
+      label: `constrained/catch-up-${newLogs}/secrets=${secretCount}`,
+      kind: AppTaggingSecretKind.CONSTRAINED,
+      secretCount,
+      newLogs,
+    })),
+  ),
+  // Deep catch-up at 100 and 1000 secrets (the negative case): round-trips grow one per probe step while tag-queries
+  // stay at the K + 1 floor. From a full window up, the WINDOW_LEN cap forces multiple rounds at any P. As with light
+  // catch-up, round-trips depend only on K and P (not N), so the two secret counts share a round-trip count and differ
+  // only in tag-queries (10x) and blocking time.
+  ...[100, 1000].flatMap(secretCount =>
+    [UNFINALIZED_TAGGING_INDEXES_WINDOW_LEN, 50, 100].map(newLogs => ({
+      label: `constrained/catch-up-${newLogs}/secrets=${secretCount}`,
+      kind: AppTaggingSecretKind.CONSTRAINED,
+      secretCount,
+      newLogs,
+    })),
+  ),
+  // Mixed (the realistic active sync): 999 idle senders + 1 deep straggler (K = 100). Tag-queries stay dominated by the
+  // idle majority, but the single straggler alone sets the round-trip count (tags are batched across all secrets per
+  // round, so the round count equals the deepest secret's), so round-trips match catch-up-100.
+  {
+    label: `constrained/mixed/secrets=1000`,
+    kind: AppTaggingSecretKind.CONSTRAINED,
+    secretCount: 1000,
+    newLogs: 0,
+    deepCohort: { count: 1, newLogs: 100 },
+  },
+  // Control: unconstrained steady state is unaffected by the optimization (windowed scan cannot first-miss).
+  {
+    label: `unconstrained/steady-state/secrets=100`,
+    kind: AppTaggingSecretKind.UNCONSTRAINED,
+    secretCount: 100,
+    newLogs: 0,
+  },
+];
+
+describeBench('syncTaggedPrivateLogs constrained-sync bench', () => {
+  const aztecNode: MockProxy<AztecNode> = mock<AztecNode>();
+
+  function makeFinalizedLog() {
+    return {
+      ...randomLogResult(/* includeEffects */ true),
+      blockNumber: FINALIZED_BLOCK_NUMBER,
+      blockTimestamp: AGED_TIMESTAMP,
+    };
+  }
+
+  async function runScenario(scenario: Scenario) {
+    const { kind, secretCount } = scenario;
+    const perSecretNewLogs = newLogsPerSecret(scenario);
+
+    aztecNode.getPrivateLogsByTags.mockReset();
+    const taggingStore = new RecipientTaggingStore(await openTmpStore('bench'));
+    const secrets = await Promise.all(Array.from({ length: secretCount }, () => randomAppTaggingSecret(kind)));
+
+    // Seed the persisted finalized indexes to simulate a recipient that already synced prior finalized messages. The
+    // per-secret writes are independent, so run them concurrently.
+    await Promise.all(
+      secrets.map(async secret => {
+        await taggingStore.updateHighestFinalizedIndex(secret, PRIOR_FINALIZED_INDEX, JOB_ID);
+        if (kind === AppTaggingSecretKind.UNCONSTRAINED) {
+          await taggingStore.updateHighestAgedIndex(secret, PRIOR_FINALIZED_INDEX, JOB_ID);
+        }
+      }),
+    );
+
+    // Tags that should resolve to a finalized log: per secret, the contiguous run
+    // (PRIOR_FINALIZED_INDEX, PRIOR_FINALIZED_INDEX + K]. Computing the tags is independent per (secret, index).
+    const hitTagList = await Promise.all(
+      secrets.flatMap((secret, s) =>
+        times(perSecretNewLogs[s], i => computeSiloedTagForIndex(secret, PRIOR_FINALIZED_INDEX + i + 1)),
+      ),
+    );
+    const hitTags = new Set(hitTagList.map(tag => tag.toString()));
+
+    aztecNode.getPrivateLogsByTags.mockImplementation(async (query: PrivateLogsQuery) => {
+      await sleep(MODELED_NODE_RPC_LATENCY_MS);
+      return extractTags(query).map(tag => (hitTags.has(tag.toString()) ? [makeFinalizedLog()] : []));
+    });
+
+    // Wrap the node so we capture round-trips and blocking time the same way the client_flows app benches do. The
+    // Proxy delegates to the underlying mock, so `mock.calls` still records every query for tag counting.
+    const benchmarkedNode = BenchmarkedNodeFactory.create(aztecNode);
+
+    const logs = await syncTaggedPrivateLogs(
+      secrets,
+      benchmarkedNode,
+      taggingStore,
+      ANCHOR_BLOCK_HEADER,
+      FINALIZED_BLOCK_NUMBER,
+      JOB_ID,
+    );
+
+    const calls = aztecNode.getPrivateLogsByTags.mock.calls;
+    const tagQueries = sum(calls.map(([query]) => extractTags(query).length));
+
+    // Round-trips and blocking time from the same instrumentation the app benches use. `syncTaggedPrivateLogs` only
+    // ever calls `getPrivateLogsByTags`, so every round-trip is that method.
+    const { roundTrips } = benchmarkedNode.getStats();
+    const rpcRoundTrips = roundTrips.roundTrips;
+    const rpcBlockingTimeMs = roundTrips.totalBlockingTime;
+
+    // First-miss floor: each secret pays its K hits + 1 miss. Unconstrained cannot first-miss, so its floor is
+    // its current cost.
+    const firstMissOptimum =
+      kind === AppTaggingSecretKind.CONSTRAINED ? sum(perSecretNewLogs.map(k => k + 1)) : tagQueries;
+
+    return {
+      ...scenario,
+      logsFound: logs.length,
+      tagQueries,
+      rpcRoundTrips,
+      rpcBlockingTimeMs,
+      firstMissOptimum,
+    };
+  }
+
+  it('reports per-sync tag-queries, round-trips, and blocking time', async () => {
+    const rows = [];
+    for (const scenario of SCENARIOS) {
+      const row = await runScenario(scenario);
+      rows.push(row);
+      logger.info(
+        `${row.label.padEnd(42)} tag-queries=${String(row.tagQueries).padStart(6)} ` +
+          `first-miss-optimum=${String(row.firstMissOptimum).padStart(6)} ` +
+          `reduction=${(row.tagQueries / row.firstMissOptimum).toFixed(1)}x ` +
+          `round-trips=${String(row.rpcRoundTrips).padStart(4)} ` +
+          `blocking=${row.rpcBlockingTimeMs.toFixed(0).padStart(4)}ms logs=${String(row.logsFound).padStart(5)}`,
+      );
+
+      // Seeding sanity check only: every seeded log must be found, or the reported numbers measure a broken harness.
+      // Scan behavior (probe schedules, round counts) is pinned by the unit tests, which run in CI while this does not.
+      expect(row.logsFound).toBe(sum(newLogsPerSecret(scenario)));
+
+      // Steady state is one round trip by construction: every secret first-misses in round one, independent of secret
+      // count, so a wide round's chunked parallel RPC calls must count as a single blocking wait. Guards the round-trip
+      // accounting at widths the unit tests don't reach (they count RPC calls, not round trips).
+      if (scenario.newLogs === 0 && !scenario.deepCohort) {
+        expect(row.rpcRoundTrips).toBe(1);
+      }
+    }
+
+    const results: BenchResult[] = rows.flatMap(row => [
+      { name: `TagSync/${row.label}/tag-queries`, value: row.tagQueries, unit: 'tag-queries' },
+      { name: `TagSync/${row.label}/rpc-round-trips`, value: row.rpcRoundTrips, unit: 'round_trips' },
+      { name: `TagSync/${row.label}/rpc-blocking-time`, value: Number(row.rpcBlockingTimeMs.toFixed(2)), unit: 'ms' },
+    ]);
+
+    if (process.env.BENCH_OUTPUT) {
+      await mkdir(path.dirname(process.env.BENCH_OUTPUT), { recursive: true });
+      await writeFile(process.env.BENCH_OUTPUT, JSON.stringify(results, null, 2));
+    }
+  }, 600_000);
+});