cartethyia 2.5.10 → 2.6.0

package/src/console/observability/store.ts

@@ -3,19 +3,25 @@ import { and, desc, eq, inArray } from "drizzle-orm";
 import type { CartethyiaDatabase } from "../../persistence/db-handle";
 import type { RedisClient } from "../../persistence/redis";
 import { apiKeys, networkPools, telemetryPayloads } from "../../persistence/schema";
+import { type TelemetryEventRecord } from "../../observability/event-log";
 import {
-  queryTelemetryEvents,
-  type TelemetryEventRecord,
-} from "../../observability/event-log";
-import { isGatewayError } from "../../observability/telemetry-status";
-import { queryIndexedBuckets, type IndexedBucket } from "../../observability/event-day-index";
-import {
-  LATENCY_BUCKETS_MS,
-  numeric,
-  percentileFromHistogram,
-} from "../../observability/event-compaction";
+  breakdownTelemetry,
+  chartTelemetry,
+  healthTelemetry,
+  queryTelemetryEvents as queryTelemetryEventPage,
+  summarizeHealthPerformance,
+  summarizeUsagePerformanceSql,
+  summarizeTelemetry,
+} from "../../observability/event-query";
+import { clientNameFromUserAgent } from "../../observability/client-name";
+
+// Re-exported from the canonical module: ingestion (`telemetryEventRow`) and
+// the backfill stamp the same value the console read side groups by, so the
+// derivation cannot drift between writer and reader. Existing console
+// imports of `clientNameFromUserAgent` keep working.
+export { clientNameFromUserAgent };
 import { cachedReadWithEtag, type CachedResult } from "../shared/read-through-cache";
-import { decodeDatedCursor, encodeCursor } from "../shared/cursor";
+import { decodeDatedCursor, encodeCursor, trimKeysetPage } from "../shared/cursor";
 import { CARTETHYIA_VERSION } from "../../version";
 import { CachedPreferencesReader, DrizzlePreferencesReader, type PreferencesReader } from "../../persistence/tenant-preferences";
 import type {
@@ -88,23 +94,6 @@ function periodStartDate(period: string): Date | undefined {
   return new Date(Date.now() - amount * unitMs);
 }
 
-
-/**
- * Client identity straight from the stored `User-Agent`: the first RFC 9110
- * product token carries both the client and its version (`bun/1.2.3`,
- * `codex-cli/0.1.0`), so no prefix catalog can drift from what the client
- * actually sent. Browser agents start with the historical `Mozilla/5.0`
- * token, which names no real client, so they report as `Browser`.
- */
-export function clientNameFromUserAgent(userAgent: string | null | undefined): string {
-  const raw = userAgent?.trim();
-  if (!raw) return "unknown";
-  if (/^mozilla\//i.test(raw)) return "Browser";
-  const product = raw.split(/\s+/, 1)[0] ?? "";
-  const slash = product.indexOf("/");
-  const token = slash <= 0 ? product : product.slice(0, slash + 1).concat(product.slice(slash + 1).split(/[;,)]/, 1)[0] ?? "");
-  return token || "unknown";
-}
 /** Time-series bucket width for the usage chart, chosen from the window length. */
 function chartBucketSeconds(period: string): number {
   const since = periodStartDate(period);
@@ -197,15 +186,13 @@ async function resolveProxyLabel(
 }
 const HEALTH_WINDOW_MS = 24 * 3_600_000;
 const TOP_N = 50;
+
 /**
- * How long a parsed window is shared. A dashboard poll fires ~8 usage
- * endpoints at once, and a slow parse outlives a 1.5 s window — so the
- * endpoints that miss it each re-read and re-parse the same files, which is how
- * one page load becomes several full parses of the retention window.
+ * Nearest-rank percentile over a sorted array — kept for the JS-side
+ * `summarizeUsagePerformance` (still used by callers that already hold a
+ * bounded event list). The usage endpoints read percentiles from SQL via
+ * `summarizeUsagePerformanceSql`, which is what the dashboard polls.
  */
-const PERIOD_MEMO_TTL_MS = 5_000;
-
-/** Nearest-rank percentile over a sorted array, matching `percentile_cont` closely enough for a health gauge. */
 function percentile(sorted: readonly number[], fraction: number): number {
   if (sorted.length === 0) return 0;
   const index = Math.min(sorted.length - 1, Math.max(0, Math.ceil(fraction * sorted.length) - 1));
@@ -254,93 +241,22 @@ export function summarizeUsagePerformance(
 }
 
 /**
- * The performance summary from indexed buckets rather than from events.
+ * The console dimension as the SQL reader's grouping key.
  *
- * The index keeps a latency sum, a sample count, and a histogram, so the mean
- * stays exact and the percentiles stay bucketed — the same precision the
- * per-event version had, without parsing the period's requests.
+ * `client` maps to `client_name` (stamped at ingestion) rather than
+ * `user_agent`: the grouping is now a plain GROUP BY on a column the writer
+ * filled, instead of a per-row `clientNameFromUserAgent` call in JavaScript.
  */
-export function summarizeBucketPerformance(
-  buckets: readonly IndexedBucket[],
-): UsagePerformanceSummary {
-  let latencySum = 0;
-  let latencyCount = 0;
-  let resolveSum = 0;
-  let resolveCount = 0;
-  let latencyMaxMs: number | null = null;
-  const histogram: number[] = [];
-  for (const bucket of buckets) {
-    latencySum += bucket.latencySum;
-    latencyCount += bucket.latencyCount;
-    // Read defensively: an index written before these fields existed carries
-    // neither, and summing `undefined` poisons the whole period with NaN —
-    // including the days that do have the data.
-    resolveSum += numeric(bucket.resolveSumMs);
-    resolveCount += numeric(bucket.resolveCount);
-    if (bucket.latencyMaxMs !== null)
-      latencyMaxMs = latencyMaxMs === null ? bucket.latencyMaxMs : Math.max(latencyMaxMs, bucket.latencyMaxMs);
-    for (let index = 0; index < bucket.latencyHistogram.length; index += 1) {
-      histogram[index] = (histogram[index] ?? 0) + (bucket.latencyHistogram[index] ?? 0);
-    }
-  }
-  const fill =
-    histogram.length > 0
-      ? histogram
-      : new Array<number>(LATENCY_BUCKETS_MS.length + 1).fill(0);
-  return {
-    avgResolveMs: resolveCount > 0 ? resolveSum / resolveCount : 0,
-    avgDurationMs: latencyCount > 0 ? latencySum / latencyCount : 0,
-    // `latencyMaxMs` is passed as the overflow answer: a percentile past the
-    // histogram's last edge is at most the slowest request actually recorded.
-    // Without it an overflow returned `null`, and `?? 0` turned a P95 of 40 s
-    // into "0 ms" — the slow tail reported as nothing at all.
-    latencyP50Ms: percentileFromHistogram(fill, 50, latencyMaxMs) ?? 0,
-    latencyP95Ms: percentileFromHistogram(fill, 95, latencyMaxMs) ?? 0,
-    latencyP99Ms: percentileFromHistogram(fill, 99, latencyMaxMs) ?? 0,
-    latencySampleCount: latencyCount,
-    resolveSampleCount: resolveCount,
-  };
-}
-
-/** The largest entries of a count map, capped. */
-function topCounts(counts: ReadonlyMap<string, number>, limit: number): Array<[string, number]> {
-  return [...counts].sort((a, b) => b[1] - a[1]).slice(0, limit);
-}
-
-/** The grouping value for one usage breakdown dimension. */
-function dimensionName(dimension: UsageDimension, event: TelemetryEventRecord): string | null {
-  if (dimension === "model") return event.requestedModel;
-  if (dimension === "key") return event.apiKeyId;
-  if (dimension === "client") return event.userAgent;
-  if (dimension === "client_ip") return event.clientIp;
-  return event.providerId;
-}
-
-interface BreakdownBucket {
-  name: string;
-  requests: number;
-  input: number;
-  output: number;
-  cached: number;
-  errors: number;
-  cost: number | null;
-  rateSum: number;
-  rateCount: number;
-}
-
-function emptyBreakdown(name: string): BreakdownBucket {
-  return { name, requests: 0, input: 0, output: 0, cached: 0, errors: 0, cost: null, rateSum: 0, rateCount: 0 };
+function breakdownDimension(dimension: UsageDimension): string {
+  if (dimension === "model") return "model";
+  if (dimension === "key") return "api_key";
+  if (dimension === "client") return "client";
+  if (dimension === "client_ip") return "client_ip";
+  return "provider";
 }
 
 /** Real Drizzle-backed observability repository used by the production console. */
 export class DrizzleObservabilityStore implements ObservabilityStore {
-  /**
-   * Per-(tenant, period) parsed event set, shared by the ~8 usage endpoints
-   * one dashboard poll fires in parallel. 1.5 s window: a poll burst shares
-   * one file parse; the next poll re-reads. In-flight shares one promise.
-   */
-  private readonly periodMemo = new Map<string, { at: number; promise: Promise<readonly TelemetryEventRecord[]> }>();
-  private readonly periodBucketMemo = new Map<string, { at: number; promise: Promise<readonly IndexedBucket[]> }>();
   constructor(
     private readonly db: CartethyiaDatabase,
     private readonly redis: RedisClient | undefined,
@@ -404,44 +320,27 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     let cacheHitRate = 0;
     let avgTokensPerSec = 0;
     try {
-      // Bounded 24h rolling window (matches the Overview copy). The file is the
-      // source, so the percentiles are computed over the retained events.
+      // Bounded 24h rolling window (matches the Overview copy), computed IN
+      // SQL. The previous shape materialized every retained event into a
+      // latency array and sorted it in JavaScript — the single largest read
+      // cost on the dashboard. `healthTelemetry` returns the totals and the
+      // database computes the percentiles.
       const since = new Date(Date.now() - HEALTH_WINDOW_MS);
-      const events = await queryTelemetryEvents({ tenantId, since });
-      const latencies: number[] = [];
-      let latencySum = 0;
-      let inputTokens = 0;
-      let cachedTokens = 0;
-      let outputTokens = 0;
-      let tokensPerSecSum = 0;
-      let tokensPerSecCount = 0;
-      for (const event of events) {
-        total += 1;
-        if (isGatewayError(event.status, event.httpStatus)) errors += 1;
-        if (event.latencyMs !== null) {
-          latencies.push(event.latencyMs);
-          latencySum += event.latencyMs;
-        }
-        if (event.cachedInputTokens !== null && event.cachedInputTokens > 0) {
-          inputTokens += event.inputTokens ?? 0;
-          cachedTokens += event.inputTokens ?? 0;
-          outputTokens += event.outputTokens ?? 0;
-        }
-        if (event.tokensPerSec !== null) {
-          tokensPerSecSum += event.tokensPerSec;
-          tokensPerSecCount += 1;
-        }
-      }
-      latencies.sort((a, b) => a - b);
-      avg = latencies.length > 0 ? Math.round(latencySum / latencies.length) : 0;
-      p95 = percentile(latencies, 0.95);
-      p99 = percentile(latencies, 0.99);
+      const totals = await healthTelemetry(this.db, tenantId, since);
+      const performance = await summarizeHealthPerformance(this.db, tenantId, since);
+      total = totals.requests;
+      errors = totals.errors;
+      avg = performance.latencyAvgMs === null ? 0 : Math.round(performance.latencyAvgMs);
+      p95 = performance.latencyP95Ms ?? 0;
+      p99 = performance.latencyP99Ms ?? 0;
+      const inputTokens = totals.inputTokens;
+      const cachedTokens = totals.cachedInputTokens;
       // Cached share of the prompt: cached/input, the same honest definition
       // usageSummary uses, so the health card and the Usage page cannot
       // disagree. (Cached share, not prompt share: input/(input+output) reads
       // ~99% and mislabels it a cache hit rate.)
       cacheHitRate = inputTokens > 0 ? Math.round((cachedTokens / inputTokens) * 10000) / 100 : 0;
-      avgTokensPerSec = tokensPerSecCount > 0 ? tokensPerSecSum / tokensPerSecCount : 0;
+      avgTokensPerSec = performance.avgTokensPerSec;
     } catch {
       databaseHealthy = false;
     }
@@ -514,23 +413,23 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
   }
 
   private async computeUsage(tenantId: string, period: string): Promise<UsageResponse> {
-    const events = await this.periodEvents(tenantId, period);
-    let requestsTotal = 0;
-    let requestsSucceeded = 0;
-    let requestsFailed = 0;
-    let tokensUsed = 0;
-    let estimatedCost = 0;
-    const models = new Map<string, number>();
-    const providers = new Map<string, number>();
-    for (const event of events) {
-      requestsTotal += 1;
-      if (event.status === "completed") requestsSucceeded += 1;
-      if (isGatewayError(event.status, event.httpStatus)) requestsFailed += 1;
-      tokensUsed += (event.inputTokens ?? 0) + (event.outputTokens ?? 0);
-      estimatedCost += event.estimatedCostUsd ?? 0;
-      if (event.requestedModel) models.set(event.requestedModel, (models.get(event.requestedModel) ?? 0) + 1);
-      if (event.providerId) providers.set(event.providerId, (providers.get(event.providerId) ?? 0) + 1);
-    }
+    const since = this.periodSince(period);
+    // Totals AND the top-model/top-provider counts come from SQL: the old
+    // shape parsed the whole retained window into an array to count two
+    // dimensions in JavaScript.
+    const [summary, models, providers] = await Promise.all([
+      summarizeTelemetry(this.db, tenantId, since),
+      breakdownTelemetry(this.db, tenantId, "model", since),
+      breakdownTelemetry(this.db, tenantId, "provider", since),
+    ]);
+    const requestsTotal = summary.requests;
+    // `completed` is a lifecycle status, not the gateway-error predicate: it
+    // is derived as requests minus the errors and the non-completed states the
+    // summary already counts.
+    const requestsFailed = summary.errors;
+    const requestsSucceeded = Math.max(0, requestsTotal - requestsFailed - summary.cancelled - summary.truncated);
+    const tokensUsed = summary.inputTokens + summary.outputTokens;
+    const estimatedCost = summary.estimatedCostUsd;
     return {
       tenantId,
       period,
@@ -539,53 +438,25 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
       requestsFailed,
       tokensUsed,
       estimatedCost,
-      topModels: topCounts(models, TOP_N).map(([modelId, count]) => ({ modelId, count })),
-      topProviders: topCounts(providers, TOP_N).map(([providerId, count]) => ({ providerId, count })),
+      topModels: models
+        .slice(0, TOP_N)
+        .map((group) => ({ modelId: group.key, count: group.requests })),
+      topProviders: providers
+        .slice(0, TOP_N)
+        .map((group) => ({ providerId: group.key, count: group.requests })),
     };
   }
 
-  /** The tenant's retained events inside a usage period. `all` keeps everything. */
   /**
-   * The indexed buckets covering a period, for one tenant.
+   * The retained window for one usage period.
    *
-   * Reads the per-day index files instead of the event files. A usage query no
-   * longer parses every request in the window and filters it in memory, so its
-   * cost no longer grows with traffic.
+   * `all` means the retained SQL window (the database's own retention), not
+   * "read every file ever written": `telemetry_events` is pruned by the same
+   * retention rule as the raw files, so the SQL store already holds exactly
+   * what a period can show.
    */
-  private periodBuckets(tenantId: string, period: string): Promise<readonly IndexedBucket[]> {
-    const key = `${tenantId}:${period}`;
-    const now = Date.now();
-    const hit = this.periodBucketMemo.get(key);
-    if (hit && now - hit.at < PERIOD_MEMO_TTL_MS) return hit.promise;
-    const since = periodStartDate(period);
-    // Fold before reading. Events appended since the scheduler last ran are
-    // otherwise invisible, and a usage page that under-reports is worse than one
-    // that is a moment slow.
-    const promise = queryIndexedBuckets({
-      tenantId,
-      foldFirst: true,
-      ...(since === undefined ? {} : { since }),
-    });
-    this.periodBucketMemo.set(key, { at: now, promise });
-    void promise.catch(() => {
-      if (this.periodBucketMemo.get(key)?.promise === promise) this.periodBucketMemo.delete(key);
-    });
-    return promise;
-  }
-
-  private periodEvents(tenantId: string, period: string): Promise<readonly TelemetryEventRecord[]> {
-    const key = `${tenantId}:${period}`;
-    const now = Date.now();
-    const hit = this.periodMemo.get(key);
-    if (hit && now - hit.at < PERIOD_MEMO_TTL_MS) return hit.promise;
-    const since = periodStartDate(period);
-    const promise = queryTelemetryEvents(since === undefined ? { tenantId } : { tenantId, since });
-    this.periodMemo.set(key, { at: now, promise });
-    // A rejected parse must not poison the memo for the next poll.
-    void promise.catch(() => {
-      if (this.periodMemo.get(key)?.promise === promise) this.periodMemo.delete(key);
-    });
-    return promise;
+  private periodSince(period: string): Date {
+    return periodStartDate(period) ?? new Date(0);
   }
 
   async usageSummary(tenantId: string, period: string): Promise<CachedResult<UsageSummaryResponse>> {
@@ -595,58 +466,36 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
   }
 
   private async computeSummary(tenantId: string, period: string): Promise<UsageSummaryResponse> {
-    const buckets = await this.periodBuckets(tenantId, period);
-    const performance = summarizeBucketPerformance(buckets);
-    let requests = 0;
-    let inputTokens = 0;
-    let cachedTokens = 0;
-    let outputTokens = 0;
-    let errors = 0;
-    let cancelled = 0;
-    let truncated = 0;
-    let estimatedCostUsd = 0;
-    let unpriced = 0;
-    let tokensPerSecSum = 0;
-    let tokensPerSecCount = 0;
-    const statuses = new Map<number, number>();
-    for (const row of buckets) {
-      // Folded from the index's own per-status counts. These used to be left
-      // empty: the loop below never filled `statuses`, so every status chip and
-      // the requests-list filters read zero for any day whose events had been
-      // folded — while `requests` beside them read the real total.
-      for (const [status, count] of Object.entries(row.byStatus ?? {})) {
-        const code = Number(status);
-        if (!Number.isFinite(code)) continue;
-        statuses.set(code, (statuses.get(code) ?? 0) + count);
-      }
-      requests += row.requests;
-      inputTokens += row.inputTokens;
-      outputTokens += row.outputTokens;
-      cachedTokens += row.cachedTokens;
-      errors += row.failed;
-      cancelled += row.cancelled;
-      truncated += row.truncated;
-      estimatedCostUsd += row.estimatedCostUsd;
-      unpriced += row.unpriced;
-      tokensPerSecSum += row.tokensPerSecSum;
-      tokensPerSecCount += row.tokensPerSecCount;
-    }
+    const since = this.periodSince(period);
+    // Totals, status breakdown, and latency percentiles all computed in SQL.
+    // This used to parse the whole retained window into JavaScript to sum six
+    // counters and sort a latency array.
+    const [summary, statuses, performance] = await Promise.all([
+      summarizeTelemetry(this.db, tenantId, since),
+      breakdownTelemetry(this.db, tenantId, "status", since),
+      summarizeUsagePerformanceSql(this.db, tenantId, since),
+    ]);
+    const inputTokens = summary.inputTokens;
+    const cachedTokens = summary.cachedInputTokens;
     return {
       period,
       totals: {
-        statusCounts: [...statuses].sort(([a], [b]) => a - b).map(([status, count]) => ({ status, count })),
-        requests,
+        statusCounts: statuses
+          .map((group) => ({ status: Number(group.key), count: group.requests }))
+          .filter((entry) => Number.isFinite(entry.status))
+          .sort((a, b) => a.status - b.status),
+        requests: summary.requests,
         inputTokens,
         cachedTokens,
-        outputTokens,
-        errors,
-        cancelled,
-        truncated,
+        outputTokens: summary.outputTokens,
+        errors: summary.errors,
+        cancelled: summary.cancelled,
+        truncated: summary.truncated,
         ...performance,
-        estimatedCostUsd,
-        partial: unpriced > 0,
+        estimatedCostUsd: summary.estimatedCostUsd,
+        partial: summary.unpriced > 0,
         cacheHitRate: inputTokens > 0 ? (cachedTokens / inputTokens) * 100 : 0,
-        avgTokensPerSec: tokensPerSecCount > 0 ? tokensPerSecSum / tokensPerSecCount : 0,
+        avgTokensPerSec: performance.avgTokensPerSec,
       },
     };
   }
@@ -661,31 +510,37 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     const since = periodStartDate(period);
     const bucketSeconds = chartBucketSeconds(period);
     const stepMs = bucketSeconds * 1000;
-    const events = await this.periodEvents(tenantId, period);
+    // Bucketing happens in SQL. The old shape parsed the retained window to
+    // floor every event's timestamp into a JS Map.
+    const buckets = await chartTelemetry(
+      this.db,
+      tenantId,
+      this.periodSince(period),
+      bucketSeconds,
+      );
     const byTime = new Map<number, UsageChartBucket>();
-    for (const event of events) {
-      const at = new Date(event.createdAt).getTime();
-      if (!Number.isFinite(at)) continue;
-      const start = Math.floor(at / stepMs) * stepMs;
-      const bucket = byTime.get(start) ?? { t: new Date(start).toISOString(), requests: 0, input: 0, cached: 0, output: 0 };
-      bucket.requests += 1;
-      bucket.input += event.inputTokens ?? 0;
-      bucket.cached += event.cachedInputTokens !== null && event.cachedInputTokens > 0 ? (event.inputTokens ?? 0) : 0;
-      bucket.output += event.outputTokens ?? 0;
-      byTime.set(start, bucket);
+    for (const bucket of buckets) {
+      byTime.set(bucket.at, {
+        t: new Date(bucket.at).toISOString(),
+        requests: bucket.requests,
+        input: bucket.inputTokens,
+        // `cached` is the cached share of the prompt: the SQL reader sums
+        // cached_input_tokens, which is the same number the old JS loop used.
+        cached: bucket.inputTokens,
+        output: bucket.outputTokens,
+      });
     }
-    const rows = [...byTime.keys()];
-    const first = rows.length > 0 ? Math.min(...byTime.keys()) : undefined;
+    const first = buckets.length > 0 ? buckets[0]!.at : undefined;
     const startMs = since ? since.getTime() : first;
     if (startMs === undefined) return { buckets: [] };
     const aligned = Math.floor(startMs / stepMs) * stepMs;
-    const buckets: UsageChartBucket[] = [];
+    const out: UsageChartBucket[] = [];
     for (let at = aligned; at <= Date.now(); at += stepMs) {
-      buckets.push(
+      out.push(
         byTime.get(at) ?? { t: new Date(at).toISOString(), requests: 0, input: 0, cached: 0, output: 0 },
       );
     }
-    return { buckets };
+    return { buckets: out };
   }
 
   async usageBy(
@@ -703,74 +558,33 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     dimension: UsageDimension,
     period: string,
   ): Promise<UsageByResponse> {
-    // The breakdown is read from the per-name sums the day index maintains, not
-    // from the events: the console asks how much went to each model, and the
-    // index already carries exactly that. Parsing the window to recount it was
-    // the whole cost of this endpoint.
-    const buckets = await this.periodBuckets(tenantId, period);
-    const groups = new Map<string, BreakdownBucket>();
-    // Fall back to the events only when the index carries nothing for the
-    // period — an install that has not folded yet, or a day whose index is
-    // missing. Never alongside it: the same request is in both, and counting it
-    // twice would double every column.
-    if (buckets.length === 0) {
-      for (const event of await this.periodEvents(tenantId, period)) {
-        const raw = dimensionName(dimension, event);
-        if (raw === null || raw.length === 0) continue;
-        const name = dimension === "client" ? clientNameFromUserAgent(raw) : raw;
-        const bucket = groups.get(name) ?? emptyBreakdown(name);
-        bucket.requests += 1;
-        bucket.input += event.inputTokens ?? 0;
-        bucket.output += event.outputTokens ?? 0;
-        if (event.cachedInputTokens !== null && event.cachedInputTokens > 0) {
-          bucket.cached += event.inputTokens ?? 0;
-        }
-        if (isGatewayError(event.status, event.httpStatus)) bucket.errors += 1;
-        if (event.estimatedCostUsd !== null) bucket.cost = (bucket.cost ?? 0) + event.estimatedCostUsd;
-        if (event.tokensPerSec !== null) {
-          bucket.rateSum += event.tokensPerSec;
-          bucket.rateCount += 1;
-        }
-        groups.set(name, bucket);
-      }
-    }
-    for (const indexed of buckets) {
-      const source =
-        dimension === "model"
-          ? indexed.dimModel
-          : dimension === "provider"
-            ? indexed.dimProvider
-            : dimension === "client"
-              ? indexed.dimClient
-              : indexed.dimClientIp;
-      for (const [raw, entry] of Object.entries(source)) {
-        if (raw.length === 0) continue;
-        const name = dimension === "client" ? clientNameFromUserAgent(raw) : raw;
-        const bucket = groups.get(name) ?? emptyBreakdown(name);
-        bucket.requests += entry.requests;
-        bucket.input += entry.inputTokens;
-        bucket.output += entry.outputTokens;
-        bucket.cached += entry.cachedTokens;
-        bucket.errors += entry.errors;
-        bucket.cost = (bucket.cost ?? 0) + entry.estimatedCostUsd;
-        bucket.rateSum += entry.tokensPerSecSum;
-        bucket.rateCount += entry.tokensPerSecCount;
-        groups.set(name, bucket);
-      }
-    }
-    const mapped = [...groups.values()]
-      .map((bucket) => ({
-        name: bucket.name,
-        requests: bucket.requests,
-        input: bucket.input,
-        output: bucket.output,
-        cached: bucket.cached,
-        total: bucket.input + bucket.output,
-        errors: bucket.errors,
-        costUsd: bucket.cost,
-        cacheHitRate: bucket.input > 0 ? (bucket.cached / bucket.input) * 100 : 0,
-        avgTokensPerSec: bucket.rateCount > 0 ? bucket.rateSum / bucket.rateCount : 0,
-      }))
+    // Grouping happens in SQL and only the (bounded) groups come back. The
+    // old shape walked every retained event to accumulate a JS Map per
+    // dimension; `client` now groups on the ingestion-stamped `client_name`
+    // instead of re-deriving a name per row.
+    const groups = await breakdownTelemetry(
+      this.db,
+      tenantId,
+      breakdownDimension(dimension),
+      this.periodSince(period),
+    );
+    const mapped = groups
+      .map((group) => {
+        const input = group.inputTokens;
+        const output = group.outputTokens;
+        return {
+          name: group.key,
+          requests: group.requests,
+          input,
+          output,
+          cached: group.cachedInputTokens,
+          total: input + output,
+          errors: group.errors,
+          costUsd: group.costUsd,
+          cacheHitRate: input > 0 ? (group.cachedInputTokens / input) * 100 : 0,
+          avgTokensPerSec: group.avgTokensPerSec,
+        };
+      })
       .sort((a, b) => b.requests - a.requests)
       .slice(0, 100);
     if (dimension === "client_ip") {
@@ -841,17 +655,10 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
   }
 
   private async computeCache(tenantId: string, period: string): Promise<UsageCacheResponse> {
-    const events = await this.periodEvents(tenantId, period);
-    let inputTokens = 0;
-    let cachedTokens = 0;
-    let outputTokens = 0;
-    for (const event of events) {
-      inputTokens += event.inputTokens ?? 0;
-      outputTokens += event.outputTokens ?? 0;
-      if (event.cachedInputTokens !== null && event.cachedInputTokens > 0) {
-        cachedTokens += event.inputTokens ?? 0;
-      }
-    }
+    const summary = await summarizeTelemetry(this.db, tenantId, this.periodSince(period));
+    const inputTokens = summary.inputTokens;
+    const outputTokens = summary.outputTokens;
+    const cachedTokens = summary.cachedInputTokens;
     const tokenTotal = inputTokens + outputTokens;
     return {
       period,
@@ -868,24 +675,16 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     limit: number,
     httpStatus?: number,
   ): Promise<CachedResult<UsageRequestsResponse>> {
-    // Ask the reader for `limit` rows instead of folding the whole window and
-    // slicing it. It stops at the newest file that fills the page, so the
-    // console's most-polled endpoint no longer parses the retention window just
-    // to show 50 rows.
-    const since = periodStartDate(period);
-    const events = await queryTelemetryEvents({
-      tenantId,
-      limit,
-      ...(httpStatus === undefined ? {} : { httpStatus }),
-      ...(since === undefined ? {} : { since }),
-    });
-    const rows = events.slice(0, limit);
     const hideClientIp = await shouldMaskClientIp(this.preferences, tenantId);
     // Cached like its siblings, so a repeated poll costs a `304` instead of
     // another read. The mask flag is part of the key: it is a privacy setting,
     // so a page must never be served from an entry computed under the other
-    // value. The rows were already read above — this bounds the label lookup
-    // and the mapping, not the telemetry read.
+    // value.
+    //
+    // The telemetry read moved INSIDE the read-through: it used to run before
+    // it, so every 304 still paid a query. Now a cache hit costs nothing at
+    // all, and a miss pays one bounded SQL page (keyset, newest first) rather
+    // than a scan of the raw window.
     return this.readThroughWithEtag(
       this.usageCacheKey(
         tenantId,
@@ -893,6 +692,13 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
         "requests",
       ),
       async () => {
+        const page = await queryTelemetryEventPage(this.db, {
+          tenantId,
+          since: this.periodSince(period),
+          limit,
+          ...(httpStatus === undefined ? {} : { httpStatus }),
+        });
+        const rows = page.events.slice(0, limit);
         const keyIds = [...new Set(rows.map((row) => row.apiKeyId).filter((id): id is string => id !== null))];
         const labels = await this.apiKeyLabels(tenantId, keyIds);
         // One indexed read for the whole page, not one per row: presence lives
@@ -947,8 +753,8 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     tenantId: string,
     requestId: string,
   ): Promise<UsageRequestDetail | undefined> {
-    const rows = await queryTelemetryEvents({ tenantId, requestId });
-    const event = rows[0];
+    const page = await queryTelemetryEventPage(this.db, { tenantId, requestId });
+    const event = page.events[0];
     if (!event) return undefined;
     // The label, proxy, and mask lookups are independent: run them together
     // instead of three sequential round trips.
@@ -1013,17 +819,15 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     // `until` is inclusive while the cursor is exclusive, so rows at the exact
     // cursor timestamp are still compared by id here — the cheap part of the
     // predicate, now over one timestamp's worth of rows instead of the window.
-    const events = await queryTelemetryEvents({
+    // Keyset page: the cursor is the last row of the previous page, and the
+    // SQL reader pushes it into the WHERE clause instead of filtering a
+    // materialized window in JS.
+    const page = await queryTelemetryEventPage(this.db, {
       tenantId,
       limit: limit + 1,
-      ...(before === undefined ? {} : { until: new Date(before.createdAt) }),
+      ...(before === undefined ? {} : { cursor: encodeCursor({ createdAt: before.createdAt, id: before.id }) }),
     });
-    const rows = (before
-      ? events.filter((event) => event.createdAt < before.createdAt || (event.createdAt === before.createdAt && event.id < before.id))
-      : events
-    ).slice(0, limit + 1);
-    const overflow = rows.length > limit;
-    const trimmed = overflow ? rows.slice(0, limit) : rows;
+    const { rows: trimmed, hasMore: overflow } = trimKeysetPage(page.events, limit);
     const entries = trimmed.map(mapTelemetryEventRow);
     const lastEntry = entries[entries.length - 1];
     const nextCursor =
@@ -1043,7 +847,8 @@ export class DrizzleObservabilityStore implements ObservabilityStore {
     // `requestId`, which is a different column that merely defaults to the
     // same value. Predicating on it lets the reader stop at the first hit
     // instead of parsing the retention window to find one row.
-    const events = await queryTelemetryEvents({ tenantId, id: eventId });
+    const page = await queryTelemetryEventPage(this.db, { tenantId, id: eventId });
+    const events = page.events;
     const event = events.find(
       (candidate) =>
         candidate.id === eventId && (createdAt === undefined || candidate.createdAt === createdAt),