Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/glean-pack/skills/glean-performance-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill glean-performance-tuning. Copyright stays with the author (MIT).

Glean Performance Tuning

Overview

Glean's enterprise search API handles search queries across multiple connectors, bulk document indexing, and connector sync throughput. Search latency compounds when querying across dozens of datasources simultaneously. Large indexing jobs (10K+ documents) require careful batching to avoid rate limits and maintain connector sync schedules. Optimizing batch sizes, caching frequent search results, and tuning connector configurations reduces search P95 latency and keeps indexing pipelines within SLA windows.

Caching Strategy

const cache = new Map<string, { data: any; expiry: number }>();
const TTL = { search: 60_000, suggestions: 30_000, datasources: 600_000 };

async function cached(key: string, ttlKey: keyof typeof TTL, fn: () => Promise<any>) {
  const entry = cache.get(key);
  if (entry && entry.expiry > Date.now()) return entry.data;
  const data = await fn();
  cache.set(key, { data, expiry: Date.now() + TTL[ttlKey] });
  return data;
}
// Search results expire fast (1 min). Datasource metadata is stable (10 min).

Batch Operations

import PQueue from 'p-queue';
const BATCH_SIZE = 100;

async function indexDocsBatched(glean: any, dsName: string, docs: any[]) {
  const batches = [];
  for (let i = 0; i < docs.length; i += BATCH_SIZE) batches.push(docs.slice(i, i + BATCH_SIZE));
  const queue = new PQueue({ concurrency: 3, interval: 500 });
  await Promise.all(batches.map(batch =>
    queue.add(() => glean.indexDocuments(dsName, batch))
  ));
}

Connection Pooling

import { Agent } from 'https';
const agent = new Agent({ keepAlive: true, maxSockets: 15, maxFreeSockets: 5, timeout: 30_000 });
// High socket count for parallel indexing across multiple datasources

Rate Limit Management

async function withGleanRateLimit(fn: () => Promise<any>): Promise<any> {
  try { return await fn(); }
  catch (err: any) {
    if (err.status === 429) {
      const retryMs = parseInt(err.headers?.['retry-after'] || '5') * 1000;
      await new Promise(r => setTimeout(r, retryMs));
      return fn();
    }
    throw err;
  }
}

Monitoring

const metrics = { searches: 0, indexOps: 0, cacheHits: 0, p95LatencyMs: 0, errors: 0 };
const latencies: number[] = [];
function trackSearch(startMs: number, cached: boolean) {
  const lat = Date.now() - startMs; latencies.push(lat); metrics.searches++;
  if (cached) metrics.cacheHits++;
  latencies.sort((a, b) => a - b);
  metrics.p95LatencyMs = latencies[Math.floor(latencies.length * 0.95)] || 0;
}

Performance Checklist

• Batch indexing calls at 100 docs per request with 3 concurrent workers
• Use incremental indexing for real-time updates (< 100 docs)
• Switch to bulkindexdocuments for daily full refreshes (> 1K docs)
• Cache repeated search queries with 1-min TTL
• Set descriptive document titles and full body text for relevance
• Keep connector sync schedules staggered to avoid burst load
• Monitor P95 search latency and indexing throughput
• Enable keep-alive connections with high socket count for parallel ops

Error Handling

Prerequisites

• A baseline for latency percentiles, error rate, freshness lag, queue depth, and authorized-search coverage by datasource.
• A staging workload made of synthetic, bounded data and an approved error budget; do not capture real queries or content as a performance trace.
• A rollback revision for cache TTL, concurrency, batching, and connector schedules before production changes.

Instructions

1. Establish a baseline and target one bottleneck at a time: source read, transform, index submission, permission sync, or query.
2. Test bounded batch and concurrency changes in staging with backpressure, idempotency, and strict retry limits.
3. Confirm cache invalidation after ACL and document updates so lower latency never serves unauthorized or stale results.
4. Canary one datasource, monitor latency, freshness, rate limits, and synthetic allow/deny probes, then promote or roll back.
5. Record the configuration revision and review cost, error budget, and authorization evidence together.

Output

Return a tuning receipt with baseline and canary percentile bands, batch/concurrency/TTL revisions, rate-limit and freshness outcomes, authorization probes, owner approval, and rollback reference. Use aggregates only.

Examples

source=sandbox-guides; p95=420ms->310ms; batch=50; concurrency=2; freshness=pass; allow=pass; deny=pass; rollback=perf-r9 documents a safe canary.

Resources

• Glean Developer Portal
• Glean Indexing API Guide

Next Steps

See glean-reference-architecture.