Imported from qdrant/skills (
skills/qdrant-performance-optimization/indexing-performance-optimization/SKILL.md). Install upstream withnpx skills add qdrant/skills --skill indexing-performance-optimization. Copyright stays with the author.
What to Do When Qdrant Indexing Is Too Slow
Qdrant does NOT build HNSW indexes immediately. Small segments use brute-force until they exceed indexing_threshold_kb (default: 10000 KB, about 10 MB). Search during this window is slower by design, not a bug.
- Understand the indexing optimizer Indexing optimizer
Uploads/Ingestion Too Slow
Use when: upload or upsert API calls are slow. Identify bottleneck: client-side (network, batching) vs server-side (CPU, disk I/O)
For client-side, optimize batching and parallelism:
- Use batch upserts (64-256 points per request) and 2-4 parallel upload streams Bulk upload
For server-side, optimize Qdrant configuration and indexing strategy:
- Create 2 to 4 shards per machine; each shard has an independent update worker, so give each one its own upload stream Multiple shards
- Create payload indexes before HNSW builds (needed for filterable vector index) Payload index
Optimizer Stuck or Taking Too Long
Use when: optimizer running for hours, not finishing.
- Check actual progress via optimizations endpoint (v1.17+) Optimization monitoring
- Large merges and HNSW rebuilds legitimately take hours on big datasets
- Check CPU and disk I/O (HNSW is CPU-bound, merging is I/O-bound, HDD is not viable)
- If
optimizer_statusshows an error, check logs for disk full or corrupted segments
HNSW Build Time Too High
Use when: HNSW index build dominates total indexing time.
- Reduce
m(default 16, good for most cases, 32+ rarely needed) HNSW params - Reduce
ef_construct(100-200 sufficient) HNSW params - Leave
max_indexing_threadsat 0 (auto) or cap it; the configuration reference suggests 8 to 16 Configuration - Use GPU for indexing (v1.13+): run a
gpu-nvidiaorgpu-amdimage tag (e.g.qdrant/qdrant:gpu-nvidia-latest) on a Vulkan 1.3 GPU and setgpu.indexing: trueGPU indexing
HNSW Index for Multi-Tenant Collections
Use when: every query filters by a tenant field and the global HNSW build dominates indexing time.
- Skip the global HNSW and build per-tenant graphs: at collection creation, set
payload_mto a non-zero value, thenm: 0. Queries without a tenant filter get slower, so only do this when cross-tenant search is rare Calibrate performance
Additional Payload Indexes Are Too Slow
Use when: HNSW builds slowly after adding payload indexes, especially on high-cardinality fields such as long text.
- Qdrant adds extra HNSW links for every payload index so filtered search keeps its quality. For a field with many unique values per point, disable those links and rely on the slower query-time ACORN strategy instead Disable extra edges
What NOT to Do
- Do not create payload indexes AFTER HNSW is built without triggering a re-index afterward -- filterable HNSW edges won't exist for the new index until segments are re-optimized
- Do not use
m=0for bulk uploads into an existing collection, it might drop the existing HNSW and cause long reindexing - Do not upload one point at a time (per-request overhead dominates)
