Qdrant Memory Usage Optimization

Understanding memory usage

Qdrant operates with two types of memory:

• Resident memory (aka RSSAnon) - memory used for internal data structures like the ID tracker, plus components that stay fully in RAM. As of Qdrant 1.20 the ID tracker has its own memory tier (default pinned; cached lets the OS evict it, cold needs no up-front RAM), where it was always resident before. On Qdrant 1.19 or newer this is controlled per-component with memory: pinned (e.g. quantized vectors, payload indexes); see memory tier legacy settings for deployments on version 1.18 or older.

• OS page cache - memory used for caching disk reads, which can be released when needed. Original vectors are normally stored in page cache, so the service won't crash if RAM is full, but performance may degrade. On Qdrant 1.19 or newer this corresponds to memory: cached (pre-warmed into page cache at startup) or memory: cold (lazy disk reads, not pre-warmed); on 1.18 or older it's controlled via the on_disk boolean on vectors, HNSW config, sparse vector index, and payload index. See Memory Tiers docs (available on 1.19+).

It is normal for the OS page cache to occupy all available RAM, but if resident memory is above 80% of total RAM, it is a sign of a problem.

Memory usage monitoring

• Qdrant exposes memory usage through the /metrics endpoint. See Monitoring docs.

<!-- ToDo: Talk about memory usage of each components once API is available -->

How much memory is needed for Qdrant?

Optimal memory usage depends on the use case.

• For regular search scenarios, general guidelines are provided in the Capacity planning docs.

For a detailed breakdown of memory usage at large scale, see Large scale memory usage example.

Payload indexes and HNSW graph also require memory, along with vectors themselves, so it's important to consider them in calculations.

Additionally, Qdrant requires some extra memory for optimizations. During optimization, optimized segments are fully loaded into RAM, so it is important to leave enough headroom. The larger max_segment_size is, the more headroom is needed.

When to put HNSW index on disk

Putting frequently used components (such as HNSW index) on disk might cause significant performance degradation. On Qdrant 1.19 or newer this is set with memory: cold in hnsw_config; on 1.18 or older with hnsw_config.on_disk: true. There are some scenarios, however, when it can be a good option:

• Deployments with low latency disks - local NVMe or similar.

• Multi-tenant deployments, where only a subset of tenants is frequently accessed, so that only a fraction of data & index is loaded in RAM at a time.

• For deployments with inline storage enabled. On Qdrant 1.20 or newer, inline storage keeps the original vectors only in the HNSW index file, so hnsw_config.memory decides their tier and vectors.memory is ignored for those segments. Unless quantization has its own explicit memory, quantized vectors are cold when the HNSW index is cold and pinned in any other tier, so set it explicitly (for example pinned) if the index is cold but you want the quantized vectors in RAM Inline storage.

How to minimize memory footprint

The main challenge is to put on disk those parts of data, which are rarely accessed. Here are the main techniques to achieve that:

• Use quantization to store only compressed vectors in RAM. On Qdrant 1.20 or newer, 8-bit TurboQuant (bits8, 4x compression) is the recommended default over scalar quantization at the same compression (use scalar for the Manhattan metric), and it doesn't rescore by default, unlike bits2, bits1_5, and bits1 Quantization docs

• Use float16 or uint8 datatypes to reduce memory usage of vectors by 2x or 4x respectively, with some tradeoff in precision. uint8 only makes sense for vectors already quantized to 0-255. On Qdrant 1.19 or newer, the TurboQuant datatypes replace the original vectors with a compressed form (dense vectors and multivectors, not sparse): turbo4 (4 bits/dimension, ~8x, from 1.19), turbo8 (~4x, closer to float32 recall, from 1.20), and turbo16 (~2x, from 1.20, closer to float32 than float16). Any of them can be paired with 1-bit quantization for cheaper rescoring than pairing 1-bit quantization with full-precision vectors. A TurboQuant datatype is not TurboQuant quantization, which keeps the originals and adds a copy. Read more about vector datatypes in documentation

• Leverage Matryoshka Representation Learning (MRL) to store only small vectors in RAM while keeping large vectors on disk. Examples of how to use MRL with Qdrant Cloud inference: MRL docs

• For multi-tenant deployments with small tenants, vectors might be stored on disk because the same tenant's data is stored together Multitenancy docs

• For deployments with fast local storage and relatively low requirements for search throughput, it may be possible to store all components of vector store on disk. Read more about the performance implications of on-disk storage in the article

• For low RAM environments, enable async I/O (io_uring) for concurrent disk reads, which can significantly improve performance of on-disk storage: storage.performance.io_uring: auto on Qdrant 1.19 or newer (applies to every cold structure), async_scorer: true on 1.18 or older (vector rescoring only). Requires Linux with a kernel that supports io_uring Async I/O

• Keep payloads on disk: memory: cold is the default on Qdrant 1.19 or newer; on 1.18 or older, set on_disk_payload: true Default tiers

• Configure payload indexes to be stored on disk: memory: cold on Qdrant 1.19 or newer, on_disk: true on 1.18 or older docs

• Configure sparse vectors to be stored on disk: memory: cold on the sparse vector index on Qdrant 1.19 or newer (defaults to pinned), on_disk: true on 1.18 or older docs