landing_page/qdrant-landing/content/course/essentials/day-4/large-scale-ingestion.md

title, short_description, description, weight, isLesson

title	short_description	description	weight	isLesson
Large-Scale Data Ingestion	Choose the right ingestion strategy for Qdrant: batched upserts, upload_points, and streaming uploads for million- and billion-scale workloads.	Master large-scale vector ingestion in Qdrant. Compare upsert, upload_points, and upload_collection, and learn how to stream a large dataset into a collection without loading it into memory.	4	true

{{< date >}} Day 4 {{< /date >}}

Large-Scale Data Ingestion

In vector search applications inserting a few thousand data points is straightforward but the dynamics change completely when dealing with millions or billions of records. Tiny inefficiencies in the ingestion process compound into significant time losses, increased memory pressure, and degraded search performance.

Every individual upsert call initiates a transaction that consumes memory and disk I/O to build parts of the index. At scale, this naive approach can overwhelm your system, causing upload times to spike and search quality to decrease. Efficiently preparing and loading your data into Qdrant is paramount for building a reliable and scalable AI application.

Choosing Your Ingestion Strategy

The Qdrant client gives you three ways to get points in. The first has you managing the batching; the other two hand that to the client.

• upsert is the basic write operation, and the one every client library has. Send points one at a time for real-time updates, or in batches for a bulk load, which minimizes the overhead of opening a connection per point. You decide the batch size and you send the requests.

• upload_points takes an iterable of models.PointStruct, the record-oriented shape: one object per point, carrying its own id, vector, and payload.

• upload_collection takes vectors, payload, and ids as separate arguments, the column-oriented shape.

Those last two are the same tool in two shapes. Both add parallelization, retries, and lazy batching on top of upsert, and because both accept iterators, neither needs the whole dataset in memory. Pick whichever matches how your data already sits: the docs note the two formats are equivalent internally and offered for convenience.

You can also skip generating embeddings yourself. With inference, you send the text or image and the model name, and Qdrant produces the vector on upsert.

upload_points and upload_collection are helpers in the client library rather than server endpoints, so what's available depends on your language:

Client	Bulk helper
Python	upload_points, upload_collection
Rust	upsert_points_chunked(request, chunk_size)
TypeScript, Go, Java, C#	Batched upsert calls

The bottleneck during upload is usually the client library, not the Qdrant server. If ingestion speed is your priority, the Rust client is the fastest option.

The Collection Configuration

When a collection is too large to hold in memory, each structure takes a memory parameter that says where it lives. pinned stays on the heap, cached is memory-mapped and pre-warmed, and cold is memory-mapped and read on demand.

from qdrant_client import QdrantClient, models
import os

client = QdrantClient(url=os.getenv("QDRANT_URL"), api_key=os.getenv("QDRANT_API_KEY"))

client.create_collection(
    collection_name="my_collection",
    vectors_config=models.VectorParams(
        size=512,
        distance=models.Distance.COSINE,
        datatype=models.Datatype.FLOAT16,
        memory=models.Memory.COLD,
    ),
    payload=models.PayloadStorageParams(memory=models.Memory.COLD),
    quantization_config=models.BinaryQuantization(
        binary=models.BinaryQuantizationConfig(memory=models.Memory.PINNED),
    ),
    hnsw_config=models.HnswConfigDiff(memory=models.Memory.PINNED),
)

Two things catch people out. pinned is rejected for dense vectors, which support only cached or cold. And max_segment_size and indexing_threshold are both measured in kilobytes rather than points. See Memory Tiers for which tier suits which structure.

memory arrived in Qdrant v1.19. If you are following older material, it replaces on_disk on the vectors, on_disk_payload on the collection, always_ram in the quantization config, and on_disk in the HNSW config. Those still work but are deprecated, and pinned has no equivalent among them.

The Upload Process

Hand the method an iterable and it takes care of the requests. Because it accepts an iterator, you can feed it a generator that reads from disk as it goes, rather than materializing the whole set first.

import tqdm

client.upload_collection(
    collection_name="my_collection",
    vectors=embeddings,
    payload=payloads,
    ids=tqdm.tqdm(ids),
    batch_size=256,
    parallel=4,
)

A few things worth knowing about these parameters:

• ids must be unique across the whole upload. Writes are idempotent, so a point sent under an id that already exists overwrites it instead of erroring. That is what you want on a retry, and what bites you if two batches reuse the same numbers.
• batch_size controls how many points go in each request. The Bulk Upload guide covers how to pick it, along with sharding and payload indexes.
• parallel starts worker processes. Each one opens its own connection, so if batches begin failing after you raise it, drop back to 1.
• tqdm around any of the iterables gives you progress. There is no show_progress parameter.
• prefer_grpc=True on the client skips JSON serialization on every batch.
• update_mode=models.UpdateMode.INSERT_ONLY (v1.17) makes a resumed upload skip points already in the collection rather than rewriting them. See Update Mode.
• On Windows and macOS, parallel greater than 1 needs the upload call behind a if __name__ == "__main__": guard in a script, because Python starts workers by re-importing your module. Without it the upload hangs instead of failing. Notebooks are unaffected.

Start small and test. Before attempting to upload your entire dataset, ingest a smaller chunk to validate your configuration and process.

Try it hands-on in the Google Colab notebook.

See it at 400 million points in the LAION-400M benchmark.