Select all dimensions that meaningfully drive cost on a managed vector database bill in 2026.

Multi-select·Easy·4.0 · 0·~1 min·Asked atPinecone Tencent Wipro·Relevant atDatabricks

Attempt it

Managed vector databases bill on the same four axes across all the major vendors: **1. Storage**: how many vectors times their dimension. A 100M-vector index at 1024 dimensions stores ~400GB of float32 data; doubling either the vector count or the dimension doubles the storage bill. **2. Reads**: queries per second served. Each query consumes some amount of compute (graph traversal in HNSW, partition probes in IVF), and vendors charge a per-query 'read unit' or bundle them into a tier. **3. Writes**: upsert operations, which insert or update vectors. Writes are more expensive than reads per operation because they typically touch the index structure (graph edges in HNSW, cluster assignment in IVF). **4. Replication tier**: whether the index is single-AZ, multi-AZ, or multi-region. Each replica adds a roughly proportional cost. **The distractors:** the cosine-similarity algorithm is a build-time configuration that determines retrieval semantics but does not change how much hardware the vendor provisions. Naming conventions (snake_case vs camelCase) are obviously cosmetic and irrelevant. Both are testing whether you can separate operational costs from cosmetic or semantic choices.

**Four cost axes apply across every major managed vector DB in 2026:** storage, reads, writes, and replication. Two distractors test the candidate's ability to separate real cost levers from cosmetic ones. **Storage (vectors × dimension).** The dominant axis. Pinecone Serverless, Turbopuffer, Qdrant Cloud, and Weaviate Cloud all charge approximately per-vector-per-month with a dimension multiplier. Doubling the embedding dimension doubles the bill linearly. Doubling the corpus does the same. **Read throughput.** Each query consumes 'read units' or charges against a QPS-tiered plan. The unit cost depends on the index type (HNSW queries are cheaper than IVF probes at the same recall target) and on the dimension (larger vectors mean more distance compute per query). **Write throughput.** Upserts are more expensive than reads because they modify the index structure. HNSW graph updates require recomputing neighbor edges; IVF requires reassignment. Bursty writes especially trigger autoscaling charges. **Replication tier.** Single-AZ, multi-AZ, or multi-region replication each multiply the underlying storage and compute cost by the replica count. Production indexes typically run multi-AZ for availability. **Distractor 1: the distance metric.** Whether you choose cosine, dot product, or Euclidean is a build-time configuration that affects retrieval semantics. It does not change the hardware footprint and does not appear on the bill. **Distractor 2: naming conventions.** Obviously cosmetic; testing whether the candidate can dismiss irrelevant detail. This question screens for whether you understand the cost model well enough to size an index before deploying it, not just well enough to read the bill after the fact.

TL;DR

Four real cost axes: storage (vectors × dim), reads (QPS), writes (upserts), and replication tier. Distance metric and naming convention are distractors.

Memory aid

Easy to grasp

Think of a managed vector database the way you would think of renting warehouse space plus paying for the trucks that load and unload it. The space charge depends on how many boxes you store and how big each box is: that is the vectors × dimension axis. The truck charges depend on how often you ship things out (queries) and how often you ship things in (upserts). And if you ask for a second warehouse in another city as backup, that doubles the bill. Which color you paint the warehouse, or what specific way the trucks weigh the boxes, does not change the rent. Those are the distractors in the question.

Key concepts

Cost modeling

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Cost modeling for managed vector databases reduces to four orthogonal axes. Every major vendor in 2026 (Pinecone, Turbopuffer, Qdrant Cloud, Weaviate Cloud, Zilliz Cloud) exposes these axes either directly (serverless plans) or bundled into capacity tiers (pod-based plans). Understanding the axes well enough to size an index before deploying it is what separates a senior practitioner from someone reading the bill after the fact.

This question screens for that understanding by mixing the four real axes with two clear distractors. A candidate who picks the distance metric or the naming convention as cost levers has not internalized the model.

Storage: vectors × dimension × bytes_per_dim

The dominant axis. Every managed vector DB charges proportionally to total stored data.

For 100M vectors at 1024 dimensions in float32, raw storage is 100e6 × 1024 × 4 = 410GB. At float16 (an option on some vendors) it is 205GB. With Product Quantization at 32 bytes per vector it is 3.2GB. The compression ratio is the single largest cost lever after the dimension itself.

The dimension knob is what most teams underestimate. Migrating from text-embedding-3-small at 1536 dims to text-embedding-3-large at 3072 dims doubles the storage bill in one deploy. Migrating from a 768-dim BERT-based embedder to a 1024-dim modern embedder adds 33 percent. These migrations happen routinely and are invisible to the application layer.

The corpus-size knob is the other half. A chunker change that halves chunk size doubles the vector count without changing the source documents. Combined with a model-side dimension change, you can hit a 4x storage bill in a single sprint.

Reads and writes: not symmetric

Replication multiplies everything

The distractors and why they trip people up

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Cost axis	What drives it	Typical knob
Storage	vectors × dimension × bytes_per_dim	Compress vectors via PQ or BQ
Reads (QPS)	Per-query candidate count × distance compute	Lower ef_search / nprobe
Writes (upserts)	Per-upsert index maintenance cost	Bulk-load offline, avoid bursty inserts
Replication	Replica multiplier across all other axes	Match tier to RPO/RTO requirements

Real products, models, and research that use this idea.

Pinecone Serverless prices storage per vector-month with a dimension multiplier, plus separate read and write unit charges, plus multi-AZ tier multiplier.
Turbopuffer publishes a gigabyte-month storage price plus a per-query price plus a per-vector-write price; the four axes are explicit in the pricing page.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you estimate the read-unit cost of a query on HNSW versus IVF at the same recall target?

Measure the candidate count touched per query (ef_search for HNSW, nprobe × cluster_size for IVF) and multiply by per-distance compute. HNSW is usually cheaper at the same recall.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating the distance metric as a cost lever. It is a build-time choice that affects retrieval semantics, not the hardware the vendor provisions.

60 second bullets to scan on the way to the call.

The four cost axes: storage, reads, writes, replication
Why dimension is the largest single storage lever

Primary sources. Browse if you want the original framing.