Add Amazon DynamoDB vector search client

[eddy-aws]

Summary

Adds an Amazon DynamoDB vector search client, benchmarking DynamoDB's GA
vector index feature (CreateTable with VectorIndexes + SearchVectors) via
the standard boto3 SDK. No external vector store is required — vectors live in
the DynamoDB table alongside operational data.

What's included

• vectordb_bench/backend/clients/dynamodb/ — dynamodb.py, config.py, cli.py
• Registered in the DB enum (init_cls / config_cls / case_config_cls) and
  the vectordbbench CLI
• dynamodb optional-dependency extra (boto3>=1.40.0)
• Offline unit tests in tests/test_dynamodb.py

Features

• Credentials: default boto3 credential chain by default; optional explicit
  access key / secret / session token.
• Metrics: cosine / euclidean (L2) / dotproduct (IP), mapped to the
  COSINE / EUCLIDEAN / DOT_PRODUCT distance functions.
• Load: batch_write_item in 25-item batches with two layers of throttling
  defence — boto3 standard retry mode plus bounded exponential backoff (with
  full jitter) on UnprocessedItems re-drive.
• Partition key (--partition-count): when > 1, defines a SearchSchema
  HASH (vector index partition key), spreads vectors uniformly across
  partition values (id % N), and scopes each SearchVectors call to one
  randomly chosen partition value — the realistic single-partition access
  pattern the feature is built for (one scoped call per query, not a
  whole-index fan-out).
  • This uniform spread is a balanced baseline (no hot/empty partitions);
    real-world partition keys are usually skewed, so the figures represent the
    best-case even distribution rather than skewed production behaviour.
  • A scoped search covers only ~1/N of the dataset, while VectorDBBench
    computes recall against whole-dataset ground truth, so recall is expected
    to be ~1/N under partitioning; QPS/latency reflect the scoped access
    pattern. Use --partition-count 1 (default) for a whole-index recall
    benchmark with no partition key.
• Filters: label equality via an INLINE_FILTER SearchSchema element
  (SearchConditionExpression supports only =, so NumGE is intentionally not
  advertised).
• TopK capped at the API maximum of 100.

Testing

Offline unit tests cover config/metric parsing, the partition-key SearchSchema,
single-partition-scoped search, filter translation, throttling backoff, and CLI
wiring.

Validated end-to-end against a live DynamoDB vector index in us-east-1:

• 20k synthetic vectors — search returns the correct nearest neighbours.

• Real VectorDBBench run — Performance1536D50K (OpenAI, 50k × 1536-dim,
  COSINE, k=10), no partition key:

  Metric	Value
  Recall	1.0
  NDCG	1.0
  Max QPS (concurrency 5)	63.3
  Serial latency p99	79.8 ms
  Load duration (50k)	488 s

Notes

• Vector indexes require on-demand capacity, so the table is created with
  BillingMode=PAY_PER_REQUEST.
• The vector index backfills asynchronously after load; searches should poll
  until results are returned (the benchmark's load→search ordering handles this).

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: eddy-aws
To complete the pull request process, please assign xuanyang-cn after the PR has been reviewed.
You can assign the PR to them by writing /assign @xuanyang-cn in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment