Benchmarking

vecparity benchmark --from memory://bench-source --to qdrant://mycollection \
    --num-records 200000 --dimension 128 --batch-size 500

vecparity benchmark seeds synthetic vectors into --from, then times a real SyncEngine run into --to, reporting throughput, batch count, and peak memory during the sync. Only the sync itself is timed; seeding is reported separately since it isn't part of what a real migration measures.

By default it seeds an in-memory source and syncs into an in-memory target, useful for exercising the harness itself. Point --to (or --from) at a real backend to measure something meaningful; the target collection/table needs to already exist with a matching dimension.

What this doesn't tell you

This is a synthetic, single-machine benchmark, not a substitute for testing at your actual data scale on your actual infrastructure. A number from a laptop against a single local Docker container and a number from a production migration between two managed services in the same region are not comparable. Treat this as a tool for producing your own numbers under your own conditions, not a source of numbers to quote.

It also doesn't count individual backend API calls, only sync batches (SyncEngine.stats.batches, one upsert() call per batch). Getting an exact network-call count would mean instrumenting six different SDKs individually; the batch count is a reasonable proxy without that cost.

A real run, for context

One data point, not a benchmark claim: 200,000 synthetic 128-dimension vectors, in-memory source, target a single local Docker Qdrant container (qdrant/qdrant:latest) over localhost, --batch-size 500.

Machine	13th Gen Intel Core i5-1335U, 15.7 GB RAM
Sync time	98.5s
Throughput	~2,030 records/sec
Batches	400
Peak memory during sync	462.6 MB

Re-run it yourself before trusting a number for anything that matters.