The whole benchmark fits on a small cloud instance. Apache-2.0 engine — the only bill is the box.
We bootstrapped 50,000,000 documents from each of five sources into OpenSearch on an engine capped at 2 vCPU / 1 GiB — then priced the same job on every comparable tool, using each vendor's own published numbers.
vendor pricing & guarantees read from official pages · 2026-08-20 · EST = derived estimate, methodology below
The whole benchmark fits on a small cloud instance. Apache-2.0 engine — the only bill is the box.
67 MB image, no Kafka cluster, no worker fleet. Peak memory across all five 50M runs: 376 MiB.
Full 50M bootstrap in 6–9 minutes per source, CPU averaging under half of the 2-core cap.
Exactly. Deterministic ids and a durable cursor make retries overwrite, never duplicate — checked doc-by-doc, five times.
Cost normalized to this benchmark's scale — ~50M documents / ~13 GB of change data per month — at each vendor's published rates. Guarantees quoted from their own documentation.
| Solution | What actually runs | Minimum footprint | Cost at 50M docs / 13 GB·mo | Joined docs | Delivery guarantee — their words |
|---|---|---|---|---|---|
| VentStreamopen source · Apache-2.0 | One Rust binary: CDC, joins, sinks — realtime and MCP roles optional | 2 vCPU / 1 GiB — 376 MiB observed peak | ~$15/mo infra · $0 license | in-engineparent + children composed | Cursor advances only after the sink's durable ack; deterministic ids make redelivery overwrite. Verified to the exact count in every published run. |
| Debeziumself-hosted · free | Kafka broker cluster + Connect workers (+ Schema Registry); you dedupe downstream | ~10 vCPU / 30–40 GiB, 5–6 nodesEST | $0 license + ~$250–400/mo infraEST; via MSK + Connect ~$530–650/mo | noflat events; join in Flink | Duplicates will always be visible to downstream consuming applications— at-least-once; consumers must dedupe. |
| Fivetranmanaged SaaS | Batch sync into warehouses; priced per Monthly Active Row | n/a | Volume pricing unpublished — sales only. Their small-volume example rate extrapolates to $25k/mo at 50M MAREST (upper bound; real quotes discount steeply) | nopost-load in dbt | Idempotent pipeline — prevents duplicate records from being loaded. |
| Airbytecloud or self-host | Managed ELT; self-hosting runs a Kubernetes pod fleet | Self-host: 4+ CPU / 8 GiB for the platform alone | ~$130+/mo at $10/GB — billed on serialized size, officially larger than raw | noflat records | An at-least-once delivery guarantee— uniqueness only in Append-Deduped mode. |
| Estuary Flowmanaged streaming | Streaming platform on Gazette; derivations for transforms | n/a | ~$207/mo — $0.50/GB + $100/connector × 2; connector fees dominate at this size | derivationsSQL / TypeScript | Generally guarantee end-to-end exactly-once… so long as the endpoint system can accommodate them. |
| Streamkapmanaged CDC | Managed CDC streaming (Kafka-based) | n/a | Starter $250/mo + $1.50/GB overage → ~$268/mo at 13 GB; Scale tier $2,100/mo | no | At-least-once-class streaming. |
| ClickPipesClickHouse Cloud | Managed Postgres CDC — ClickHouse destination only | n/a | ~$150/mo — $0.20/hr compute + $0.20/GB replicatedEST | nojoin at query time | Replication with ClickHouse merge/dedup semantics. |
| PGSyncopen source | Python daemon + Redis; Postgres → Elasticsearch only | 1 process + Redis | $0 license + small infra | nested docsfrom JSON schema | Depends on ordered checkpoint replay; single-maintainer, ~200 open issues. |
The honest frame isn't "cheaper" — it's which curve you're signing up for.
The vendor runs everything; the bill scales with your data. This benchmark's modest 13 GB/month lands between $130 and several thousand dollars — and doubles when your data does.
No license, but the price is a broker-and-worker estate (~10 vCPU / 30–40 GiB), the engineers who keep it healthy, and dedup logic in every consumer — because duplicates are documented behavior.
The 50M benchmark ran inside 2 vCPU / 1 GiB at 376 MiB peak. Volume costs minutes, not dollars: a bigger dataset takes longer on the same $15 instance — joins, idempotent delivery, and verification built into the engine.
All vendor claims read from official pages on 2026-08-20: Fivetran pricing · Airbyte pricing · Airbyte self-host requirements · Airbyte delivery semantics · Estuary pricing · Estuary guarantees · Streamkap pricing · Debezium FAQ — architecture & at-least-once · AWS MSK pricing · Confluent Cloud pricing · ClickPipes CDC billing · PGSync · Fivetran on idempotence.
Items marked ESTare derived: the Debezium footprint is standard practice (only its architecture is vendor-documented); Fivetran's figure extrapolates their own published small-volume example and is an upper bound; the ~13 GB normalization assumes this benchmark's 64-byte payloads — wider documents shift every per-GB bill upward, and VentStream's not at all. Benchmark data: BOOTSTRAP-50M-RESULTS-2026-08-20.md — reproducible with run-bootstrap.sh.
The harness is in the repo — run it against your own data and check the counts.