Infino vs OpenSearch
Both are Apache-2.0 search engines. OpenSearch stores a copy of your data on a JVM cluster you size and run. Infino searches one copy that stays in your object storage bucket, with full-text, vector, and SQL in the same engine.
infino diff opensearch --summary
the short answer
OpenSearch moves your data onto its own nodes: primary shards, replica shards, JVM heap, block storage, a cluster provisioned for the busiest hour and billed around the clock. Infino leaves the data in your bucket as Parquet and brings the engine to it: a stateless gateway in front of per-database workers, with storage and compute billed apart. At 10 million documents on real S3, warm full-text search measured 6.20 ms.
The founders ran OpenSearch at AWS, its creator and chief architect among them. Infino is what they built next: the same retrieval problems, without the cluster.
diff ARCHITECTURE.md
How is Infino different from OpenSearch?
The difference is where the data lives while you search it.
- OpenSearch is a Lucene cluster on the JVM. Documents are indexed into shards on the cluster’s own disks, replicas double them, and snapshots copy them out to a bucket for recovery.
- Infino inverts that: the bucket is the primary. Tables are Parquet files with the search indexes embedded, workers are compute with a cache, and losing a node loses nothing.
- Growing OpenSearch means adding nodes and rebalancing shards. Growing Infino means writing more files; compute is added separately, when queries need it.
infino diff opensearch --full
Infino vs OpenSearch at a glance
| dimension | opensearch | infino |
|---|---|---|
| runs as | a JVM / Lucene cluster you size, shard, and upgrade | a service: stateless gateway, per-database workers |
| your data lives | on cluster disks, doubled by replicas, snapshotted to a bucket | in your bucket, one copy, the source of truth |
| file format | Lucene segments, engine-private | standard Parquet; DuckDB, pandas, Spark read the same files |
| full-text search | BM25 | BM25 · 6.20 ms warm p50 at 10M docs on S3 |
| vector search | k-NN plugin, graph held in node RAM | built in, index inside the files · 13.81 ms warm p50 |
| hybrid search | search pipelines + normalization processors | one query, fused with RRF · 16.57 ms warm p50 |
| SQL | SQL plugin / PPL | DataFusion; search functions inside the query |
| scaling | add nodes, rebalance shards, reindex | storage and compute scale separately |
| idle cost | the full cluster, around the clock | storage only |
| license | Apache-2.0 | Apache-2.0 engine · commercial platform |
infino cost --vs opensearch
Is Infino cheaper than OpenSearch?
Yes, and the reason is structural, not a discount.
- An OpenSearch domain bills for provisioned nodes whether they are busy or not: data nodes sized for the peak hour (a common choice, r6g.2xlarge.search, lists at $0.669/hr), three dedicated cluster managers, EBS volumes, snapshot storage, and cross-AZ traffic.
- A replica is a second full copy of the data, on nodes you also pay for. AWS’s own sizing guidance puts the disk footprint near 3× the source data.
- Infino meters three things: storage at object-store rates, writes, and queries. At 10 million documents the measured numbers were $0.15 per month of storage and $0.14 per million full-text queries.
Model your own workload in the calculator → · Every measured number →
infino diff opensearch --their-wins
When is OpenSearch the better choice?
Some workloads should stay where they are.
- Observability. If you live in OpenSearch Dashboards, alerting, and anomaly detection for logs and traces, that suite is mature and Infino does not replace it.
- A RAM-hot cluster answers warm queries fast too. If your cluster is sized right and the bill is fine, latency alone is not a reason to move.
- Cold starts are real: at 10 million documents, Infino’s first query after a restart pays about half a second to open the table, and a cold vector index takes a few seconds. OpenSearch keeps everything hot because you pay it to.
- Deep plugin dependence. Ingest pipelines, security integrations, and years of tooling do not port automatically.
infino migrate --from opensearch
How do you migrate from OpenSearch to Infino?
In three steps, running both until the numbers convince you.
- Copy documents out with the scroll or search-after API, or replay them from the original source.
- Ingest into Infino tables in your bucket. Indexing happens on write; at 10 million documents the measured rate was 76.9K docs per second on one node.
- Point reads at Infino, compare results and cost side by side, then turn the cluster off. Enterprise and Cloud are Elasticsearch / OpenSearch DSL compatible, and an engineer walks you through the cutover.
Bring your doc count and current stack, we come back with numbers →
cat FAQ.md
Is Infino a drop-in replacement for OpenSearch?
No. Infino is a different engine with a different data layout, and pretending otherwise would waste your time.
- the commercial tiers are Elasticsearch / OpenSearch DSL compatible on the query side
- your data is reindexed into Parquet in your bucket; Lucene segments do not copy over
- migrations are engineer-guided: bring your doc count and current stack, we come back with numbers
Is Infino cheaper than OpenSearch?
Yes. You stop paying for idle peak capacity and for replica copies of your data.
- an OpenSearch domain bills its provisioned nodes whether anyone is querying or not, plus EBS volumes, snapshots, and cross-AZ traffic
- Infino meters storage, writes, and queries: measured $0.14 per million full-text queries at 10M docs
- the pricing page has a calculator with both cost models side by side
Does Infino support vector search like OpenSearch k-NN?
Yes. Vector and hybrid search are part of the engine, not a plugin.
- measured at 10M docs × 1024 dimensions on real S3: 13.81 ms warm p50, recall 0.96+
- OpenSearch k-NN holds its graph in node RAM; Infino keeps the index inside the Parquet files and pages in what a query touches
- hybrid keyword + vector runs as one query, fused with RRF: 16.57 ms measured
Can Infino handle the data volumes people run on OpenSearch?
Yes. Infino has run 1 billion documents of full-text search on a single node over Azure Blob.
- ingest at 1B measured 110.6K docs per second
- warm queries at 1B ran 97 ms to 6.31 s depending on shape; at 100M, 9.6 ms to 745 ms
- the benchmarks page lists every run, cold starts included
Is Infino open source like OpenSearch?
Yes. The engine is Apache-2.0, the same license as OpenSearch, with no feature gates on search.
- full-text, vector, hybrid, and SQL are all in the open core
- the managed platform (gateway, workers, console) is the commercial part
- pip install infino · npm install @infino-ai/infino · cargo add infino
When should I stay on OpenSearch?
Log analytics and observability, deep plugin dependence, or a cluster that is already sized right and paid for.
- OpenSearch Dashboards, alerting, and anomaly detection have no Infino equivalent
- a RAM-hot cluster answers cold queries that cost Infino more: about half a second to open a 10M-doc table after a restart
Who builds Infino?
The team that built and ran OpenSearch at AWS, including its creator and chief architect.
- they spent years sizing and tuning clusters at AWS; Infino removes the cluster instead
- the engine is developed in the open on GitHub
Which object stores does Infino run on?
Amazon S3 and S3-compatible stores, Azure Blob Storage, Google Cloud Storage, and local disk.
- managed SaaS or your own cloud account, from one build
- self-hosted, run it against your own bucket