loading
loading
SereneDB launches superfast S3 data search engine. German startup SereneDB’s Krummelanke is claimed to be the fastest database search engine in the world, with search and analytics in one product. The company says AI agents fire hundreds of hybrid find-and-count queries per second, yet every database in production was designed for a human user. Its Krummelanke engine is built to cope with high-speed agent-scale database search. Krummelanke is open-source under Apache 2.0, and both Postgres- and Elastic-compatible, and it runs full-text search over data sitting on S3. Co-founder and CEO Alexander Malandin said: “The agentic future is, above all, a data-load problem. Today an AI agent is a single bee visiting a single flower. What is coming is the swarm: more agents than employees in every organization, each firing hundreds of queries a second. Infrastructure built for human users is neither fast enough nor cost-efficient enough to survive that load. The agentic future requires a new generation of database: ultra-fast search and ultra-fast analytics in one engine, where search operations are measured in nanoseconds and a question over billions of records comes back in milliseconds. One that treats agents as first-class users, withstands their scale, and reaches data wherever it lives. That is what we built, and with SereneDB Krummelanke we are putting it into production.” SereneDB says analytical engines like Snowflake, BigQuery, DuckDB over Parquet, and lakehouses generally, read columnar data straight from S3 as a matter of course. The large sequential scans suit object storage's latency profile whereas full text search does not. It requires an inverted index which lists content items, such as words or numbers, and their mapping to their locations in a table, document or document set. Inverted index search needs lots of ransom accesses. Generally Lucene-based search engines store the index on fast local storage, separate from the main lakehouse data store. SereneDB says "searchable" data in an Elastic data estate means having a second copy on flash, “the most expensive tier you own.” We asked how SereneDB’s search differed from general object storage data search, such as Cloudian HyperStore search. CTO Andrey Abramov told us: “Cloudian is an S3-compatible storage vendor and HyperSearch is a service attached to the bucket. It collects the metadata of stored objects and indexes that metadata in a search component under the hood, this is a separate indexing system fed by copying. Many services offer search this way and they work well.” However, at the search engine level: “There are only a handful of core search technologies (core libraries) in the world: Lucene, Tantivy, IResearch (by SereneDB). Elasticsearch, OpenSearch, Solr and many more are Lucene derivatives and are copy-first. They cannot search data they don't hold locally, so anything you want searchable has to be extracted, shipped and duplicated into the search cluster.” SereneDB is different: “SereneDB indexes data where it already lives: Parquet files and Iceberg tables in S3 stay the source of truth. We hold only the derived index. The search itself - full-text with bm25 rangin/vector/geo over runs against that index with a small local hot set. The Iceberg case is the concrete example: a customer writes changes to an Iceberg table on S3, SereneDB picks them up, serves search and analytics over them and you can point as many SereneDB nodes as you like at the same table.” We asked if an object storage, database or lakehouse vendor could switch to using SereneDB’s search engine. For example, instead of Lucene and Lucene derivatives? Abramov said: “Yes, absolutely! We can not only replace ElasticSearch, OpenSearch, Solr, CrateDB (which are pure Lucene derivatives), but also augment them with analytical capabilities. For example, we have a dedicated page in our documentation on migrating from ElasticSearch." The company has run benchmarks comparing its search engine’s speed against popular alternatives, and beating them. Abramov said: "Our goal with these benchmarks was to demonstrate that SereneDB is a faster and modern alternative to the traditional search stack.” Benchmark Details SereneDB’s SearchBench includes 92 search and analytics queries over 100 million and 1 billion OpenTelemetry log records (the TextBench dataset published by ClickHouse, 15 columns per record, all of them stored) run in a single server, with all candidate engines tuned with published configuration data, and a 60-second limit per query. If the response takes longer than that the query search is classed as “did not finish.” The benchmark results for more than a dozen popular search engines can be seen on a dashboard here. The engines include Elasticsearch in four configurations, OpenSearch, ParadeDB, CrateDB, ArangoDB, TigerData (row and columnstore) and vanilla Postgres - with more coming. The timing is three runs per query, best of the last two, one query at a time from a single client, and result caches off. The dashboard measures latency, not throughput. Here is a (large) screenshot of its current state for the 100million test runs; Each results cell in the dashboard is the ratio to the fastest engine on that query, so the fastest is 1.0x, with the ratio calculated as a geomean; the geometric mean of those 92 ratios per engine. SereneDB says this is the usual way to average ratios so one outlier query doesn't dominate. - SereneDB's geomean is 1.05x at 100 million and 1.07x at 1 billion.It’s the fastest engine, or within a few percent of it, on average across the 92 queries. ElasticSearch is the next fastest, averaging 2.8x at 100 million. - TigerData columnstore, averages 2.24x at 1 billion, with a long list of timeouts); - Vanilla Postgres is 204x at 100 million and 50.69X at 1 billion. We wanted to compare the supplier scores more simply than this complicated screen dashboard, and built a chart to show the suppliers’ ordered geomean scores on the 100 million benchmark, taking data from the screenshot above; The score range is so great that we excluded the geomean scores >10 and redrew the chart; As SereneDB says, its search engine is the fastest of the bunch. SereneDB founding and funding SereneDB was founded in Berlin in 2025 by Alexander Malandin (CEO, ex-EMC, Dell and ArangoDB enterprise sales), Andrey Abramov (CTO, ex-Quest Software and EMC search R&D, ex-ArangoDB) and Valery Mironov (core engineer, ex-ArangoDB, ex-YDB). The eight-person engineering team has worked together for over 16 years; the search core, IResearch, has been in development since 2014. It announced $2.1 million pre-seed funding in December 2025, led by Entourage (Pieterjan Bouten's fund) and High-Tech Gründerfonds. SereneDB Krummelanke is open source under the Apache 2.0 license and available now on GitHub. Bootnote There is a lake in south east Berlin called Krumme Lanke. The word “Krumme” means crooked, bent or curved with “Lanke” meaning an elongated lake.
document fetched 2026-09-23 02:58 · canonical https://blocksandfiles.com/ai-ml/2026/09/22/serenedb-launches-superfast-s3-data-search-engine/5298324