Real-Time Analytics
Sub-second dashboards and data products on petabytes of data at any concurrency
Data Warehousing
Sub-second analytics on open lakehouse formats with no vendor lock-in
Observability in the AI Era
The most cost-effective alternative to Elasticsearch observability
Context Engineering
Hybrid search and fresh context for RAG, agents, and LLMs
Financial Services
Real-time analytics and fraud detection for banking, capital markets, and on-chain data
Ad & Media tech
Advertiser-facing reporting, content search, and game telemetry on one engine
Automobile & Transportation
Fleet telemetry, real-time recommendations, and multimodal autonomous driving search
Your advertisers refresh a dashboard and judge you by what loads. When creative copies sit in a search engine, and performance data sits in an OLAP store, every advertiser query joins across two systems and your customer waits. VeloDB, the real-time analytics and search database from the creators of Apache Doris, serves customer-facing reporting, content search, and log analytics in one engine, proven at a Billion-row scale.
Submit your inquiries here, and our team will get back to you
Kwai, the short-video and live-streaming platform behind Kling AI, runs a TikTok-scale ads business for more than 400 million daily users. Creative data lived in Elasticsearch and performance data in ClickHouse, joined through external tables. One engine now serves full-text search and aggregation on trillion-row tables.
The data team behind Naraka: Bladepoint, Where Winds Meet, and Knives Out ran Hive, Spark, Trino, Elasticsearch, HBase, and ClickHouse behind one platform. It consolidated onto a Doris and Iceberg lakehouse, cut the stack to one engine, and made ingested data queryable within seconds.
Tencent Music (NYSE: TME) ran content library search on Elasticsearch beside a separate OLAP layer, storing the same data twice. Inverted indexes now serve artist, track, and text search from the same tables as the analytics, and full ingestion dropped from more than 10 hours to under 3.
MiniMax, the AI company behind Talkie, a top-five free entertainment app in the US, outgrew a Loki-based logging stack where regex queries scanned entire datasets and spiked compute. Its Doris-based platform now serves petabyte-scale log search at over 99.9% availability.
Powering real-time data for 10,000+ teams
Ad and media platforms sell reporting to their own customers: campaign metrics, audience insight, catalog analytics. When working with multi-modal media assets and metrics, multiple systems are typically needed to handle the search and analytic requests, resulting in latency that is felt by the end user.
Batch processing, ad hoc queries, real-time lookups, and analytics each get their own engine, and each engine brings its own pipelines, its own specialists, and its own on-call rotation. Data crosses several systems before an analyst can query it.
At platform scale the bills arrive from every direction: search indexes inflate storage, full data writes stretch past 10 hours, and full-text scans spike compute at query time. The analytics product gets more expensive to run exactly as it succeeds.
Inverted indexes serve full-text queries next to SQL aggregations, so creative search, catalog lookup, and performance analytics run on the same tables. Apache Doris 4.0 added BM25 relevance scoring, and 4.1 adds a search() function with Elasticsearch-compatible DSL. Kwai cut query latency 64% to 90% after moving both sides of its ad stack onto one engine, and improved storage efficiency roughly 60% against Elasticsearch.
Sub-second latency holds at thousands of concurrent queries, and point lookups reach 30,000 QPS per node. Advertiser dashboards, in-app reporting, and internal analysis share one engine through traffic peaks. Kwai keeps its slow query rate under 5% across 4,000+ query templates. Tencent Music runs search and analytics in separate resource groups on one cluster, which took its alert volume from more than 20 a day to single digits a month.
Point lookups, complex joins, high-concurrency aggregations, and log search run in one engine, and the next use case is a new table, not a new database. VeloDB also queries Iceberg, Hudi, and Paimon tables in place, which is how NetEase Games consolidated its 6-system stack into a lakehouse serving 15 million queries a day. That team reports 99%+ compatibility in production converting its existing Trino, Presto, Hive, PostgreSQL, and ClickHouse SQL, so consolidating did not mean rewriting the analytics.
Columnar storage with ZSTD compression cuts the search-engine storage overhead: Tencent Music now holds in 195 GB what Elasticsearch stored in 698 GB, taking operating costs down 80%. MiniMax compresses logs 5:1 and tiers anything older than 7 days out to object storage, for 70% lower storage cost. Merge-on-write ingestion sustains heavy writes while the data stays queryable.
Kwai serves every external advertiser and e-commerce seller on its platform from one engine, with write throughput up 3x and unified observability cutting troubleshooting time 80%. Read here
KwaiWhen a game ships an update, NetEase Games tracks the funnel from patch download to login across 1.4 billion records. That query dropped from 20 seconds to under 2, with peak memory down from 54 GB to 4.2 GB, on a platform running 20+ Doris clusters across hundreds of nodes. Read here
NetEase GamesTencent Music removed its separate Elasticsearch layer, cutting operating costs 80% and full ingestion from more than 10 hours to under 3. Queries that took minutes with LIKE matching now return in seconds. Read here
Tencent MusicMiniMax queries 1 billion log records within 2 seconds while writing at 10 GB/s, on one platform serving every business line at over 99.9% availability. Read here ByteDance serves 1 billion 768-dimension vectors from roughly 500 GB of memory on a single server, at about 400 ms p95 for hybrid retrieval.
MiniMaxVeloDB is a real-time analytics and search database from the creators of Apache Doris. One engine handles point lookups, complex joins, high-concurrency aggregations, full-text search, and vector search on data that changes by the second.
Ad analytics needs full-text search over creative and keyword data next to aggregations over performance metrics, under heavy writes and constant advertiser queries. VeloDB serves both from one engine. Kwai runs trillion-row ad tables this way, ingesting 300 million new rows a day at write peaks of 3 million rows per second per node. Query latency fell 64% to 90% against its previous Elasticsearch and ClickHouse stack.
Yes. Inverted indexes serve full-text search from the same tables as SQL analytics, so teams keep one copy of the data. Tencent Music cut operating costs 80% and improved write performance 4x after removing its Elasticsearch layer, holding in 195.4 GB what Elasticsearch stored in 697.7 GB. Apache Doris 4.0 added BM25 relevance scoring, and 4.1 adds a search() function with Elasticsearch-compatible DSL.
Yes. VeloDB serves log search and log analytics from one engine, with inverted indexes handling full-text and regex queries. MiniMax moved off Grafana Loki after regex scans kept spiking compute. It now queries 1 billion log records within 2 seconds and writes at 10 GB/s at over 99.9% availability, with 5:1 compression and tiered storage cutting storage cost 70%.
Yes. One engine covers point lookups, complex joins, high-concurrency aggregations, and log search, and it reads Iceberg, Hudi, and Paimon tables in place. NetEase Games consolidated Hive, Spark, Trino, Elasticsearch, HBase, and ClickHouse onto a VeloDB and Iceberg lakehouse, serving 200+ projects and 15 million queries a day on petabyte-scale storage. The team ran the move in two phases, replacing the real-time engines first and the batch layer second.
The SQL usually ports, and most of the effort goes into validating it. NetEase Games reports 99%+ compatibility in production converting Trino, Presto, Hive, PostgreSQL, and ClickHouse SQL dialects, and mapped its Trino resource limits onto VeloDB workload groups. Teams moving from Elasticsearch map Keyword fields to non-tokenized inverted indexes and Text fields to tokenized ones. Kwai and Tencent Music both ran parallel pipelines to validate consistency before cutting over, then decommissioned their old clusters.
