Rockset

Short answer: Rockset is gone as a product. OpenAI announced the acquisition on June 21, 2024, and Rockset ended customer service on September 30, 2024. The team now builds retrieval infrastructure inside OpenAI. There is no managed cloud, no self-hosted build, and no open-source fork. If you ran Rockset in production, ClickHouse Cloud is the default replacement, and Tinybird is the closest match on developer experience.

What Rockset was, in plain terms

Rockset was a real-time analytics database. You streamed messy JSON into it from Kafka, DynamoDB, MongoDB, or S3, and you got fast SQL back within a second or two. No schema design meeting. No nightly batch job. No separate search cluster.

That sounds ordinary now. In 2018 it was not. Most teams had to choose: a warehouse for aggregations, a key-value store for lookups, and Elasticsearch for search. Rockset argued you should not have to choose at all.

Founders Venkat Venkataramani and Dhruba Borthakur started the company in 2016 after years at Facebook. Dhruba wrote RocksDB, the embedded storage engine that sits under a surprising share of modern databases. That lineage mattered, because Rockset's whole pitch rested on storage-layer tricks that most application teams could never build themselves.

The Converged Index: the idea that made Rockset different

Rockset's core invention was the Converged Index. Every field you ingested was written three times, into three different index structures:

• A row index, so point lookups by key stayed fast.
• A column index, so aggregations over one field did not scan the whole record.
• An inverted index, so search and highly selective filters resolved instantly.

Writing three copies costs storage and ingest throughput. Rockset traded both away on purpose. In exchange you got a database where no query pattern was a surprise, which is exactly what you want when a product team keeps inventing new filters on a live dashboard.

The other half of the design was schemaless ingest with strong typing. Rockset inferred types per field per document, so a field that was a string in some rows and an integer in others simply worked. Anyone who has spent a weekend fixing a broken warehouse load knows how much pain that removed.

Why OpenAI bought Rockset

OpenAI did not buy Rockset for its revenue. It bought a retrieval team.

Grounding a model in current, private, fast-changing data is the same problem Rockset had been solving for eight years: index new records within seconds, serve selective queries with low latency, and keep it all cheap enough to run continuously. Hiring that skill set from the open market would have taken years. Buying the company took a quarter.

The public product was collateral. Running a multi-tenant analytics cloud demands support engineers, compliance work, uptime commitments, and a sales motion — none of which OpenAI wanted. So the service was wound down on a roughly hundred-day clock while the engineers moved onto internal retrieval work.

The wind-down, step by step

• June 21, 2024 — OpenAI announced the acquisition. New signups closed shortly after.
• July 2024 — Existing customers received end-of-life notices with individual migration timelines.
• Through Q3 2024 — Export tooling ran; enterprise accounts negotiated pro-rated credits.
• September 30, 2024 — Customer service ended. Query endpoints stopped answering.
• After that date — No exports, no support, no reactivation. Anything not migrated was lost.

Compared with the average AI acquihire, this was handled well. Customers got a clear date, working export paths, and refunds. It was still, for anyone who had built a product on top of it, a forced migration inside one quarter.

The best Rockset alternatives in 2026

No single engine copies the Converged Index. Instead, pick based on the job Rockset actually did for you.

• ClickHouse Cloud is the default answer for dashboards, event analytics, and anything scan-heavy. It is faster than Rockset on large aggregations and dramatically cheaper at rest. The tradeoff is that you now think about schemas, ordering keys, and materialized views again.
• Tinybird feels closest to Rockset day to day. Ingest a stream, write SQL, publish it as a versioned HTTP endpoint. It runs on ClickHouse underneath, so you inherit that performance without operating it.
• StarRocks is the pick if your Rockset queries relied on real joins. Most columnar engines punish joins; StarRocks was designed for them, so you can stop denormalizing everything at ingest.
• Apache Pinot wins when thousands of end users hit analytics inside your product at once. Concurrency is its whole reason for existing.
• SingleStore covers teams who want transactional writes and analytical reads from one engine, which suits operational dashboards over live application state.
• Materialize is the streaming answer. You define a view once and it stays correct as new events land, instead of recomputing on every query.
• Elasticsearch still makes sense when text relevance matters more than aggregation speed, which was one third of what the Converged Index gave you.
• DuckDB deserves a mention because plenty of "real-time analytics" workloads were never big enough to need a cluster. Under roughly 100 GB, a single node will embarrass a distributed system on both cost and latency.

For deeper comparisons, see our ranked Rockset alternatives guide and the broader best tools like Rockset roundup.

Migration playbook: leaving Rockset without breaking production

1. Inventory your query patterns first, not your data. List every distinct query shape your app fires. That list, not your table sizes, decides which engine fits.
2. Split lookups from aggregations. Rockset hid this difference. Your new engine will not, so label each query before you migrate it.
3. Design one ordering key per table. In ClickHouse or StarRocks, the sort order is most of your performance. Pick it based on your most common filter.
4. Rebuild ingest as a stream, not a batch. Keep the freshness guarantee your users already expect, or they will notice the regression immediately.
5. Shadow-run for two weeks. Send production traffic to both systems, compare results row by row, and log latency percentiles rather than averages.
6. Move dashboards last. Internal reporting tolerates a day of drift. Customer-facing endpoints do not.
7. Set a cost alarm on day one. Usage-based analytics pricing surprises teams in month two, not month one.
8. Write down why you chose the engine. The next migration is always easier when the tradeoffs were recorded honestly.

Who actually used Rockset, and what they lost

Rockset never chased the general warehouse market. Its customers clustered into four recognisable shapes, and each felt the shutdown differently.

Product analytics teams embedded Rockset behind customer-facing dashboards. They cared about freshness and concurrency, and they lost the most, because their replacement had to match both at once.

Real-time personalisation and fraud teams used it as a feature store you could query with plain SQL. For them, latency budgets were measured in single-digit milliseconds, which pushed most toward Pinot or a purpose-built store.

Operational teams syncing DynamoDB or MongoDB used Rockset purely to make an operational database queryable without ETL. That group had the easiest exit, since change-data-capture pipelines into ClickHouse now cover the same ground.

Small engineering teams with no data platform used it because it needed no tuning. They lost the most convenience. There is no engine in 2026 that is simultaneously as fast, as schemaless, and as hands-off as Rockset was.

The cost question nobody enjoyed

Rockset was never cheap, and the Converged Index was the reason. Writing three index structures for every field meant storage was several times the raw data size, and ingest burned compute continuously rather than in scheduled bursts.

Teams typically paid somewhere between a few hundred and several thousand dollars a month for workloads that ClickHouse now serves for a fraction of that. The counter-argument was always the same, and it was a fair one: Rockset replaced two or three systems plus the engineer-months to run them. Compared against a warehouse, a cache, and a search cluster, the bill often looked reasonable.

That maths stopped working as ClickHouse Cloud, Tinybird, and StarRocks matured. By 2024, a competent team could get most of the benefit for far less, which quietly shrank the space Rockset had to itself. The acquisition ended the story before the market did, but the pressure was real.

What replaced the Converged Index, in practice

Nothing replaced it directly. Instead, teams rebuilt its three jobs with three deliberate choices:

• Point lookups moved to a well-chosen ordering key, a materialized view, or in some cases straight back to the operational database.
• Aggregations went to a columnar engine, where they were always going to be faster anyway.
• Selective filters and text search went to either a secondary index in ClickHouse or a dedicated search engine.

The honest summary is that Rockset traded money for the removal of a design decision. Once it was gone, teams had to make that decision themselves — and most discovered their workload was narrower than they assumed.

What the Rockset story tells you about buying infrastructure

Rockset was well funded, technically excellent, and run by people who had built this class of system before. It still disappeared with about a hundred days of notice, because a larger buyer wanted the team more than the market wanted the product.

The practical lesson is not "avoid startups." It is keep your exit cheap. Prefer engines with an open-source core or a compatible protocol, keep transformation logic in version control rather than in a vendor UI, and re-check every year whether you could actually move.

Related reading on this site: is Rockset dead, what happened to Rockset, and why Rockset failed. For the wider pattern of AI acquihires, see Multi, Adept AI, and the July 2026 trend report.

Frequently asked questions about Rockset

Is Rockset still available in 2026?

No. Rockset stopped serving customers on September 30, 2024, about 100 days after OpenAI announced the acquisition on June 21, 2024. There is no managed Rockset cloud, no new signups, and no self-hosted edition. The engine now lives inside OpenAI's internal retrieval infrastructure.

Why did OpenAI acquire Rockset?

OpenAI needed better retrieval. Rockset's team had spent eight years building low-latency indexing and search over rapidly changing data, which is exactly the problem OpenAI faces when grounding models in fresh enterprise content. The deal bought both the engineering team and a mature indexing engine rather than a customer base.

Who founded Rockset?

Rockset was founded in 2016 by Venkat Venkataramani and Dhruba Borthakur, both formerly of Facebook. Dhruba created RocksDB, the embedded key-value store that underpins a large share of modern databases, and that lineage shaped Rockset's storage design.

What was Rockset's Converged Index?

The Converged Index stored every field three ways at once: a row index, a column index, and an inverted search index. That removed the usual need to pick between point lookups, aggregations, and search, and it let teams query semi-structured JSON without defining a schema up front.

What is the best Rockset alternative?

For most analytics workloads, ClickHouse Cloud is the default replacement on raw speed and cost. Tinybird is the closest match if you valued Rockset's developer experience and instant API endpoints. StarRocks and Apache Pinot fit high-concurrency user-facing analytics, and SingleStore suits mixed transactional plus analytical workloads.

Did Rockset customers get refunds?

Yes. Rockset offered pro-rated credits and refunds for prepaid contracts and worked with enterprise customers on individual migration timelines through the September 30, 2024 end-of-service date. Self-serve users simply lost access when the service ended.

Can I still export data from Rockset?

No. Export tooling was available only during the 2024 migration window. Any collections not exported before September 30, 2024 are gone. If you still have an old export, it will be newline-delimited JSON that loads cleanly into ClickHouse, DuckDB, or BigQuery.

Is Rockset open source?

Rockset itself was never open source, though it was built on top of the open-source RocksDB storage engine. Because the service was closed and cloud-only, there is no community fork to run today. Teams wanting a similar open engine usually land on ClickHouse, StarRocks, or Apache Pinot.

Sources

• OpenAI — OpenAI acquires Rockset (June 21, 2024)
• RocksDB project site
• ClickHouse Cloud, Tinybird, StarRocks, Apache Pinot
• Internal: Rockset alternatives, best tools like Rockset, is Rockset dead