If you watch long enough, the story repeats almost step by step:
The licence is open, the project wins developers, then a cloud provider (or a fast competitor) sells it as a service and keeps most of the money. The company answers with a licence that blocks this. The community forks the last open version. And sometimes, years later, the company adds an open-source licence back.
Here are the dated events I mean, as of November 2026:
| When | Project | What changed |
|---|
| Oct 2018 | MongoDB | Community Server moved from AGPLv3 to the SSPL, written to cover offering MongoDB as a service (MongoDB) |
| Jan 2021 | Elasticsearch, Kibana | Apache-2.0 → dual licence, SSPL or the Elastic License, announced in January (Elastic); AWS forked the last Apache version as OpenSearch (AWS) |
| Mar 2024 | Redis | BSD-3 → RSALv2 or SSPL from 7.4 (Redis); the Linux Foundation started the Valkey fork (Linux Foundation) |
| 2024 | ArangoDB 3.12 | Source under BSL 1.1; prebuilt binaries under the ArangoDB Community License, which caps production data at 100 GiB and says "you cannot use it for any commercial purposes, only internal business purposes" (ArangoDB docs) |
| Aug 2024 | Elasticsearch, Kibana | AGPL added as a third licence option (Elastic) |
| Aug 2024 (effective Nov 2024) | CockroachDB | Free Core edition retired with 24.3 (announced in August, effective in November); one Enterprise licence, free for companies under $10M annual revenue (Cockroach Labs) |
| May 2025 | Redis 8 | AGPLv3 added next to RSALv2 and SSPL (Redis) |
| Oct 2026 | Pinecone | Pinecone Local, the Docker emulator, reached end of life; the product is managed-only (Pinecone docs) |
Graph databases have their own version of this. Memgraph, SurrealDB and ArangoDB are under the BSL, FalkorDB is SSPL, and Neo4j keeps closed-source components in a commercial Enterprise edition.
I want to be fair here, because the pressure behind these moves is not imaginary.
Building a database is the hard path. It takes years of engineering before anybody pays, and the people who pay are usually large companies that want support, SLAs and features for scale. Meanwhile a hyperscaler can take your open-source engine, put it behind their console and their bill, and sell it to your future customers. You did the R&D. They own the distribution. That's a real problem, and a licence that says "you can't resell this as a service" is the cheapest tool in the drawer.
So when Mongo, Elastic or Redis changed licences, I don't think it was greed. They had investors, payroll and a competitor who didn't pay for the engine. That's fair anyway.
The licence became the business moat because the business moat wasn't anywhere else. If the only thing between you and a competitor reselling your database is a legal clause, then the product on top of the engine is not doing its job. The clause holds back the competitor, but it doesn't make your own service better.
And the market answered pretty clearly. AWS forked Elasticsearch into OpenSearch. The Linux Foundation forked Redis into Valkey within weeks, with cloud vendors behind it. Then both Elastic and Redis added the AGPL back. Relicensing didn't remove the competitor. It created a second project with the competitor's money behind it.
There's another cost that doesn't show up on a revenue chart, and it's trust. Every team that lived through a licence change now reads the LICENSE file before the README. They ask what happens in three years, not what the database does today. For a database, where migrating away is the most expensive thing a team can do, credibility and trust are the product. Once you spend them, they're hard to buy back.
This part I feel directly. Right now I'm running a benchmark suite that compares fifteen databases on one machine, all pinned to exact image digests, all with the same CPU and memory limits. It's still cooking; when it's done I'll publish the suite and the results, so anybody can re-run every number. Cloud-only products can't be part of that. There's no version to pin, no hardware budget to match, and a network round-trip in every query. When Pinecone retired its local emulator, the question of how to include it on equal terms literally had no answer.
Developers hit the same wall in smaller ways. You can't run the database in CI without an account. You can't read the code when a query behaves strangely. You can't keep a copy that works if the vendor changes plans. And for AI agents it gets worse, IMO. An agent that works with your data should be able to run next to it, on your machine or in your VPC, not only through somebody's API with somebody's rate limits.
More closed boxes -> more plumbing around them -> more entropy for the team that has to glue it all together.
To be honest, RushDB was part of this trend for a while. The 2.x platform core and dashboard were under the Elastic License 2.0: source available, but not open source. When we rebuilt RushDB as a standalone engine in Rust, I had to decide again. RushDB 3 is Apache-2.0, the whole workspace, one binary you can run anywhere. You can read the full story of the engine in the 3.0 release notes and why we left Neo4j in time to move on.
So where is the business? In the harness, not in the licence. RushDB Cloud runs the same engine and adds what is actually hard to operate yourself: managed embeddings, natural-language search, relationship suggestions, a hosted MCP server with OAuth, and team access. If somebody wants to self-host the engine and never pay us, that's fine. If somebody wants to host it for others, they'll have to build all of that too, or compete on something else.
And there's a product reason, not just a business one. RushDB is a database for AI, and its whole idea is agnosticism: any data becomes AI-ready, whatever shape it has and wherever it comes from. A database that wants to take any data can't be picky about where it runs. A closed, cloud-only engine would contradict the pitch.
Maybe it is, because a hyperscaler could package RushDB tomorrow and the licence wouldn't stop them. I've thought about it, and my answer is kinda simple: if a cloud provider decides RushDB is worth reselling, the engine has already won, and our job is to make RushDB Cloud the better place to run it. I'd rather compete on the service than on a clause.
What I don't want is to build trust for five years and then spend it in one licence change. So the plan is boring on purpose:
- The engine stays Apache-2.0.
- The money comes from RushDB Cloud and support, see pricing.
- Benchmarks get published with pinned versions, the method and re-run commands for every number.
Other companies will keep making the opposite call, and they'll have good reasons. I just think the databases people trust in ten years will be the ones they could always read, run and leave.