Supabase's New Storage Engine Is a Dropdown You Cannot Un-Click
The pitch is a faster Postgres. The dropdown does not mention that the choice is permanent, that Point-in-Time Recovery is off the table, and that your pgvector index now runs through an experimental bridge.
A storage engine went to public beta. That is not the news.
On 1 October Supabase moved OrioleDB, an alternative storage engine for Postgres, from alpha to public beta. The changelog entry is short and almost entirely good news: you can create OrioleDB projects on the Pro, Team and Enterprise plans, not only Free. OrioleDB projects now support compute and disk resizing, pause and restore, Postgres version upgrades, read replicas, and scheduled backups.
Read that list again and notice what it is: a list of things that used to be missing and are now present. It is a maturity announcement, and it is a real one.
The part that matters to anyone about to create a Supabase project sits in the documentation rather than the changelog. Supabase puts it in two sentences: you choose OrioleDB when you create a project, and you cannot add OrioleDB to an existing project or remove it later.
That is not a setting. That is a one-way door, and it opens at the moment you know least about the app you are about to build.
What this class of claim always leaves out
Performance pitches for infrastructure are made about the steady state. Some engine is faster, or scales better, or handles a contention pattern the default one handles badly. Those claims are usually true, usually narrow, and usually verifiable.
What they describe is the good day. The questions that decide whether a choice was right are all about the bad day: the migration that dropped the wrong column, the agent that rewrote a policy at 2am, the index that turned out to matter more than you thought. Almost no product page is organised around the bad day, because the bad day does not sell.
So when something is offered as a performance upgrade, the useful move is not to argue with the benchmark. It is to ask two things: what the bad-day options look like afterwards, and whether you can change your mind. For OrioleDB in public beta, as of 1 October 2026, both have clear answers.
The two gaps that matter
Point-in-Time Recovery is not there. Supabase states it directly: Point-in-Time Recovery is not available for OrioleDB projects. Scheduled backups are available, and the two get conflated constantly, so it is worth separating them. A scheduled backup returns you to a snapshot boundary. Point-in-Time Recovery returns you to a timestamp you choose — the second before the bad thing happened.
If your schema is stable and changes monthly, the gap between those is an inconvenience. If your schema is being written and rewritten by an AI agent several times a week, which is a common condition of a Lovable app in its first ninety days, the gap between those is the difference between losing four minutes and losing a day.
Your vector index is on the experimental path. Supabase is specific here too: native B-tree indexes give the best performance, and other index types, including pgvector HNSW indexes, run through an experimental bridge.
If your product has semantic search, recommendations or any retrieval feature, it sits on a vector index. Which means the one index your differentiating feature depends on is the one running through a bridge described as experimental, on an engine described as public beta.
The same question, asked of both engines
| The question you ask on a bad day | Standard Postgres project | OrioleDB project (public beta, 1 Oct 2026) |
|---|---|---|
| Can I roll back to the second before a bad migration? | Point-in-Time Recovery | Not available |
| Can I restore from a backup? | Scheduled backups | Scheduled backups |
| Can I resize compute and disk? | Yes | Yes |
| Can I add read replicas? | Yes | Yes |
| Can I run a Postgres version upgrade? | Yes | Yes |
| Is my pgvector HNSW index on a supported path? | Native | Experimental bridge |
| Can I change my mind in month three? | Not applicable | No — chosen at creation, cannot be added or removed |
Four questions to answer before you click create
One. Does this project hold data you cannot reconstruct? User accounts, payments, anything a customer typed. If yes, the absence of Point-in-Time Recovery is a standing cost for the life of the project, not a beta caveat that expires.
Two. Will an agent be writing your migrations? If a tool is generating schema changes faster than you review them, you depend on fine-grained rollback more than most teams, whether or not you have needed it yet.
Three. Does a core feature depend on a vector index? If yes, you are putting your most important query path on the experimental bridge.
Four. Can you name the contention problem OrioleDB solves for you? Not in general — for this app, at the load you actually have. If you cannot name it, you are accepting a permanent cost for a benefit you have not measured.
A no to question three and a confident yes to question four is a reasonable case for trying it. Anything else, and the default heap engine is the cheaper decision, because it is the reversible one.
If you already created one
There is no toggle. Because the engine cannot be removed, changing your mind means standing up a second Supabase project on the default engine and moving the data into it: a dump and restore, plus re-pointing every key, policy, webhook and environment variable your app and your host are holding. That is a careful day on a small app and considerably more on a live one — and it is work that would have cost nothing if the dropdown had been read for ten seconds.
This is the general shape of it. The cheapest moment to get an infrastructure decision right is before the thing exists, and the bill for getting it wrong arrives in a month when you are busy with something else.
What we do with this
We read these changelogs so our clients do not have to, and the four questions above get asked before a project is created rather than after. IdeaFlo has completed 80 jobs on Upwork, and the recurring pattern is not that founders choose wrong — it is that nobody told them a choice was permanent.
If you want a second pair of eyes before you commit, the Strategy Audit is $499, one-time, and credited toward any build.