Supabase says it examined the software the Postgres community uses to run clusters of database servers and found that, under certain conditions, every one of these tools could lose data. Its answer is Multigres, a ground-up rewrite that the company says came through 1,000 failovers in a row without losing any data. Supabase also released a faster storage engine for Postgres, OrioleDB, as a public beta.
Supabase hosts Postgres, a widely used open-source database, for developers and for the growing number of apps built with AI coding tools. Co-founders Paul Copplestone and Ant Wilson discussed the launches with swyx on Latent Space, in an episode published October 28, 2026 and recorded the day before the releases. Supabase described three database efforts: performance, scaling up and scaling out. OrioleDB covers performance, and Multigres covers scaling up. Scaling out, which means running very large numbers of small databases, is a separate project.
What an operator does, and why Supabase rewrote one
A Postgres database usually starts life on one server, or node. Serious deployments soon need several nodes: copies of the data on standby, called replicas, so that if the main server dies another can take over. That handover is called a failover. The cluster also needs regular backups. The software that runs all of this is called an operator.
The Postgres ecosystem has several operators. Supabase said it tested them and found that each one could lose data in some situation, and that this finding led it to rewrite the whole structure of an operator instead of adopting an existing one.
The Multigres developers have published one example of how such a loss can happen. In a May 28, 2026 demonstration, Rafael Chacon starts with a primary server and two replicas, then adds a third replica. During that change, the processes coordinating the cluster can disagree about which machines belong to it. If the database confirms a write to an application by counting the new replica, and a later failover relies on an older view of the membership, that confirmed write can vanish. Multigres treats changes to membership and durability rules as decisions the cluster must agree on: a new replica cannot count toward confirming writes or choosing a new primary until the cluster has durably agreed it is a member.
From a sharding project to an operating system for the cluster
Multigres began with a narrower goal: to give the Postgres community sharding. Sharding splits one large database across several servers, each holding part of the data, so a workload can grow past what one machine can handle. Supabase said the project grew into something closer to an "entire operating system" for the cluster.
Besides the 1,000 sequential failovers, Supabase said it had attacked the system in every way it could think of to see whether it would ever drop data, and that it was now confident in its stability. The company said Multigres would be completely open source and available on its own platform by invitation, so that customers with serious workloads could get help scaling up.
Supabase's hosted Multigres documentation describes automatic failover, connection strings that stay the same, and writes that are confirmed only after the cluster agrees they are durable. It also sets out the limits. Hosted access is a private alpha for paid organizations, with no uptime guarantee, and it is not aimed at production workloads. The alpha does not support Realtime, point-in-time recovery, cross-region replicas, resizing after creation, or running OrioleDB at the same time, though daily backups remain. There is no managed path back from Multigres to a standard Postgres project. Sharding is still listed as a roadmap item.
The lesson from Vitess
Multigres invites comparison with Vitess, which does a similar job for MySQL, another popular open-source database. According to the Vitess project history, it was created at YouTube in 2010. Read replicas eased the load on YouTube's MySQL databases at first, but write volume eventually required sharding. YouTube initially wrote the logic for choosing a shard into its application code; Vitess moved that routing into a proxy, a layer that sits between the application and the databases.
Supabase said Vitess users always had to change some of their queries when they moved to sharding, and that Vitess was always catching up on compatibility. When Vitess's original creator came to work on Multigres, the company said, he made compatibility the first goal. Supabase said the rewrite is designed so that once sharding arrives, customers keep running the same queries, with "zero changes to your code."
That goal is not yet fully met at the proxy layer. In a June 4, 2026 post on compatibility, Haritabh Gupta describes running Postgres's own regression test suite through the Multigres proxy. The table shown there, labeled as live nightly data, reports 164 of 222 tests passing. For comparison, it gives snapshots of PgBouncer, a widely used connection pooler: 155 of 222 in transaction mode and 221 of 222 in session mode. Session mode scores higher because it keeps each client tied to its own backend connection, which gives up the connection sharing that makes pooling useful. Remaining Multigres gaps include some catalog views and asynchronous-message behavior. Multigres supports session settings, SQL prepared statements, temporary tables and LISTEN/NOTIFY, but it blocks some operations, including creating databases and logical-replication subscriptions.
Sharding is still to come
When the conversation credited Supabase with building Multigres in two years, the reply was a correction: the work had taken one year, and sharding had not been released. Early prototypes and designs exist, the answer went on, and some of the sharding work should land "probably early next year," depending on what else was happening on Supabase's platform.
Ant Wilson said he hoped Multigres would not be only for Supabase customers. Anyone "running Postgres at scale," he said, should be able to use it. Supabase said it runs a design partner program with organizations including Booz Allen, Postman and Perplexity, which need help scaling Postgres.
Later in the episode, Supabase put Multigres in the context of its business. Large companies, it said, almost never move existing production databases. Multigres gives Supabase a chance to win those large workloads if they ever migrate, but the company said it would not be a core focus for about the next five years. Its enterprise effort will concentrate on new projects inside large companies.
OrioleDB: a different way to store rows
The performance effort is OrioleDB. A storage engine is the part of a database that decides how rows are laid out and kept on disk. Postgres's standard engine is called heap. When a row in a heap table is updated, the old version stays behind as a dead row until a background process called vacuum clears it, and tables can swell with this leftover data, a problem known as bloat.
Supabase called OrioleDB a drop-in replacement for heap. It said OrioleDB runs about 1.5 to 2 times faster with nothing else to change, and removes the need for vacuum and the bloat it deals with. According to Supabase's beta announcement, the engine uses tables organized around their index and an undo log, which reduce the cleanup of dead rows.
Supabase said it spent the past year testing OrioleDB: running it through pg_regress, Postgres's regression test suite, for full compatibility, and through Antithesis, nightly benchmarks modeled on TPC-C, an industry benchmark for transaction processing, and Jepsen, a set of tests that look for data loss. The aim, the company said, was to make sure it could never lose data.
The announcement's benchmark figures show where the speed-up is largest. The tests, run by DBArena on four hosted configurations across seven compute sizes, used a workload derived from TPC-C that does not meet the official TPC-C requirements. When the data was four times larger than memory, OrioleDB delivered more transactions per dollar at every size. On the largest machine tested, with 32 virtual CPUs and 128 GiB of memory, it reached 113,850 transactions a minute, about 1.8 times heap's throughput; on the smallest, its throughput per dollar was about 46% higher. When the data fit in memory, OrioleDB was about 13% ahead on the Large size and 29% ahead on the largest, but heap won on the Small size. Supabase says the gains are concentrated in update-heavy workloads and those built around primary keys.
Supabase said it acquired the OrioleDB technology, kept developing it, paid for a patent, and is now giving it away under the Apache 2 license. The OrioleDB repository lets users choose between the Apache License 2.0 and the PostgreSQL License, and points to a separate patent grant. That grant gives worldwide, royalty-free rights to use, modify and distribute the implementations, but excludes patent claims infringed only because of further modifications, and it ends for anyone who brings the specified kind of patent action against the code.
The repository describes the public beta as suitable for experiments, testing and benchmarking, not production. On Supabase's platform, the beta has no service-level guarantee or point-in-time recovery, uses an experimental bridge for indexes other than B-trees, and must be chosen when a project is created.