Changelog

Pick how idle branches wake: stop to zero (1.1 to 1.4 s median in our tests), a warm pool (about 600 ms) or paused in memory (about 9 ms). Also new: audit log export and self-serve account deletion.

Self-hosted installs need a signed license key, verified offline in your cluster. It is free for organizations under $1M in annual revenue. If a key lapses, running databases keep serving and new branches are refused until you install a valid key.

A test runs a random workload and compares the pages SkelaDB serves, byte for byte, with the same pages from a standard PostgreSQL standby. We run it while developing each release; it does not check pages inside your installation.

Large databases can spread across storage nodes, and a busy shard splits on its own, so storage throughput grows with the data. Tested on a local Kubernetes cluster.

History is compacted and cleaned up automatically once it passes your retention window.

Add read replicas that stream from the same replicated log, without copying the database.

Fork a branch from any past moment within your retention window, down to the second.

Every commit is written to three WAL nodes and acknowledged once two have it on disk. Losing one node loses no acknowledged commit.

Database history is stored as immutable layers in your object storage bucket, with a local cache in front for fast reads.

Create an isolated, writable copy of any database. In the object storage mode, branches share storage with their parent until they change.

In the object storage mode, Postgres computes read every page from the storage layer, so a compute can start anywhere and hold no data of its own.

Projects, branches, scale to zero with wake on connect, backups with point-in-time restore, retention, Prometheus metrics and an installer with profiles for EKS, GKE and AKS. Those profiles are still in preview and not yet verified on the managed services.

PostgreSQL 18.4 with three small patches and the SkelaDB storage extension, in one image for primaries, replicas and recovery.

TLS, SCRAM-SHA-256 pass-through, transaction pooling, prepared statements, cancel requests and routing by branch. Clients connect to one address and the proxy wakes the right compute.