Postgres 17.4 available
Projects can upgrade to Postgres 17.4 from the dashboard; the previous minor version stays available for 90 days.
Downstream product
One normalized stream, assembled from every collector the control plane watches. Served from the last healthy snapshot of each source, so it keeps returning correct data even while a collector is mid-incident. /api/feed serves the same thing as JSON.
Projects can upgrade to Postgres 17.4 from the dashboard; the previous minor version stays available for 90 days.
Read replica connection strings now respect the project's IP allowlist instead of accepting traffic from anywhere.
Log timestamps in the dashboard now render in the project's configured timezone instead of always UTC.
`nimbus logs --follow` now streams over a persistent connection instead of polling every two seconds. Reconnects are transparent and preserve cursor position, so you no longer lose lines during a deploy.
Suspended computes now resume in under a second for most branch sizes, down from a typical two-to-three second wake.
The dashboard now surfaces replication slot lag per subscriber, so a stuck consumer is visible before it grows the WAL.
A visual editor for RLS policies in the dashboard, with a preview of which rows a policy actually admits.
New projects can opt into Postgres 17 at creation time; existing projects upgrade via the usual maintenance window.
The v1 storage endpoints stop accepting new integrations next quarter; existing usage keeps working through the sunset window.
Compute autoscaling now scales up to 16 compute units on Scale-plan projects, up from 8, for bursty analytical workloads.
`neonctl branches reset` discards a branch's changes and re-forks it from its parent's current head in one step.
Read replicas move out of beta and are available in every supported region for Team and Enterprise plans.
Tokens issued before a workspace rename could fail to refresh, producing a confusing 401 several hours into a session. The refresh path now resolves the workspace by id rather than by slug.
Reconnecting after a dropped websocket could leave the old channel subscribed, doubling event delivery. Channels now close cleanly first.
A shared runtime pool keeps warm instances across invocations, cutting p95 cold-start latency roughly in half.
Spinner frames and ANSI codes leaked into piped output in some shells. All writers now check both NO_COLOR and TTY status before emitting escape sequences.
Pro-tier projects can now restore to any second within the retention window, not just daily snapshots.
Variables can be declared once at the workspace level and inherited by every project inside it. Project-level declarations still win, so existing setups are unaffected.
The config loader no longer walks the entire directory tree looking for a lockfile. It stops at the first workspace boundary, which removes most of the startup cost in large monorepos.
`nimbus push` has printed a deprecation notice since v2.4 and is now gone. Use `nimbus deploy`, which accepts the same flags.
Attempting to deploy into a region that is at capacity returned a bare 503. The CLI now names the region, lists the nearest alternatives, and exits with code 75.
The manifest is validated locally against the published schema before anything is sent, so typos in a service name fail in under a second instead of thirty.
Path separators were normalised inconsistently, producing different archive checksums for identical trees across platforms. Builds are now byte-identical regardless of host OS.
Layers upload concurrently with a default fan-out of four, tunable via `--upload-concurrency`. Large images see roughly a 3x improvement on typical connections.
HTTPS_PROXY and NO_PROXY were read only by the auth client, not by the upload client. Both now share a single transport configuration.
A single command that checks CLI version, credential validity, clock skew, and network reachability, then prints a copy-pasteable summary for bug reports.