Skip to main content

v1.2.2

A maintenance release with fixes only. It stops writes being lost when a restart overlaps the old process, stops CPU climbing under heavy UI load, and stops high-churn fleets from wrapping entity keys or holding a startup heap sized by retention. It also corrects several checks, search queries and query results, and updates dependencies to clear known vulnerabilities. Read Considerations before upgrading: notification endpoints that redirect now fail.

Improvements​

  • Startup memory no longer grows with retention: The indexer's cache loaded every connection and subscription in retention to answer whether one was seen last epoch, and a Go map keeps its size after deletes, so the heap stayed at that startup peak. With high subscription churn, that peak was many times what was live. Connections, subscriptions, leaf nodes, routes and gateways now load only what each current server last reported.

  • Live means data is arriving: The toolbar read Live whenever the scraper process was up, so a scraper whose scrapes all failed looked healthy. Live now also requires a recent epoch, and the page switches to Stale on its own once epochs stop landing.

  • Parallel reads per system: Streamed queries on one system ran one after another, so a page's parallel panels finished in sequence. They now run concurrently up to db.query-stream.concurrency (default: the CPUs the process may use). checks.findings gets its own pool, db.checks.concurrency, so one slow check over a long history no longer holds every other check read on that system.

  • account-002 reads only the last hour: The check compared each server against its previous sample anywhere in retained history, so every epoch scanned the whole account history. It now uses the latest sample in the hour before the epoch. This also stops a server returning from a long absence from reporting all the slow consumers it accrued while away as new.

  • Dependencies: The embedded NATS server (the sink and the simulator) moves to nats-server 2.15.1, and the database engine to DuckDB 1.5.6. Go 1.26.11 carries the fixes for GO-2026-6809 (crypto/x509 trusting compromised intermediate certificates) and GO-2026-6859. golang.org/x/net 0.62.0 clears CVE-2026-78661, klauspost/compress 1.20.2 clears GO-2026-6865, brotli 1.2.7 clears GO-2026-6888, and the HTTP router update clears GO-2026-5774, GO-2026-5775 and GO-2026-5777.

Fixes​

  • Writes lost after a restart: A restart that overlapped the old process hit a lock conflict, deleted the old process's write-ahead log, and lost its unflushed writes once it exited. Insights now deletes the log only when DuckDB fails to replay it.

  • CPU climbing under heavy UI load: When a browser left before Insights read its page request, the page's stream was never cancelled and re-ran its queries on every epoch for as long as the process lived. On a busy instance these accumulated until they held most of the CPU. Fixed.

  • Stats attached to the wrong entity on high-churn fleets: Subscription, connection, consumer and user keys grow with every new entity, so a fleet with heavy subscription churn could pass the 32-bit limit. The key was stored wrapped to a negative number without an error, and after a restart new keys collided with those rows. These keys are now 64-bit.

  • Route and gateway counters split across identities: A route's or gateway's start time compared unequal to itself after a restart, a reconnect, or on every epoch for a server in a timezone offset that is not a whole hour. Each miss recorded the unchanged connection as a new one, splitting its counters. Fixed.

  • Uptime and idle of a day or more stored as zero: nats-server reports long durations with day and year units, which were not parsed, so connection, route and gateway uptime and idle past a day read as 0. connection-004 can now fire for long-idle connections.

  • Meta leader checks reporting events that never happened: When the meta leader's name could not be resolved, a follower's row read as that follower claiming leadership. system-004 counted leader changes that did not occur, system-011 reported split leadership, and system-007 and server-058 could show an offline follower as an online leader. Fixed.

  • user-001 never fired: Insights always recorded users as non-bearer, and nats-server 2.11.17, 2.12.8 and 2.14.0 onward also withhold a bearer user's JWT. Insights now infers bearer users from what the server does report, and corrects existing rows on upgrade. The check's description states the cases it still cannot tell apart.

  • Refused logins counted twice: From nats-server 2.12 a client refused at login is published both as an auth error and as a disconnect, and a client whose credentials a reload revokes is published twice on every version, so the realtime events chart showed each failure twice. Each now counts once.

  • An nkey seed without a JWT was ignored: The connection went out with no credentials, so the scraper failed to authenticate or landed in the wrong account. A seed on its own now authenticates as an nkey user.

  • Chart gaps over complete data: Charts placed epochs on a grid at exactly the scrape interval, and epochs land a few seconds over it, so every few hours a slot was empty and a gap appeared. Gaps are now drawn only where an epoch is actually missing.

  • Search matching the wrong things: kvstore@name: and objectstore@name: matched the backing stream's name rather than the store's, so the name the page shows found nothing. A query such as name:x OR stream@subjects:y listed every server, account and connection. A qualified term naming an attribute the entity lacks matched every entity of that kind; it is now an error.

  • Query results showing raw storage: A DECIMAL in $INS.db.query JSON reached MCP and HTTP clients as an object instead of a number string. In streamed CLI output, HUGEINT, BOOLEAN and UUID values printed as base64 or 0/1 and small decimals printed in exponent form, and the web UI read a large UHUGEINT as negative. Fixed, including inside LIST and STRUCT values.

  • Client commands failing in the container image: Commands that resolve a NATS context failed with user: Current requires cgo or $USER set in environment when run as the image's non-root user. The image now sets XDG_CONFIG_HOME, and mounted contexts go under /tmp/nats/context/.

Considerations​

  • Notification endpoints must not redirect: Delivery no longer follows a 3xx, because Go forwards an endpoint's credentials, and its configured URL as the Referer, to wherever the redirect points. An endpoint whose url redirects, such as http:// to https:// or a missing trailing slash, now fails every delivery with the redirect's status and target host in the delivery log. Set url to the final address.

  • Check reads have a deadline: checks.findings is now bounded by db.query-timeout (default 2m). A read that runs longer fails with a timeout error instead of running on after its client has gone.

  • New findings after upgrade: user-001 and connection-004 can now fire where they could not before, so expect new findings, and notifications for endpoints subscribed to them, on the first epochs after upgrade.

  • Automatic migrations: On first start existing user rows are rechecked for bearer users, and the subscription, connection, consumer and user key columns, with the columns that reference them, are widened to 64-bit. Both run without intervention. Widening rewrites those columns, so the first start takes longer in proportion to the stored subscription, connection, consumer and user history: unnoticeable on a small database, minutes on a large one. Health checks fail until it finishes, so allow for it in any startup probe.