Released: 2026-11-19
Varnish version: 6.0.18r8
Artifact Firewall version: 0.11.0
cluster.peer_group URL that names an SRV record, such as http://_http._tcp.<release>-varnish-runtime-headless.<namespace>.svc.cluster.local, places each peer in the hash ring by its pod name instead of its IP address, so a restarted pod keeps the objects it owned. In a cluster already configured this way, the upgrade gives many objects a new owner once, and the new owner fetches them from the remote again.--schema, marks what a reload does with each setting: x-reload where it is stated, and x-reload-index on the root listing the settings a reload applies, the ones it refuses, and that anything else needs a restart.enable_firewall applies the firewall’s rules to crates. A denied or hidden version is marked yanked in the index, so cargo does not pick it, a lockfile that pins a denied version is refused at the download, and a crate denied outright cannot be resolved. A download is checked only when its URL names the crate and version, as crates.io’s and Artifactory’s do.not_found_ttl caches the 404s from its remotes for that many seconds, so a client that keeps asking for a missing tag or package is answered from the cache./artifactory/<repo>/ are classified as jfrog-generic, except in formats that use those paths, such as Maven and RPM. The firewall does not evaluate them, and remote_auth applies to them.jfrog-generic files are revalidated on every request, so an overwrite in Artifactory shows at once. other_ttl caches them longer.enable_firewall applies the firewall’s oci and docker rules to image pulls, by tag and by digest, including images already in the cache. Layers are not checked, so a client that already knows a layer’s digest can still download it.auth_policy: drop or remote_auth. A firewall set with firewall.remote_address must be Artifact Firewall 0.11.0 or later, or the client’s credential reaches the remote.403 Blocked by firewall instead of 403 Forbidden. Clients such as pnpm, Gradle and pip print only that line, not the body that names the rule, so their users took the refusal for a login problem.varnish.params to the running Varnish, without restarting it, so the cache is kept. If Varnish refuses a value, the reload is refused and nothing in it applies. A few parameters Varnish only takes when its cache process restarts, such as listen_depth, vsl_space and reuseport, still need a Supervisor restart, as do cli_timeout, shutdown_delay and http_max_hdr.varnish.http, otel, git_mirror or one of those parameters, is now refused, as a change to varnish.path already was. The refusal names the settings, and the Supervisor keeps warning that it runs a config other than the one on disk until it is restarted. Before, the reload reported the setting as applied while the running process kept the old value. A reload applies virtual_registry, varnish.vcls, varnish.params, license, cluster (except cluster.discovery), firewall, acme.domains, and supervisor.log_level, log_output and config_watch.enable_firewall on a registry) or ACME (the first acme.domains entry), when it was off at startup, reporting success while nothing started it: requests went to a firewall that was not running, and no certificate was obtained. Such a reload is now refused, as it needs a restart.error parsing REF pkt-line on a git index last fetched through the registry by git, within manifest_ttl or before the remote went down./sumdb/<name>/...) failing with 503 on a cache with no route to sum.golang.org, even through an Artifactory that serves them. They go to the registry’s remotes first, and only a 404 or 410 from those sends them to the database itself, now without the client’s Authorization header.sumdb segment in its path, which was refused with 404.