Files
tailscale/ipn
David Bond c8ba9da543 ipn/store/kubestore: cache TLS certs briefly on write replicas (#22103)
The write replica of an HA ProxyGroup (cert share "rw") doesn't keep TLS
certs in memory, so that it notices a recreated, empty cert Secret and
issues a new cert. As a result, it read the cert Secret from the API
server on every TLS handshake, with a 30s timeout. When the API server
was slow or unavailable, handshakes on that replica waited for the read,
and feature/acme repeated the read under the per-domain lock. Clients
with a 10s handshake timeout, such as kubectl, failed.

This commit adds a small per-domain cache of the last-known-good cert
and key for write replicas. The replica serves a cached cert for 30s
before it reads the Secret again. When it has a cached cert, the read
uses a 3s timeout, and on an error other than not found or forbidden
the replica serves the cached cert and waits another 30s before the
next read. A read that finds the Secret missing, forbidden or empty
removes the cached cert and returns ipn.ErrStateNotExist as before, so
a recreated Ingress still gets a new cert, at most 30s later.
Concurrent reads of the same Secret share one GET, and a read that is
in flight when a cert is written does not replace the written cert.

A Secret watch remains the long-term fix.

Fixes #22009

Change-Id: Id107bf0a7628a1a026f94e621173d7c5d02fb451

Signed-off-by: David Bond <davidsbond93@gmail.com>
2026-11-20 12:32:02 +00:00
..
2026-11-14 09:50:33 -08:00