Repository navigation
SECURITY
- Go 1.27.3 build.
Go released a critical security fix, so the build now requires Go 1.27.3
via GOTOOLCHAIN, and the Docker build uses the same version. The linter was
updated to work with it.
RELIABILITY FIXES
-
A dial timeout no longer drops a healthy agent (#924).
Before, if sending a dial request to an agent timed out, the server retired
the whole agent connection and cut every other tunnel using it. Now only
that one dial fails. Waiting for the send lock can also be cancelled, so
expired requests are never sent. -
Client Close() no longer blocks.
conn.Close() used to wait up to 10 seconds for the server to confirm the
close. A webhook call with a 10s timeout could therefore take 20s and hold
an apiserver goroutine the whole time. Close() now returns right away and
waits for the confirmation in the background. -
The agent no longer closes the wrong connection.
When a server connection died, two cleanup paths could each try to remove
it. If the agent reconnected in between, the second removal killed the new,
healthy connection. Removal now checks that it is removing the same client. -
Fixed an agent crash on auth setup failure (#927).
If the auth context failed to initialize, the agent could crash with a nil
pointer error and lose the real error. Both problems are fixed. -
Pending-dial counts stay accurate.
Some cleanup paths skipped updating the pending_backend_dials gauge or the
per-agent count of dials in progress. Every path now updates them.
FEATURES AND TUNING
-
The server prefers faster agents for new dials.
The server used to send every agent an equal share of new dials, even one
too short on CPU to dial quickly. It now estimates how long each agent will
take to dial and avoids clearly slow ones. In testing, mean dial time fell
from 31ms to 5ms. -
New --keepalive-timeout flag on server and agent.
You can now set how long each side waits for a keepalive ping reply before
closing the connection (default 20s, same as before). With short keepalive
values, dead connections behind load balancers or NAT are found in seconds
instead of minutes. -
Fewer wasted reconnects with --sync-forever.
With --count-server-leases, the agent stops dialing once it is connected to
every server. Before, it opened a new connection every sync interval,
causing churn on the load balancer and NAT and noise in the logs. -
Forward HTTP proxying.
The HTTP frontend now forwards plain http:// requests like GET and POST
through agent tunnels, not just CONNECT requests. The
--backend-dial-timeout help text now covers these dials.
OBSERVABILITY
- New metrics for connection loss.
Agent: server_connection_lost_total (by reason) and
server_connection_attempts_total (by result).
Server: established_connections_closed_total (by reason) and
backend_connection_duration_seconds.
These record short outages that a gauge can miss between scrapes.
TOOLING AND DOCS
- Troubleshooting toolkit.
examples/kubernetes-troubleshooting has scripts and write-ups for
reproducing proxy failures in a kind cluster, such as connection resets,
CPU-starved agents and NAT behaviour. The tools tunnel-probe, webhook-load
and test-webhook now live under cmd/.