Upgrading to Varnish 7.5

Logs#

The optional reason field of BackendClose records changed from a description (for example “Receive timeout”) to a reason tag (for example RX_TIMEOUT). This will break VSL queries based on the reason field.

Using a tag should however make queries more reliable:

# before
varnishlog -q 'BackendClose ~ "Receive timeout"'

# after
varnishlog -q 'BackendClose[4] eq RX_TIMEOUT'

Such queries would no longer break when a description is changed.

Timeouts#

The value zero for timeouts could mean either timing out immediately, never timing out, or not having a value and falling back to another. The semantic changed so zero always means non-blocking, timing out either immediately after checking whether progress is possible, or after a millisecond in parts where this can’t practically be done in a non-blocking fashion.

In order to disable a timeout, that is to say never time out, the value INFINITY is used (or tested with isinf()).

In the places where a timeout setting may fall back to another value, like VCL variables falling back to parameters, NAN is used to convey the lack of timeout setting.

VCL#

All the timeouts backed by parameters can now be unset, meaning that setting the value zero no longer falls back to parameters.

Parameters#

All the timeout parameters that can be disabled are now documented with the timeout flag, meaning that they can take the special value never for that purpose:

varnishadm param.set pipe_timeout never

The parameters idle_send_timeout and timeout_idle are now limited to a maximum of 1 hour.

VRT#

For VMOD authors, it means that the vtim_dur type expects INFINITY to never time out or NAN to not set a timeout.

For VMOD authors managing backends, there is one exception where a timeout fallback did not change from zero to NAN in struct vrt_backend. The vtim_dur fields must take a negative value in order to fall back to the respective parameters due to a limitation in how VCL is compiled.

As a convenience, a new macro VRT_BACKEND_INIT() behaves like INIT_OBJ but initializes timeouts to a negative value.

VCL COLD events have been fixed for directors vs. VMODs: VDI COLD now comes before VMOD COLD.

Other changes relevant for VMOD / VEXT authors#

Transports are now responsible for calling VDP_Close() in all cases.

Storage engines are now responsible for deciding which fetch_chunksize to use. When Varnish-Cache does not know the expected object size, it calls the objgetspace stevedore function with a zero sz argument.

(struct req).filter_list has been renamed to vdp_filter_list.

eof