DOKK / manpages / debian 12 / varnish / varnish-cli.7.en
VARNISH-CLI(7) VARNISH-CLI(7)

varnish-cli - Varnish Command Line Interface

Varnish has a command line interface (CLI) which can control and change most of the operational parameters and the configuration of Varnish, without interrupting the running service.

The CLI can be used for the following tasks:

You can upload, change and delete VCL files from the CLI.
You can inspect and change the various parameters Varnish has available through the CLI. The individual parameters are documented in the varnishd(1) man page.
Bans are filters that are applied to keep Varnish from serving stale content. When you issue a ban Varnish will not serve any banned object from cache, but rather re-fetch it from its backend servers.
You can stop and start the cache (child) process though the CLI. You can also retrieve the latest stack trace if the child process has crashed.

If you invoke varnishd(1) with -T, -M or -d the CLI will be available. In debug mode (-d) the CLI will be in the foreground, with -T you can connect to it with varnishadm or telnet and with -M varnishd will connect back to a listening service pushing the CLI to that service. Please see varnishd(1) for details.

The Varnish CLI is similar to another command line interface, the Bourne Shell. Commands are usually terminated with a newline, and they may take arguments. The command and its arguments are tokenized before parsing, and as such arguments containing spaces must be enclosed in double quotes.

It means that command parsing of

help banner


is equivalent to

"help" banner


because the double quotes only indicate the boundaries of the help token.

Within double quotes you can escape characters with \ (backslash). The \n, \r, and \t get translated to newlines, carriage returns, an tabs. Double quotes and backslashes themselves can be escaped with \" and \\ respectively.

To enter characters in octals use the \nnn syntax. Hexadecimals can be entered with the \xnn syntax.

Commands may not end with a newline when a shell-style here document (here-document or heredoc) is used. The format of a here document is:

<< word

here document word


word can be any continuous string chosen to make sure it doesn't appear naturally in the following here document. Traditionally EOF or END is used.

Integrating with the Varnish CLI can be sometimes surprising when quoting is involved. For instance in Bourne Shell the delimiter used with here documents may or may not be separated by spaces from the << token:

cat <<EOF
hello
world
EOF
hello
world


With the Varnish CLI, the << and EOF tokens must be separated by at least one blank:

vcl.inline boot <<EOF
106 258
Message from VCC-compiler:
VCL version declaration missing
Update your VCL to Version 4 syntax, and add

vcl 4.0; on the first line of the VCL files. ('<vcl.inline>' Line 1 Pos 1) <<EOF ##--- Running VCC-compiler failed, exited with 2 VCL compilation failed


With the missing space, the here document can be added and the actual VCL can be loaded:

vcl.inline test << EOF
vcl 4.0;
backend be {

.host = "localhost"; } EOF 200 14 VCL compiled.


A big difference with a shell here document is the handling of the << token. Just like command names can be quoted, the here document token keeps its meaning, even quoted:

vcl.inline test "<<" EOF
vcl 4.0;
backend be {

.host = "localhost"; } EOF 200 14 VCL compiled.


When using a front-end to the Varnish-CLI like varnishadm, one must take into account the double expansion happening. First in the shell launching the varnishadm command and then in the Varnish CLI itself. When a command's parameter require spaces, you need to ensure that the Varnish CLI will see the double quotes:

varnishadm param.set cc_command '"my alternate cc command"'
Change will take effect when VCL script is reloaded


Otherwise if you don't quote the quotes, you may get a seemingly unrelated error message:

varnishadm param.set cc_command "my alternate cc command"
Unknown request.
Type 'help' for more info.
Too many parameters
Command failed with error code 105


If you are quoting with a here document, you must wrap it inside a shell multi-line argument:

varnishadm vcl.inline test '<< EOF
vcl 4.0;
backend be {

.host = "localhost"; } EOF' VCL compiled.


Another difference with a shell here document is that only one here document can be used on a single command line. For example, it is possible to do this in a shell script:

#!/bin/sh
cat << EOF1 ; cat << EOF2
hello
EOF1
world
EOF2


The expected output is:

hello
world


With the Varnish CLI, only the last parameter may use the here document form, which greatly restricts the number of commands that can effectively use them. Trying to use multiple here documents only takes the last one into account.

For example:

command argument << EOF1 << EOF2
heredoc1
EOF1
heredoc2
EOF2


This conceptually results in the following command line:

  • "command"
  • "argument"
  • "<<"
  • "EOF1"
  • "heredoc1\nEOF1\nheredoc2\n"

Other pitfalls include variable expansion of the shell invoking varnishadm but this is not directly related to the Varnish CLI. If you get the quoting right you should be fine even with complex commands.

A number of commands with informational responses support a -j parameter for JSON output, as specified below. The top-level structure of the JSON response is an array with these first three elements:

  • A version number for the JSON format (integer)
  • An array of strings that comprise the CLI command just received
  • The time at which the response was generated, as a Unix epoch time in seconds with millisecond precision (floating point)

The remaining elements of the array form the data that are specific to the CLI command, and their structure and content depend on the command.

For example, the response to status -j just contains a string in the top-level array indicating the state of the child process ("running", "stopped" and so forth):

[ 2, ["status", "-j"], 1538031732.632, "running"
]


The JSON responses to other commands may have longer lists of elements, which may have simple data types or form structured objects.

JSON output is only returned if command execution was successful. The output for an error response is always the same as it would have been for the command without the -j parameter.

Authenticate.


List backends.

-p also shows probe status.

-j specifies JSON output.

Unless -j is specified for JSON output, the output format is five columns of dynamic width, separated by white space with the fields:

  • Backend name
  • Admin: How health state is determined:
  • healthy: Set healthy through backend.set_health.
  • sick: Set sick through backend.set_health.
  • probe: Health state determined by a probe or some other dynamic mechanism.
  • deleted: Backend has been deleted, but not yet cleaned up.

Admin has precedence over Health

  • Probe X/Y: X out of Y checks have succeeded

    X and Y are backend specific and may represent probe checks, other backends or any other metric.

    If there is no probe or the director does not provide details on probe check results, 0/0 is output.

  • Health: Probe health state
  • healthy
  • sick

If there is no probe, healthy is output.

Last change: Timestamp when the health state last changed.

The health state reported here is generic. A backend's health may also depend on the context it is being used in (e.g. the object's hash), so the actual health state as visible from VCL (e.g. using std.healthy()) may differ.

For -j, the object members should be self explanatory, matching the fields described above. probe_message has the format [X, Y, "state"] as described above for Probe. JSON Probe details (-j -p arguments) are director specific.



Set health status of backend(s) matching <backend_pattern>.
  • With auto, the health status is determined by a probe or some other dynamic mechanism, if any
  • healthy sets the backend as usable
  • sick sets the backend as unsable



Mark obsolete all objects where all the conditions match.

See vcl(7)_ban for details



List the active bans.

Unless -j is specified for JSON output, the output format is:

  • Time the ban was issued.
  • Objects referencing this ban.
  • C if ban is completed = no further testing against it.
  • if lurker debugging is enabled:
  • R for req.* tests
  • O for obj.* tests
  • Pointer to ban object

Ban specification

Durations of ban specifications get normalized, for example "7d" gets changed into "1w".



Print welcome banner.


Show command/protocol help.

-j specifies JSON output.



Clear the last panic, if any, -z will clear related varnishstat counter(s)


Return the last panic, if any.

-j specifies JSON output -- the panic message is returned as an unstructured JSON string.



Reset parameter to default value.


Set parameter value.

The JSON output is the same as param.show -j <param> and contains the updated value as it would be represented by a subsequent execution of param.show.

This can be useful to later verify that a parameter value didn't change and to use the value from the JSON output to reset the parameter to the desired value.



Show parameters and their values.

The long form with -l shows additional information, including documentation and minimum, maximum and default values, if defined for the parameter. JSON output is specified with -j, in which the information for the long form is included; only one of -l or -j is permitted. If a parameter is specified with <param>, show only that parameter. If changed is specified, show only those parameters whose values differ from their defaults.



Show the pid of the master process, and the worker if it's running.

-j specifies JSON output.



Keep connection alive.

The response is formatted as JSON if -j is specified.



Close connection.


Start the Varnish cache process.


Check status of Varnish cache process.

-j specifies JSON output.



Stop the Varnish cache process.


List storage devices.

-j specifies JSON output.



List all loaded configuration and their dependencies.

Unless -j is specified for JSON output, the output format is up to two columns of dynamic width separated by white space with the fields:

  • VCL: a VCL program
  • Dependency: another VCL program it depends on

Only direct dependencies are listed, and VCLs with multiple dependencies are listed multiple times.



Unload the named configurations (when possible).

Unload the named configurations and labels matching at least one name pattern. All matching configurations and labels are discarded in the correct order with respect to potential dependencies. If one configuration or label could not be discarded because one of its dependencies would remain, nothing is discarded. Each individual name pattern must match at least one named configuration or label.



Compile and load the VCL data under the name provided.

Multi-line VCL can be input using the here document ref_syntax.



Apply label to configuration.

A VCL label is like a UNIX symbolic link, a name without substance, which points to another VCL.

Labels are mandatory whenever one VCL references another.



List all loaded configuration.

Unless -j is specified for JSON output, the output format is five or seven columns of dynamic width, separated by white space with the fields:

  • status: active, available or discarded
  • state: label, cold, warm, or auto
  • temperature: init, cold, warm, busy or cooling
  • busy: number of references to this vcl (integer)
  • name: the name given to this vcl or label
  • [ <- | -> ] and label info last two fields)
  • -> <vcl> : label "points to" the named <vcl>
  • <- (<n> label[s]): the vcl has <n> label(s)




Compile and load the VCL file under the name provided.


Display the source code for the specified configuration.


Force the state of the named configuration.


Dump the VCL symbol-tables.


Switch to the named configuration immediately.


A backend pattern can be a backend name or a combination of a VCL name and backend name in "VCL.backend" format. If the VCL name is omitted, the active VCL is assumed. Partial matching on the backend and VCL names is supported using shell-style wildcards, e.g. asterisk (*).

Examples:

backend.list def*
backend.list b*.def*
backend.set_health default sick
backend.set_health def* healthy
backend.set_health * auto


A ban expression consists of one or more conditions. A condition consists of a field, an operator, and an argument. Conditions can be ANDed together with "&&".

A field can be any of the variables from VCL, for instance req.url, req.http.host or obj.http.set-cookie.

Operators are "==" for direct comparison, "~" for a regular expression match, and ">" or "<" for size comparisons. Prepending an operator with "!" negates the expression.

The argument could be a quoted string, a regexp, or an integer. Integers can have "KB", "MB", "GB" or "TB" appended for size related fields.

A VCL program goes through several states related to the different commands: it can be loaded, used, and later discarded. You can load several VCL programs and switch at any time from one to another. There is only one active VCL, but the previous active VCL will be maintained active until all its transactions are over.

Over time, if you often refresh your VCL and keep the previous versions around, resource consumption will increase, you can't escape that. However, most of the time you want to pay the price only for the active VCL and keep older VCLs in case you'd need to rollback to a previous version.

The VCL temperature allows you to minimize the footprint of inactive VCLs. Once a VCL becomes cold, Varnish will release all the resources that can be be later reacquired. You can manually set the temperature of a VCL or let varnish automatically handle it.

Load a multi-line VCL using shell-style here document:

vcl.inline example << EOF
vcl 4.0;
backend www {

.host = "127.0.0.1";
.port = "8080"; } EOF


Ban all requests where req.url exactly matches the string /news:

ban req.url == "/news"


Ban all documents where the serving host is "example.com" or "www.example.com", and where the Set-Cookie header received from the backend contains "USERID=1663":

ban req.http.host ~ "^(?i)(www\\.)?example\\.com$" && obj.http.set-cookie ~ "USERID=1663"


This manual page was originally written by Per Buer and later modified by Federico G. Schwindt, Dridi Boukelmoune, Lasse Karstensen and Poul-Henning Kamp.