VMOD directors - Varnish Directors Module

SYNOPSIS

import directors [as name] [from "path"]

new xround_robin = directors.round_robin()

    VOID xround_robin.add_backend(BACKEND)

    VOID xround_robin.remove_backend(BACKEND)

    BACKEND xround_robin.backend()

new xfallback = directors.fallback(BOOL sticky=0)

    VOID xfallback.add_backend(BACKEND)

    VOID xfallback.remove_backend(BACKEND)

    BACKEND xfallback.backend()

new xrandom = directors.random()

    VOID xrandom.add_backend(BACKEND, REAL)

    VOID xrandom.remove_backend(BACKEND)

    BACKEND xrandom.backend()

new xhash = directors.hash()

    VOID xhash.add_backend(BACKEND, REAL weight=1.0)

    VOID xhash.remove_backend(BACKEND)

    BACKEND xhash.backend(STRING)

new xshard = directors.shard()

    VOID xshard.set_warmup(REAL probability=0.0)

    VOID xshard.set_rampup(DURATION duration=0)

    VOID xshard.associate(BLOB param=0)

    BOOL xshard.add_backend(BACKEND backend, [STRING ident], [DURATION rampup], [REAL weight])

    BOOL xshard.remove_backend([BACKEND backend], [STRING ident])

    BOOL xshard.clear()

    BOOL xshard.reconfigure(INT replicas=67)

    INT xshard.key(STRING)

    BACKEND xshard.backend([ENUM by], [INT key], [BLOB key_blob], [INT alt], [REAL warmup], [BOOL rampup], [ENUM healthy], [BLOB param], [ENUM resolve])

    VOID xshard.debug(INT)

new xshard_param = directors.shard_param()

    VOID xshard_param.clear()

    VOID xshard_param.set([ENUM by], [INT key], [BLOB key_blob], [INT alt], [REAL warmup], [BOOL rampup], [ENUM healthy])

    STRING xshard_param.get_by()

    INT xshard_param.get_key()

    INT xshard_param.get_alt()

    REAL xshard_param.get_warmup()

    BOOL xshard_param.get_rampup()

    STRING xshard_param.get_healthy()

    BLOB xshard_param.use()

BACKEND lookup(STRING)

DESCRIPTION

vmod_directors enables backend load balancing in Varnish.

The module implements load balancing techniques, and also serves as an example on how one could extend the load balancing capabilities of Varnish.

To enable load balancing you must import this vmod (directors).

Then you define your backends. Once you have the backends declared you can add them to a director. This happens in executed VCL code. If you want to emulate the previous behavior of Varnish 3.0 you can just initialize the directors in vcl_init{}, like this:

sub vcl_init {
    new vdir = directors.round_robin();
    vdir.add_backend(backend1);
    vdir.add_backend(backend2);
}

As you can see there is nothing keeping you from manipulating the directors elsewhere in VCL. So, you could have VCL code that would add more backends to a director when a certain URL is called.

Note that directors can use other directors as backends.

new xround_robin = directors.round_robin()

Create a round robin director.

This director will pick backends in a round robin fashion.

Example:

new vdir = directors.round_robin();

VOID xround_robin.add_backend(BACKEND)

Add a backend to the round-robin director.

Example:

vdir.add_backend(backend1);

VOID xround_robin.remove_backend(BACKEND)

Remove a backend from the round-robin director.

Example:

vdir.remove_backend(backend1);

BACKEND xround_robin.backend()

Pick a backend from the director.

Example:

set req.backend_hint = vdir.backend();

new xfallback = directors.fallback(BOOL sticky=0)

Create a fallback director.

A fallback director will try each of the added backends in turn, and return the first one that is healthy.

If sticky is set to true, the director will keep using the healthy backend, even if a higher-priority backend becomes available. Once the whole backend list is exhausted, it’ll start over at the beginning.

Example:

new vdir = directors.fallback();

VOID xfallback.add_backend(BACKEND)

Add a backend to the director.

Note that the order in which this is done matters for the fallback director.

Example:

vdir.add_backend(backend1);

VOID xfallback.remove_backend(BACKEND)

Remove a backend from the director.

Example:

vdir.remove_backend(backend1);

BACKEND xfallback.backend()

Pick a backend from the director.

Example:

set req.backend_hint = vdir.backend();

new xrandom = directors.random()

Create a random backend director.

The random director distributes load over the backends using a weighted random probability distribution.

The “testable” random generator in varnishd is used, which enables deterministic tests to be run (See: d00004.vtc).

Example:

new vdir = directors.random();

VOID xrandom.add_backend(BACKEND, REAL)

Add a backend to the director with a given weight.

Each backend will receive approximately 100 * (weight / (sum(all_added_weights))) per cent of the traffic sent to this director.

Example:

# 2/3 to backend1, 1/3 to backend2.
vdir.add_backend(backend1, 10.0);
vdir.add_backend(backend2, 5.0);

VOID xrandom.remove_backend(BACKEND)

Remove a backend from the director.

Example:

vdir.remove_backend(backend1);

BACKEND xrandom.backend()

Pick a backend from the director.

Example:

set req.backend_hint = vdir.backend();

new xhash = directors.hash()

Create a hashing backend director.

The director chooses the backend server by computing a hash/digest of the string given to xhash.backend().

Commonly used with client.ip or a session cookie to get sticky sessions.

Example:

new vdir = directors.hash();

VOID xhash.add_backend(BACKEND, REAL weight=1.0)

Add a backend to the director with a certain weight.

Weight is used as in the random director. Recommended and default value is 1.0 unless you have special needs.

Example:

vdir.add_backend(normal_backend);
vdir.add_backend(larger_backend, 1.5);

VOID xhash.remove_backend(BACKEND)

Remove a backend from the director.

Example::

vdir.remove_backend(larger_backend);

BACKEND xhash.backend(STRING)

Pick a backend from the hash director.

Use the string or list of strings provided to pick the backend.

Example::

# pick a backend based on the cookie header from the client set req.backend_hint = vdir.backend(req.http.cookie);

new xshard = directors.shard()

Create a shard director.

Introduction

The shard director selects backends by a key, which can be provided directly or derived from strings. For the same key, the shard director will always return the same backend, unless the backend configuration or health state changes. Conversely, for differing keys, the shard director will likely choose different backends. In the default configuration, unhealthy backends are not selected.

The shard director resembles the hash director, but its main advantage is that, when the backend configuration or health states change, the association of keys to backends remains as stable as possible.

In addition, the rampup and warmup features can help to further improve user-perceived response times.

Sharding

This basic technique allows for numerous applications like optimizing backend server cache efficiency, Varnish clustering or persisting sessions to servers without keeping any state, and, in particular, without the need to synchronize state between nodes of a cluster of Varnish servers:

  • Many applications use caches for data objects, so, in a cluster of application servers, requesting similar objects from the same server may help to optimize efficiency of such caches.

    For example, sharding by URL or some id component of the url has been shown to drastically improve the efficiency of many content management systems.

  • As special case of the previous example, in clusters of Varnish servers without additional request distribution logic, each cache will need store all hot objects, so the effective cache size is approximately the smallest cache size of any server in the cluster.

    Sharding allows to segregate objects within the cluster such that each object is only cached on one of the servers (or on one primary and one backup, on a primary for long and others for short etc…). Effectively, this will lead to a cache size in the order of the sum of all individual caches, with the potential to drastically increase efficiency (scales by the number of servers).

  • Another application is to implement persistence of backend requests, such that all requests sharing a certain criterion (such as an IP address or session ID) get forwarded to the same backend server.

When used with clusters of varnish servers, the shard director will, if otherwise configured equally, make the same decision on all servers. In other words, requests sharing a common criterion used as the shard key will be balanced onto the same backend server(s) no matter which Varnish server handles the request.

The drawbacks are:

  • the distribution of requests depends on the number of requests per key and the uniformity of the distribution of key values. In short, while this technique may lead to much better efficiency overall, it may also lead to less good load balancing for specific cases.

  • When a backend server becomes unavailable, every persistence technique has to reselect a new backend server, but this technique will also switch back to the preferred server once it becomes healthy again, so when used for persistence, it is generally less stable compared to stateful techniques (which would continue to use a selected server for as long as possible (or dictated by a TTL)).

Method

When xshard.reconfigure() is called explicitly (or implicitly at the end of any task containing reconfigurations like xshard.add_backend()), a consistent hashing circular data structure gets built from the last 32 bits of SHA256 hash values of <ident><n> (default ident being the backend name) for each backend and for a running number n from 1 to the replicas argument to xshard.reconfigure(). Hashing creates the seemingly random order for placement of backends on the consistent hashing ring. When xshard.add_backend() was called with a weight argument, replicas is scaled by that weight to add proportionally more copies of the that backend on the ring.

When xshard.backend() is called, a load balancing key gets generated unless provided. The smallest hash value in the circle is looked up that is larger than the key (searching clockwise and wrapping around as necessary). The backend for this hash value is the preferred backend for the given key.

If a healthy backend is requested, the search is continued linearly on the ring as long as backends found are unhealthy or all backends have been checked. The order of these “alternative backends” on the ring is likely to differ for different keys. Alternative backends can also be selected explicitly.

On consistent hashing see:

Error Reporting

Failing methods should report errors to VSL with the Error tag, so when configuring the shard director, you are advised to check:

varnishlog -I Error:^vmod_directors.shard

Additional information may be provided as Notices, which can be checked using

varnishlog -I Notice:^vmod_directors.shard

VOID xshard.set_warmup(REAL probability=0.0)

Set the default warmup probability. See the warmup parameter of xshard.backend(). If probability is 0.0 (default), warmup is disabled.

VOID xshard.set_rampup(DURATION duration=0)

Set the default rampup duration. See rampup parameter of xshard.backend(). If duration is 0 (default), rampup is disabled.

VOID xshard.associate(BLOB param=0)

Associate a default directors.shard_param() object or clear an association.

The value of the param argument must be a call to the xshard_param.use() method. No argument clears the association.

The association can be changed per backend request using the param argument of xshard.backend().

BOOL xshard.add_backend(BACKEND backend, [STRING ident], [DURATION rampup], [REAL weight])

BOOL xshard.add_backend(
      BACKEND backend,
      [STRING ident],
      [DURATION rampup],
      [REAL weight]
)

Add a backend backend to the director.

ident: Optionally specify an identification string for this backend, which will be hashed by xshard.reconfigure() to construct the consistent hashing ring. The identification string defaults to the backend name.

ident allows to add multiple instances of the same backend.

rampup: Optionally specify a specific rampup time for this backend. Otherwise, the per-director rampup time is used (see xshard.set_rampup()).

weight: Optionally specify a weight to scale the xshard.reconfigure() replicas parameter. weight is limited to at least 1. Values above 10 probably do not make much sense. The effect of weight is also capped such that the total number of replicas does not exceed UINT32_MAX.

BOOL xshard.remove_backend([BACKEND backend], [STRING ident])

BOOL xshard.remove_backend(
      [BACKEND backend=0],
      [STRING ident=0]
)

Remove backend(s) from the director. Either backend or ident must be specified. ident removes a specific instance. If backend is given without ident, all instances of this backend are removed.

BOOL xshard.clear()

Remove all backends from the director.

BOOL xshard.reconfigure(INT replicas=67)

Explicitly reconfigure the consistent hashing ring to reflect backend changes to become effective immediately.

If this method is not called explicitly, reconfiguration happens at the end of the current task (after vcl_init {} or when the current client or backend task is finished).

INT xshard.key(STRING)

Convenience method to generate a sharding key for use with the key argument to the xshard.backend() method by hashing the given string with SHA256.

To generate sharding keys using other hashes, use a custom vmod like vmod blobdigest with the key_blob argument of the xshard.backend() method.

BACKEND xshard.backend([ENUM by], [INT key], [BLOB key_blob], [INT alt], [REAL warmup], [BOOL rampup], [ENUM healthy], [BLOB param], [ENUM resolve])

BACKEND xshard.backend(
      [ENUM {HASH, URL, KEY, BLOB} by=HASH],
      [INT key],
      [BLOB key_blob],
      [INT alt=0],
      [REAL warmup=-1],
      [BOOL rampup=1],
      [ENUM {CHOSEN, IGNORE, ALL} healthy=CHOSEN],
      [BLOB param],
      [ENUM {NOW, LAZY} resolve]
)

Lookup a backend on the consistent hashing ring.

This documentation uses the notion of an order of backends for a particular shard key. This order is deterministic but seemingly random as determined by the consistent hashing algorithm and is likely to differ for different keys, depending on the number of backends and the number of replicas. In particular, the backend order referred to here is _not_ the order given when backends are added.

  • by how to determine the sharding key

    • HASH:

      • when called in backend context and in vcl_pipe {}: Use the varnish hash value as set by vcl_hash{}

      • when called in client context other than vcl_pipe {}: hash req.url

    • URL: hash req.url / bereq.url

    • KEY: use the key argument

    • BLOB: use the key_blob argument

  • key lookup key with by=KEY

    the xshard.key() method may come handy to generate a sharding key from custom strings.

  • key_blob lookup key with by=BLOB

    Currently, this uses the first 4 bytes from the given blob in network byte order (big endian), left-padded with zeros for blobs smaller than 4 bytes.

  • alt alternative backend selection

    Select the alt-th alternative backend for the given key.

    This is particularly useful for retries / restarts due to backend errors: By setting alt=req.restarts or alt=bereq.retries with healthy=ALL, another server gets selected.

    The rampup and warmup features are only active for alt==0

  • rampup slow start for servers which just went healthy

    If alt==0 and the chosen backend is in its rampup period, with a probability proportional to the fraction of time since the backup became healthy to the rampup period, return the next alternative backend, unless this is also in its rampup period.

    The default rampup interval can be set per shard director using the xshard.set_rampup() method or specifically per backend with the xshard.add_backend() method.

  • warmup probabilistic alternative server selection

    possible values: -1, 0..1

    -1: use the warmup probability from the director definition

    Only used for alt==0: Sets the ratio of requests (0.0 to 1.0) that goes to the next alternate backend to warm it up when the preferred backend is healthy. Not active if any of the preferred or alternative backend are in rampup.

    warmup=0.5 is a convenient way to spread the load for each key over two backends under normal operating conditions.

  • healthy

    • CHOSEN: Return a healthy backend if possible.

      For alt==0, return the first healthy backend or none.

      For alt > 0, ignore the health state of backends skipped for alternative backend selection, then return the next healthy backend. If this does not exist, return the last healthy backend of those skipped or none.

    • IGNORE: Completely ignore backend health state

      Just return the first or alt-th alternative backend, ignoring health state, rampup and warmup.

    • ALL: Check health state also for alternative backend selection

      For alt > 0, return the alt-th alternative backend of all those healthy, the last healthy backend found or none.

  • resolve

    default: LAZY in vcl_init{}, NOW otherwise

    • NOW: look up a backend and return it.

      Can not be used in vcl_init{}.

    • LAZY: return an instance of this director for later backend resolution.

      LAZY mode is required for referencing shard director instances, for example as backends for other directors (director layering).

      In vcl_init{} and on the client side, LAZY mode can not be used with any other argument.

      On the backend side and in vcl_pipe {}, parameters from arguments or an associated parameter set affect the shard director instance for the backend request irrespective of where it is referenced.

  • param

    Use or associate a parameter set. The value of the param argument must be a call to the xshard_param.use() method.

    default: as set by xshard.associate() or unset.

    • for resolve=NOW take parameter defaults from the directors.shard_param() parameter set

    • for resolve=LAZY associate the directors.shard_param() parameter set for this backend request

      Implementation notes for use of parameter sets with resolve=LAZY:

      • A param argument remains associated and any changes to the associated parameter set affect the sharding decision once the director resolves to an actual backend.

      • If other parameter arguments are also given, they have preference and are kept even if the parameter set given by the param argument is subsequently changed within the same backend request.

      • Each call to xshard.backend() overrides any previous call.

VOID xshard.debug(INT)

intentionally undocumented

new xshard_param = directors.shard_param()

Create a shard parameter set.

A parameter set allows for re-use of xshard.backend() arguments across many shard director instances and simplifies advanced use cases (e.g. shard director with custom parameters layered below other directors).

Parameter sets have two scopes:

  • per-VCL scope defined in vcl_init{}

  • per backend request scope

The per-VCL scope defines defaults for the per backend scope. Any changes to a parameter set in backend context and in vcl_pipe {} only affect the respective backend request.

Parameter sets can not be used in client context except for vcl_pipe {}.

The following example is a typical use case: A parameter set is associated with several directors. Director choice happens on the client side and parameters are changed on the backend side to implement retries on alternative backends:

sub vcl_init {
  new shard_param = directors.shard_param();

  new dir_A = directors.shard();
  dir_A.add_backend(...);
  dir_A.reconfigure();
  dir_A.associate(shard_param.use()); # <-- !

  new dir_B = directors.shard();
  dir_B.add_backend(...);
  dir_B.reconfigure();
  dir_B.associate(shard_param.use()); # <-- !
}

sub vcl_recv {
  if (...) {
    set req.backend_hint = dir_A.backend(resolve=LAZY);
  } else {
    set req.backend_hint = dir_B.backend(resolve=LAZY);
  }
}

sub vcl_backend_fetch {
  # changes dir_A and dir_B behaviour
  shard_param.set(alt=bereq.retries, by=URL);
}

VOID xshard_param.clear()

Reset the parameter set to default values as documented for xshard.backend().

  • in vcl_init{}, resets the parameter set default for this VCL in

  • backend context and in vcl_pipe {}, resets the parameter set for this backend request to the VCL defaults

Restricted to: vcl_pipe, backend, housekeeping.

VOID xshard_param.set([ENUM by], [INT key], [BLOB key_blob], [INT alt], [REAL warmup], [BOOL rampup], [ENUM healthy])

VOID xshard_param.set(
      [ENUM {HASH, URL, KEY, BLOB} by],
      [INT key],
      [BLOB key_blob],
      [INT alt],
      [REAL warmup],
      [BOOL rampup],
      [ENUM {CHOSEN, IGNORE, ALL} healthy]
)

Change the given parameters of a parameter set as documented for xshard.backend().

  • in vcl_init{}, changes the parameter set default for this VCL

  • in backend context and in vcl_pipe {}, changes the parameter set for this backend request, keeping the defaults set for this VCL for unspecified arguments.

Restricted to: vcl_pipe, backend, housekeeping.

STRING xshard_param.get_by()

Get a string representation of the by enum argument which denotes how a shard director using this parameter object would derive the shard key. See xshard.backend().

INT xshard_param.get_key()

Get the key which a shard director using this parameter object would use. See xshard.backend().

INT xshard_param.get_alt()

Get the alt parameter which a shard director using this parameter object would use. See xshard.backend().

REAL xshard_param.get_warmup()

Get the warmup parameter which a shard director using this parameter object would use. See xshard.backend().

BOOL xshard_param.get_rampup()

Get the rampup parameter which a shard director using this parameter object would use. See xshard.backend().

STRING xshard_param.get_healthy()

Get a string representation of the healthy enum argument which a shard director using this parameter object would use. See xshard.backend().

BLOB xshard_param.use()

For use with the param argument of xshard.backend() to associate this shard parameter set with a shard director.

Restricted to: vcl_pipe, backend, housekeeping.

BACKEND lookup(STRING)

Lookup a backend by its name.

Restricted to: housekeeping.

ACKNOWLEDGEMENTS

Development of a previous version of the shard director was partly sponsored by Deutsche Telekom AG - Products & Innovation.

Development of a previous version of the shard director was partly sponsored by BILD GmbH & Co KG.