avldya/nuster

A web caching proxy server based on HAProxy

Nuster, a web caching proxy server.

Table of Contents

• Introduction
• Performance
• Installation
  • Download
  • Build
• Usage
• Directives
  • cache
  • filter cache
  • cache-rule
• FAQ
• Example
• Conventions
• Contributing
• TODO
• License

Introduction

Nuster is a simple yet powerful web caching proxy server based on HAProxy. It is 100% compatible with HAProxy, and takes full advantage of the ACL functionality of HAProxy to provide fine-grained caching policy based on the content of request, response or server status, such as,

• request url: cache only if the request url equals to X
• request query: cache only if the request query contains X and equls to Y
• cookie: cache only if cookie contains X and equls to Y
• response header: cache only if the response contains header X equals to Y
• rate: cache only if the request rate is greater than X
• etc, any combination of above

Performance

Nuster is very fast, some test shows nuster is almost three times faster than nginx when both using single core, and nearly two times faster than nginx and three times faster than varnish when using all cores.

See detailed benchmark

Installation

Download

Download stable version from releases page for production use, otherwise git clone the source code.

Build

make TARGET=linux2628
make install

use make TARGET=linux2628 USE_PTHREAD_PSHARED=1 to use pthread lib

See HAProxy README for details.

Usage

Nuster is based on HAProxy, all directives from HAProxy are supported in nuster.

In order to use cache functionality, cache on should be declared in global section and a cache filter along with some cache-rules should be added into backend or listen section.

If cache off is declared or there is no cache on|off directive, nuster acts just like HAProxy, as a TCP and HTTP load balancer.

Directives

cache

syntax: cache on|off [share on|off] [data-size size] [dict-size size]

default: none

context: global

Determines whether to use cache or not.

share

share on

A memory zone with a size of data-size + dict-size will be created. Except for temporary data created and destroyed within request, all cache related data including http response data, keys and overheads are stored in this memroy zone and shared between all processes.

If no more memory can be allocated from this memory zone, new requests that should be cached according to defined cache rules will not be cached unless some memory are freed.

share off

Cache data are stored in a memory pool which allocates memory dynamically from system in case there is no available memory in the pool.

A global internal counter monitors the memory usage of all http response data across all processes, new requests will not be cached if the counter exceeds data-size.

By default, share is set to on in multiple processes mode, and off in single process mode.

data-size

With share on, it determines the size of memory zone along with dict-size.

With share off, it detemines the maximum memory used by cache.

It accepts units like m, M, g and G. By default, the size is 1024 * 1024 bytes, which is also the minimal size.

dict-size

Determines the size of memory used by hash table in share on mode.

It has no effect in share off mode, the hash table resize itself if full.

It accepts units like m, M, g and G. By default, the size is 1024 * 1024 bytes, which is also the minimal size.

Note that it only decides the memory used by hash table not keys. In fact, keys are stored in memory zone which is limited by data-size.

dict-size is different from number of keys. New keys can still be added to hash table even if the number of keys exceeds dict-size as long as there are enough memory.

Nevertheless it may lead to a performance drop if number of keys is greater than dict-size.

An approximate number of keys multiplied by 8 (normally) as dict-size should work.

filter cache

syntax: filter cache [on|off]

default: on

context: backend, listen

Define a cache filter, additional cache-rule should be defined. It can be turned off separately by including off. If there are multiple filters, make sure that cache filter is put after all other filters.

cache-rule

syntax: cache-rule name [key KEY] [ttl TTL] [code CODE] [if|unless condition]

default: none

context: backend, listen

Define cache rule. It is possible to declare multiple rules in the same section. The order is important because the matching process stops on the first match.

acl pathA path /a.html
filter cache
cache-rule all ttl 3600
cache-rule path01 ttl 60 if pathA

cache-rule path01 will never match because first rule will cache everything.

name

Define a name for this cache-rule.

key KEY

Define the key for cache, it takes a string combined by following keywords with . separator:

• method: http method, GET/POST...
• scheme: http or https
• host: the host in the request
• path: the URL path of the request
• query: the whole query string of the request
• header_NAME: the value of header NAME
• cookie_NAME: the value of cookie NAME
• param_NAME: the value of query NAME
• body: the body of the request

By default the key is method.scheme.host.path.query.body

Example

GET http://www.example.com/q?name=X&type=Y

http header:
GET /q?name=X&type=Y HTTP/1.1
Host: www.example.com
ASDF: Z
Cookie: logged_in=yes; user=nuster;

Should result:

• method: GET
• scheme: http
• host: www.example.com
• path: /q
• query: name=X&type=Y
• header_ASDF: Z
• cookie_user: nuster
• param_type: Y
• body: (empty)

So default key produces GEThttpwww.example.com/qname=X&type=Y, and key method.scheme.host.path.header_ASDF.cookie_user.param_type produces GEThttpwww.example.com/qZnusterY

If a request has the same key as a cached http response data, then cached data will be sent to the client.

ttl TTL

Set a TTL on key, after the TTL has expired, the key will be deleted. It accepts units like d, h, m and s. Default ttl is 3600 seconds. Set to 0 if you don't want to expire the key.

code CODE1,CODE2...

Cache only if the response status code is CODE. By default, only 200 response is cached. You can use all to cache all responses.

cache-rule only200
cache-rule 200and404 code 200,404
cache-rule all code all

if|unless condition

Define when to cache using HAProxy ACL. See 7. Using ACLs and fetching samples section in HAProxy configuration

FAQ

How to debug?

Set debug in global section, or start haproxy with -d.

Cache related debug messages start with [CACHE].

How to cache POST request?

Enable option http-buffer-request.

By default, the cache key includes the body of the request, remember to put body in key field if you use a customized key.

Note that the body of the request maybe incomplete, refer to option http-buffer-request section in HAProxy configuration for details.

Also it might be a good idea to put it separately in a dedicated backend as example does.

Example

global
    cache on data-size 100m
    #daemon
    ## to debug cache
    #debug
defaults
    retries 3
    option redispatch
    timeout client  30s
    timeout connect 30s
    timeout server  30s
frontend web1
    bind *:8080
    mode http
    acl pathPost path /search
    use_backend app1a if pathPost
    default_backend app1b
backend app1a
    balance roundrobin
    # mode must be http
    mode http

    # http-buffer-request must be enabled to cache post request
    option http-buffer-request

    acl pathPost path /search

    # enable cache for this proxy
    filter cache

    # cache /search for 120 seconds. Only works when POST/PUT
    cache-rule rpost ttl 120 if pathPost

    server s1 10.0.0.10:8080
backend app1b
    balance     roundrobin
    mode http

    filter cache on

    # cache /a.jpg, not expire
    acl pathA path /a.jpg
    cache-rule r1 ttl 0 if pathA

    # cache /mypage, key contains cookie[userId], so it will be cached per user
    acl pathB path /mypage
    cache-rule r2 key method.scheme.host.path.query.cookie_userId ttl 60 if pathB

    # cache /a.html if response's header[cache] is yes
    http-request set-var(txn.pathC) path
    acl pathC var(txn.pathC) -m str /a.html
    acl resHdrCache1 res.hdr(cache) yes
    cache-rule r3 if pathC resHdrCache1

    # cache /heavy for 100 seconds if be_conn greater than 10
    acl heavypage path /heavy
    acl tooFast be_conn ge 100
    cache-rule heavy ttl 100 if heavypage tooFast 

    # cache all if response's header[asdf] is fdsa
    acl resHdrCache2 res.hdr(asdf)  fdsa
    cache-rule resCache ttl 0 if resHdrCache1

    server s1 10.0.0.10:8080

frontend web2
    bind *:8081
    mode http
    default_backend app2
backend app2
    balance     roundrobin
    mode http

    # disable cache on this proxy
    filter cache off
    cache-rule all

    server s2 10.0.0.11:8080

listen web3
    bind *:8082
    mode http

    filter cache
    cache-rule everything

    server s3 10.0.0.12:8080

Conventions

1. Files with same name: those with .md extension belong to Nuster, otherwise HAProxy

Contributing

• Join the development
• Give feedback
• Report issues
• Send pull requests
• Spread nuster

TODO

• Purge cache(web/api)
• Cache stats(web/api)
• etc

License

Copyright (C) 2017, Jiang Wenyuan, < koubunen AT gmail DOT com >

All rights reserved.

Licensed under GPL, the same as HAProxy

HAProxy and other sources license notices: see relevant individual files.