Skip to content

Latest commit

 

History

History
358 lines (263 loc) · 11.6 KB

readme.md

File metadata and controls

358 lines (263 loc) · 11.6 KB

NSQPHP

PHP client for NSQ.

NSQ basics

You can read all about NSQ via the readme on Github, or via the Bitly blog post describing it. More details on nsqd, nsqlookupd are provided within each folder within the project.

Here's some thing I have learned:

  • Clustering is provided only in the sense that nsqlookupd will discover which machines hosts messages for a particular topic. To consume from a cluster, you simply ask an instance of nslookupd where to find messages and then connect to every nsqd it tells you to (this is one of the things that makes nsq good).
  • HA (for pub) is easy due to the fact that each nsqd instance is isolated; you can simply connect to any and send publish your message (I have built this into the client).
  • Resilience is provided by simply writing to more than one nsqd and then de-duplicating on subscribe (I have built this into the client).
  • nsq is not designed as a work queue (for long running tasks) out of the box. The default setting of msg-timeout is 60,000ms (60 seconds). This is the time before nsq will automatically consider a message to have failed and hence requeue it. Our "work" should take much less time than this. Additionally, PHP is a blocking language, and although we are using a non-blocking IO event loop, any work you do to process a message will block the client from being able to reply to any heartbeats etc.

Installation

nsqphp is available to add to your project via composer. Simply add the following to your composer.json.

{
    ...
    "require": {
        ...
        "davegardnerisme/nsqphp": "dev-master"
    }
    ...
}

You can also simply clone it into your project:

git clone git://github.com/davegardnerisme/nsqphp.git
cd nsqphp
git submodule update --init --recursive

To use nsqphp in your projects, just include the bootstrap.php file, or setup autoloading via composer. The design lends itself to a dependency injection container (all dependencies are constructor injected), although you can just setup the dependencies manually when you use it.

Testing it out

Follow the getting started guide to install nsq on localhost.

Publish some events:

php cruft/test-pub.php 10

Fire up a subscriber in one shell:

php cruft/test-sub.php mychannel > /tmp/processed-messages

Then tail the redirected STDOUT in another shell, so you can see the messages received and processed:

tail -f /tmp/processed-messages

Note

In these tests I'm publishing first since I haven't yet got the client to automatically rediscover which nodes have messages for a given topic; hence if you sub first, there are no nodes found with messages for the topic.

Other tests

Multiple channels

The blog post describes a channel:

| Each channel receives a copy of all the messages for a topic. In | practice, a channel maps to a downstream service consuming a topic.

So each message in a topic will be delivered to each channel.

Fire up two subscribers with different channels (one in each shell):

php cruft/test-sub.php mychannel
php cruft/test-sub.php otherchannel

Publish some messages:

php cruft/test-pub.php 10

Each message will be delivered to each channel. It's also worth noting that the API allows you to subscribe to multiple topics/channels within the same process.

Multiple nsqds

Setup a bunch of servers running nsqd and nsqlookupd with hostnames nsq1, nsq2 ... Now publish a bunch of messages to both:

php cruft/test-pub.php 10 nsq1
php cruft/test-pub.php 10 nsq2

Now subscribe:

php cruft/test-sub.php mychannel > /tmp/processed-messages

You will receive 20 messages.

Resilient delivery

Same test as before, but this time we deliver the same message to two nsqd instances and then de-duplicate on subscribe.

php cruft/test-pub.php 10 nsq1,nsq2
php cruft/test-sub.php mychannel > /tmp/processed-messages

This time you should receive only 10 messages.

To do

  • Requeue failed messages using a back-off strategy (currently only simple fixed-delay requeue strategy)
  • Continuously re-evaluate which nodes contain messages for a given topic (that is subscribed to) and establish new connections for those clients (via event loop timer)

The PHP client interface

Messages

Messages are encapsulated by the nsqphp\Message\Message class and are referred to by interface within the code (so you could implement your own).

Interface:

public function getPayload();
public function getId();
public function getAttempts();
public function getTimestamp();

Publishing

The client supports publishing to N nsqd servers, which must be specified explicitly by hostname. Unlike with subscription, there is no facility to lookup the hostnames via nslookupd (and we probably wouldn't want to anyway for speed).

Minimal approach:

    $nsq = new nsqphp\nsqphp;
    $nsq->publishTo('localhost')
        ->publish('mytopic', new nsqphp\Message\Message('some message payload'));

It's up to you to decide if/how to encode your payload (eg: JSON).

HA publishing:

    $nsq = new nsqphp\nsqphp;
    $nsq->publishTo(array('nsq1', 'nsq2', 'nsq3'), nsqphp\nsqphp::PUB_QUORUM)
        ->publish('mytopic', new nsqphp\Message\Message('some message payload'));

We will require a quorum of the publishTo nsqd daemons to respond to consider this operation a success (currently that happens in series). This is assuming I have 3 nsqds running on three hosts which are contactable via nsq1 etc.

This technique is going to log messages twice, which will require de-duplication on subscribe.

Subscribing

The client supports subscribing from N nsqd servers, each of which will be auto-discovered from one or more nslookupd servers. The way this works is that nslookupd is able to provide a list of auto-discovered nodes hosting messages for a given topic. This feature decouples our clients from having to know where to find messages.

So when subscribing, the first thing we need to do is initialise our lookup service object:

    $lookup = new nsqphp\Lookup\Nsqlookupd;

Or alternatively:

    $lookup = new nsqphp\Lookup\Nsqlookupd('nsq1,nsq2');

We can then use this to subscribe:

    $lookup = new nsqphp\Lookup\Nsqlookupd;
    $nsq = new nsqphp\nsqphp($lookup);
    $nsq->subscribe('mytopic', 'somechannel', function($msg) {
        echo $msg->getId() . "\n";
        })->run();

Warning: if our callback were to throw any Exceptions, the messages would not be retried using these settings - read on to find out more.

Or a bit more in the style of PHP (?):

    $lookup = new nsqphp\Lookup\Nsqlookupd;
    $nsq = new nsqphp\nsqphp($lookup);
    $nsq->subscribe('mytopic', 'somechannel', 'msgCallback')
        ->run();

    function msgCallback($msg)
    {
        echo $msg->getId() . "\n";
    }

We can also subscribe to more than one channel/stream:

    $lookup = new nsqphp\Lookup\Nsqlookup;
    $nsq = new nsqphp\nsqphp($lookup);
    $nsq->subscribe('mytopic', 'somechannel', 'msgCallback')
        ->subscribe('othertopic', 'somechannel', 'msgCallback')
        ->run();

Retrying failed messages

The PHP client will catch any thrown Exceptions that happen within the callback and then either (a) retry, or (b) discard the messages. Usually you won't want to discard the messages.

To fix this, we need a requeue strategy - this is in the form of any object that implements nsqphp\RequeueStrategy\RequeueStrategyInterface:

    public function shouldRequeue(MessageInterface $msg);

The client currently ships with one; a fixed delay strategy:

    $requeueStrategy = new nsqphp\RequeueStrategy\FixedDelay;
    $lookup = new nsqphp\Lookup\Nsqlookupd;
    $nsq = new nsqphp\nsqphp($lookup, NULL, $requeueStrategy);
    $nsq->subscribe('mytopic', 'somechannel', 'msgCallback')
        ->run();

    function msgCallback($msg)
    {
        if (rand(1,3) == 1) {
            throw new \Exception('Argh, something bad happened');
        }
        echo $msg->getId() . "\n";
    }

De-duplication on subscribe

Recall that to achieve HA we simply duplicate on publish into two different nsqd servers. To perform de-duplication we simply need to supply an object that implements nsqphp\Dedupe\DedupeInterface.

public function containsAndAdd($topic, $channel, MessageInterface $msg);

The PHP client ships with two mechanisms for de-duplicating messages on subscribe. Both are based around the opposite of a bloom filter. One maintains a hash map as a PHP array (and hence bound to a single process); the other calls out to Memcached and hence can share the data structure between many processes.

We can use this thus:

    $requeueStrategy = new nsqphp\RequeueStrategy\FixedDelay;
    $dedupe = new nsqphp\Dedupe\OppositeOfBloomFilterMemcached;
    $lookup = new nsqphp\Lookup\Nsqlookupd;
    $nsq = new nsqphp\nsqphp($lookup, $dedupe, $requeueStrategy);
    $nsq->subscribe('mytopic', 'somechannel', 'msgCallback')
        ->run();

    function msgCallback($msg)
    {
        if (rand(1,3) == 1) {
            throw new \Exception('Argh, something bad happened');
        }
        echo $msg->getId() . "\n";
    }

You can read more about de-duplication on my blog, however it's worth keeping the following in mind:

  • With Memcached de-duplication we can then happily launch N processes to subscribe to the same topic and channel, and only process the messages once.
  • De-duplication is not guaranteed (in fact far from it) - the implementations shipped are based on a lossy hash map, and hence are probabilistic in how they will perform. For events fed down at a similar time, they will usually perform acceptably (and they can be tuned to trade off memory usage for de-duplication abilities)
  • nsq is designed around the idea of idempotent subscribers - eg: your subscriber must be able to cope with processing a duplicated message (writing into Cassandra is an example of a system that copes well with executing something twice).

Logging

The final optional dependency is a logger, in the form of some object that implements nsqphp\Logger\LoggerInterface (there is no standard logger interface shipped with PHP to the best of my knowledge):

    public function error($msg);
    public function warn($msg);
    public function info($msg);
    public function debug($msg);

The PHP client ships with a logger that dumps all logging information to STDERR. Putting all of this together we'd have something similar to the test-sub.php file:

    $requeueStrategy = new nsqphp\RequeueStrategy\FixedDelay;
    $dedupe = new nsqphp\Dedupe\OppositeOfBloomFilterMemcached;
    $lookup = new nsqphp\Lookup\Nsqlookupd;
    $logger = new nsqphp\Logger\Stderr;
    $nsq = new nsqphp\nsqphp($lookup, $dedupe, $requeueStrategy, logger);
    $nsq->subscribe('mytopic', 'somechannel', 'msgCallback')
        ->run();

    function msgCallback($msg)
    {
        if (rand(1,3) == 1) {
            throw new \Exception('Argh, something bad happened');
        }
        echo $msg->getId() . "\n";
    }

Design log

  • main client based on event loop (powered by React PHP) to allow us to handle multiple connections to multiple nsqd instances