DDevelopment & Design

HAProxy load balancing with sticky sessions based on request body

Integrating systems you have no influence on needs a lot of workarounds. Recently we could not scale Python service consuming SOAP messages with a new hardware. It just didn’t benefit from more processing cores. On the other hand (and this happens often with older software) setting up several instances gave almost linear scalability. Only thing left – configure a loadbalancer and we are done.

Easier said than done. We had to make sure messages are loadbalanced but also that all messages related to given customer USSD conversation always hit the same backend service. So, we had to use application layer information to configure sticky sessions. This is not straightforward in HAProxy when you have to look into http payload and parse some specific information. We used HAProxy 1.6 and simple LUA script to do just that:

core.Alert("LUA script parsing SOAP element loaded");

function parseElement(txn, salt)

    local payload = txn.req:dup()

    -- parses integer value from element named "element"
    local value = string.match(string.match(payload, "element>%d+<"), "%d+")
    core.Info("value: " .. value)
    return value
end

-- register HAProxy "fetch"
core.register_fetches("parseElement", parseElement)

Put this script into a file and it can be loaded in HAProxy configuration using lua-load directive.

Script registers new HAProxy fetch which can be used to configure session stickiness.

balance roundrobin
stick-table type string size 30k expire 30m
stick on "lua.parseElement" table nodes

You have to also make sure all payload is loaded before you start parsing it. This can be achieved with option http-buffer-request configuration directive.

— Previous article

Formatting Java Time with Spring Boot using JSON

Next article —

Sputnik lands on GitHub!

You May Also Like
DDevelopment & Design

Recently at storm-users

I've been reading through storm-users Google Group recently. This resolution was heavily inspired by Adam Kawa's post "Football zero, Apache Pig hero". Since I've encountered a lot of insightful and very interesting information I've decided to describe some of those in this post.

  • nimbus will work in HA mode - There's a pull request open for it already... but some recent work (distributing topology files via Bittorrent) will greatly simplify the implementation. Once the Bittorrent work is done we'll look at reworking the HA pull request. (storm’s pull request)

  • pig on storm - Pig on Trident would be a cool and welcome project. Join and groupBy have very clear semantics there, as those concepts exist directly in Trident. The extensions needed to Pig are the concept of incremental, persistent state across batches (mirroring those concepts in Trident). You can read a complete proposal.

  • implementing topologies in pure python with petrel looks like this:

class Bolt(storm.BasicBolt):
    def initialize(self, conf, context):
       ''' This method executed only once '''
        storm.log('initializing bolt')

    def process(self, tup):
       ''' This method executed every time a new tuple arrived '''       
       msg = tup.values[0]
       storm.log('Got tuple %s' %msg)

if __name__ == "__main__":
    Bolt().run()

  • Fliptop is happy with storm - see their presentation here

  • topology metrics in 0.9.0: The new metrics feature allows you to collect arbitrarily custom metrics over fixed windows. Those metrics are exported to a metrics stream that you can consume by implementing IMetricsConsumer and configure with Config.java#L473. Use TopologyContext#registerMetric to register new metrics.

  • storm vs flume - some users' point of view: I use Storm and Flume and find that they are better at different things - it really depends on your use case as to which one is better suited. First and foremost, they were originally designed to do different things: Flume is a reliable service for collecting, aggregating, and moving large amounts of data from source to destination (e.g. log data from many web servers to HDFS). Storm is more for real-time computation (e.g. streaming analytics) where you analyse data in flight and don't necessarily land it anywhere. Having said that, Storm is also fault-tolerant and can write to external data stores (e.g. HBase) and you can do real-time computation in Flume (using interceptors)

That's all for this day - however, I'll keep on reading through storm-users, so watch this space for more info on storm development.

I've been reading through storm-users Google Group recently. This resolution was heavily inspired by Adam Kawa's post "Football zero, Apache Pig hero". Since I've encountered a lot of insightful and very interesting information I've decided to describe some of those in this post.

  • nimbus will work in HA mode - There's a pull request open for it already... but some recent work (distributing topology files via Bittorrent) will greatly simplify the implementation. Once the Bittorrent work is done we'll look at reworking the HA pull request. (storm’s pull request)

  • pig on storm - Pig on Trident would be a cool and welcome project. Join and groupBy have very clear semantics there, as those concepts exist directly in Trident. The extensions needed to Pig are the concept of incremental, persistent state across batches (mirroring those concepts in Trident). You can read a complete proposal.

  • implementing topologies in pure python with petrel looks like this:

class Bolt(storm.BasicBolt):
    def initialize(self, conf, context):
       ''' This method executed only once '''
        storm.log('initializing bolt')

    def process(self, tup):
       ''' This method executed every time a new tuple arrived '''       
       msg = tup.values[0]
       storm.log('Got tuple %s' %msg)

if __name__ == "__main__":
    Bolt().run()

  • Fliptop is happy with storm - see their presentation here

  • topology metrics in 0.9.0: The new metrics feature allows you to collect arbitrarily custom metrics over fixed windows. Those metrics are exported to a metrics stream that you can consume by implementing IMetricsConsumer and configure with Config.java#L473. Use TopologyContext#registerMetric to register new metrics.

  • storm vs flume - some users' point of view: I use Storm and Flume and find that they are better at different things - it really depends on your use case as to which one is better suited. First and foremost, they were originally designed to do different things: Flume is a reliable service for collecting, aggregating, and moving large amounts of data from source to destination (e.g. log data from many web servers to HDFS). Storm is more for real-time computation (e.g. streaming analytics) where you analyse data in flight and don't necessarily land it anywhere. Having said that, Storm is also fault-tolerant and can write to external data stores (e.g. HBase) and you can do real-time computation in Flume (using interceptors)

That's all for this day - however, I'll keep on reading through storm-users, so watch this space for more info on storm development.