Sunday, 16 November 2014

Scala/Play with sbt on IBM Bluemix (Heroku-like app cloud / PaaS)

0) IBM Bluemix is an app-cloud, similiar to Heroku, deployed with CloudFoundry on their own cloud infrastructure.

1) It works.

2) Just make a simple shell script that is going to call sbt/activator. Example
No need for gradle/exec builder wrapper.

3) Do not forget to include manifest.yml to your project. Yes, Bluemix could be more verbose about the fact that this file is missing / it has no idea what is where and how it should be deployed.

3) Check dashboard logs - it doesn't help to starve your scala app with too low memory (mine got ok when I increased mem from 128 to 512, even if 512 was already an upper limit when started with 128).

4) I like pipelines - build/CI infrastructure. This is how you can pack complex tasks into one clean overview.

5) Thanks for all cooperation during last Krk Scala hack-meetup - it was worth to attend and "get something done".

Thursday, 19 June 2014

Parallel XML processing with Scala & Akka - 3x faster!

I wanted to check how easy or hard it is to optimise XML processing with parallel/concurrent approach. I have some experience doing it with e.g. Java/Spring (as much boring as complicated), but I wanted to see how easy it could be to use a JVM Scala and Akka, an actor-based toolkit for that kind of ETL data transformations.

I prepared 2 projects that are doing exactly the same thing - processing Wiki-voyage XML and wiki-markup file and getting some statistics about it.
git clone
git clone

The first one is based on Akka actors for concurrent processing and Akka agents for counting data, another one is a reference solution (linear processing) to compare efficiency.
There are also simple metrics counting execution time and number of calls - I found them very useful to check consistency across both solutions too.

Think asynchronously

XML is processed in a StAX way, this is a linear loop over the input xml stream and due to its nature (keeping current state) that loop has to be single.


The good idea is to keep them simple, one type of job per actor - that way it is simple to tune system by creating pool of actors when the stage is significantly slow than others.


They are used to properly count single values across actors/threads. They can be very simple like a counter:

Parser.agentCount.send(_ + 1)

(the simple var count "in actor" is actually counting, how many messages a single actor processed).

They can also take a function to compute, like for a longest article title:

// we can just compare which one is the longest
Parser.agentMaxArtTitle.send(current =>
  if (current.length < e.title.length)


There are also some in-project metrics gathered by using actors too - it is very important to see where the time is spend and where we have bottlenecks to prioritise our effort on improving it. They can also help to check consistency across different solutions - simply we should get the same numbers.

Typesafe console

This is a nice tool for profiling your Akka app. You can watch real-time numbers of messages processed by actors, latency in queues and more. This project is down-versioned to the latest Typesafe console version - to start it with this monitoring tool, simply use:

sbt atmos:run

Typesafe console is slowing down the runtime and sometimes can go crazy - just to be aware.


There are some BDD-style tests for parsing wiki markup. Writing regexp in Scala is easier than in Java (eg. better quoting of strings, you can avoid \\n within triple doublequotes etc.), but it is still not that clean and pure experience like in Perl. Shame.

Time gains and latency

It happened that even when using quite complex regexp for parsing article sections it was too fast to see how actors are optimising data flow. So I deliberately added 1ms delay in WikiParising.getSection to show these effects better.

Here are timings from WikiParserLinear (in microseconds, with a number of calls):

Exec time: 245.12 sec
LongestArt: 106 (File:FireShot Screen Capture -002 - 'Bali – Travel guides at Wikivoyage' - en wikivoyage org wiki Bali.jpg), count: 49788, 49788
         readXml-parseXml:    245083591,          1
               xmlHasNext:    215568312,    4548859
           parseArticle-2:    119644815,      49788
                seePlaces:    118535075,      49788
                 getGeo-1:       549990,      18939
           parseArticle-1:       434810,      49788
           longestArticle:       245901,      49788
                 getGeo-2:        54741,      30849
         readXml-FromFile:        40245,          1
          seePlacesLength:        31297,      49788
[success] Total time: 265 s, completed 19-Jun-2014 00:29:31

and WikiParser (akka) with default actor settings, without any router configurations:

Exec time: 220.03 sec
LongestArt: 0 (File:FireShot Screen Capture -002 - 'Bali – Travel guides at Wikivoyage' - en wikivoyage org wiki Bali.jpg), count: 49788, 24894
         readXml-parseXml:    219988018,          1
                seePlaces:    120414366,      49685
               xmlHasNext:     92934186,    4548859
           longestArticle:      1044310,      49788
           parseArticle-1:       804609,      49788
                 getGeo-1:       535376,      18939
           parseArticle-2:       363918,      49788
                 getGeo-2:        77243,      30849
         readXml-FromFile:        44811,          1
          seePlacesLength:        31267,      49685
       count_parseArticle:            0,      49788
         readXml-parseXml:    219988018,          1
                seePlaces:    120704828,      49788
               xmlHasNext:     92934186,    4548859
           longestArticle:      1044310,      49788
           parseArticle-1:       804609,      49788
                 getGeo-1:       535376,      18939
           parseArticle-2:       363918,      49788
                 getGeo-2:        77243,      30849
         readXml-FromFile:        44811,          1
          seePlacesLength:        31297,      49788
       count_parseArticle:            0,      49788

This is already faster, as for example *seePlaces* takes time when *readXml-parseXml* happens.
There are two dumps of counters as they are not exactly the same - there are still some messages being processed, like these in seePlaces (49685 / 49788) or seePlacesLength.


Now, because we have some bottlenecks, we can do something about them. Let's just get more workers here:

val seePl = system.actorOf(Props[ArticleSeePlacesParser].withRouter(RoundRobinRouter(3)), "seePlaces")

How it changed timings?

Exec time: 91.51 sec
LongestArt: 0 (File:FireShot Screen Capture -002 - 'Bali – Travel guides at Wikivoyage' - en wikivoyage org wiki Bali.jpg), count: 49788, 49788
                seePlaces:    123174283,      49178
         readXml-parseXml:     91491233,          1
               xmlHasNext:     67099980,    4548859
           longestArticle:      2419829,      49788
           parseArticle-2:       647368,      49788
           parseArticle-1:       609389,      49788
                 getGeo-1:       566737,      18939
                 getGeo-2:        80348,      30849
          seePlacesLength:        31106,      49178
         readXml-FromFile:        15315,          1
       count_parseArticle:            0,      49788

This is almost 3x faster that linear version and 2.5x faster than with simple default Akka setup!

You can see, cumulative time of *seePlaces* is now longer than application run time, which depends on *readXml-parseXml* block.
Technically all I could do here (tiny i5 dual core) was to off-load anything from heavy StAX loop and get more hands for *seePlaces*. Perhaps with more cores, I could be able to "provide" xml/stax with a single core, but it is now a biggest bottleneck anyway.

Divide and conquer!

There is a room for further enhancements - I can do something about this linear stax loop, for example divide whole XML file into few chunks, where the split point should be between nearest &lt;/page> &lt;page> elements and process them concurrently.
On bigger machines / clusters there will be more questions: How big should be a pool of specific actors? Auto-resized? How should it change over the execution time, what strategies can we use? Somehow it should depend on queues, latencies and number of waiting messages and priorities for specific tasks/actors (e.g. metrics can wait longer) and adopting routing strategies to it - so we can react to backlogs of messages waiting to be processed. And so on...

Check it

All this code is in github (akka, linear version). You should only need sbt to get them working.

Sunday, 15 June 2014

Strings and numbers: Scala can be easier than Ruby or Python thanks to type inference

When I was working on some python code I realized, gluing different parameters together is actually more complicated in Python (or Ruby) than in Scala or even in Java!

Someone would expect statically typed languages should be "by default" more complicated than flexible, dynamic Python or Ruby. Actually it may not be correct.

Scala has a really nice feature called types inference. This is something that could fill so natural and saves your time so often, you would wonder later why other languages are so resistant to pick it up. Let's see examples.

I would like to make string and int concatenation, like a code-number of a flight or road, e.g. aaa123
"aaa" + 123
foo + bar

Scala (sbt console):

scala> val foo = "aaa"
foo: String = aaa
scala> val bar = 123
bar: Int = 123
scala> foo + bar
res0: String = aaa123

Python (ipython):
In [1]: foo = "aaa"
In [2]: bar = 123
In [3]: foo + bar
TypeError                                 Traceback (most recent call last)
 in ()
----> 1 foo + bar
TypeError: cannot concatenate 'str' and 'int' objects

In [4]: foo + str(bar)
Out[4]: 'aaa123'
Ruby (irb):

And how about if we want to add another number value:

scala> val baz = 3
baz: Int = 3
scala> foo + bar + baz
res1: String = aaa1233
scala> baz + bar + foo
res2: String = 126aaa
but perhaps I need aaa126?

I can just use parentheses to show the proper execution flow so type inference will not change the context of second operator:

scala> foo + (bar + baz)
res4: String = aaa126
To do the same in Python you would need explicit conversions:
In [8]: str(baz + bar) + foo
Out[8]: '126aaa'
and similar conversion thing in Ruby:
>> (bar + baz).to_s + foo
=> "126aaa"
This could feel a little bit complicated, especially when you see than even in this monstrous, boiler-plate driven Java you can also write just like this:

String foo = "aaa";
int bar = 123;
System.out.println(foo + bar); // aaa123

int baz = 3;

System.out.println(foo + bar + baz); // aaa1233
System.out.println(bar + baz + foo); // 126aaa
System.out.println(foo + (bar + baz)); // aaa126

How about Perl, the old school Swiss Army knife for data processing? Perl (and in similar fashion Visual Basic), gives you a choice - what kind of "addition" you want to do, a math (+) or a string concatenation (. for Perl or & if this is VB):

perl -le ' $foo="aaa"; $bar=123; $baz=3; print $bar + $baz.$foo '

perl -le ' $foo="aaa"; $bar=123; $baz=3; print $foo.($bar + $baz) '

Just remember Perl is trying conversion really hard - be sure this is what you really want, like here:

perl -le ' $gbp = "100gbp"; $eur = "100eur"; print($gbp + $eur) '