Home Ask Login Register

Developers Planet

Your answer is one click away!

SinDeus February 2016

Write Parquet files from Spark RDD to dynamic folders

Given the following snippet (Spark version: 1.5.2):


which saves RDD data to flattened Parquet files, I would like my storage to have a structure like:


The data itself contains a country column and a timestamp one, so I started with this method. However, since I only have a timestamp in my data, I can't partition the whole thing by year/yearmonth/yearmonthday as those are not columns per se...

And this solution seemed pretty nice, except I can't get to adapt it to Parquet files...

Any idea?


SinDeus February 2016

I figured it out. In order for the path to be dynamically linked to the RDD, one first has to create a tuple from the rdd:

rdd.map(model => (model.country, model))

Then, the records will all have to be parsed, to retrieve the distinct countries:

val countries = rdd.map {
        case (country, model) => country

Now that the countries are known, the records can be written according to their distinct country:

countries.map {
    country => {
        val countryRDD = rdd.filter {
                case (c, model) => c == country
        countryRDD.toDF().write.parquet(pathToStorage + "/" + country)

Of course, the whole collection has to be parsed twice, but it is the only solution I found so far.

Regarding the timestamp, you will just have to do the same process with a 3-tuple (the third being something like 20160214); I went with the current timestamp finally.

Post Status

Asked in February 2016
Viewed 3,313 times
Voted 13
Answered 1 times


Leave an answer

Quote of the day: live life