Hi Xinh
sorry for my late reply
it`s slow because of two reasons (at least to my knowledge)
1. lots of IOs - writing as json, then reading and writing again as parquet
2. because of nested rdd I can`t run the cycle and filter by event_type
in parallel - this applies to your solution (3rd step)
I ended up with the suggestion you proposed - in realtime partition by
event_type and store as jsons (which is pretty fast) and with another
job which runs less frequently read jsons and store them as parquet
thank you very much
best regards
Michal
On 05/05/2016 06:02 PM, Xinh Huynh wrote:
Hi Michal,
Why is your solution so slow? Is it from the file IO caused by storing
in a temp file as JSON and then reading it back in and writing it as
Parquet? How are you getting "events" in the first place?
Do you have the original Kafka messages as an RDD[String]? Then how about:
1. Start with eventsAsRDD : RDD[String] (before converting to DF)
2. eventsAsRDD.map() --> use a RegEx to parse out the event_type of
each event
For example, search the string for {"event_type"="[.*]"}
3. Now, filter by each event_type to create a separate RDD for each
type, and convert those to DF. You only convert to DF for events of
the same type, so you avoid the NULLs.
Xinh
On Thu, May 5, 2016 at 2:52 AM, Michal Vince <vince.mic...@gmail.com
<mailto:vince.mic...@gmail.com>> wrote:
Hi Xinh
For (1) the biggest problem are those null columns. e.g. DF will
have ~1000 columns so every partition of that DF will have ~1000
columns, one of the partitioned columns can have 996 null columns
which is big waste of space (in my case more than 80% in avg)
for (2) I can`t really change anything as the source belongs to
the 3rd party
Miso
On 05/04/2016 05:21 PM, Xinh Huynh wrote:
Hi**Michal,
For (1), would it be possible to partitionBy two columns to
reduce the size? Something like partitionBy("event_type", "date").
For (2), is there a way to separate the different event types
upstream, like on different Kafka topics, and then process them
separately?
Xinh
On Wed, May 4, 2016 at 7:47 AM, Michal Vince
<vince.mic...@gmail.com <mailto:vince.mic...@gmail.com>> wrote:
Hi guys
I`m trying to store kafka stream with ~5k events/s as
efficiently as possible in parquet format to hdfs.
I can`t make any changes to kafka (belongs to 3rd party)
Events in kafka are in json format, but the problem is there
are many different event types (from different subsystems
with different number of fields, different size etc..) so it
doesn`t make any sense to store them in the same file
I was trying to read data to DF and then repartition it by
event_type and store
events.write.partitionBy("event_type").format("parquet").mode(org.apache.spark.sql.SaveMode.Append).save(tmpFolder)
which is quite fast but have 2 drawbacks that I`m aware of
1. output folder has only one partition which can be huge
2. all DFs created like this share the same schema, so even
dfs with few fields have tons of null fields
My second try is bit naive and really really slow (you can
see why in code) - filter DF by event type and store them
temporarily as json (to get rid of null fields)
val event_types =events.select($"event_type").distinct().collect() //
get event_types in this batch
for (row <- event_types) {
val currDF =events.filter($"event_type" === row.get(0))
val tmpPath =tmpFolder + row.get(0)
currDF.write.format("json").mode(org.apache.spark.sql.SaveMode.Append).save(tmpPath)
sqlContext.read.json(tmpPath).write.format("parquet").save(basePath)
}
hdfs.delete(new Path(tmpFolder),true)
Do you have any suggestions for any better solution to this?
thanks
