[
https://issues.apache.org/jira/browse/AVRO-1720?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Janosch Woschitz updated AVRO-1720:
-----------------------------------
Attachment: AVRO-1720-with-extended-unittests.patch
I tried to change the visibility of the {{hasNextBlock}} method but
unfortunately this was not sufficient enough since the DataFileStream does not
provide a real iterator functionality for blocks.
The {{nextBlock}} method just forwards to the next block but does not
initializes it.
The loop would have looked like this:
{{code}}
count += streamReader.getBlockCount();
while (streamReader.hasNext() && streamReader.hasNextBlock()) {
streamReader.nextBlock();
count += streamReader.getBlockCount();
}
{{code}}
Therefore I decided against exposing {{hasNextBlock}} method as public and left
the code of the counter tool like it was before but I added an explanatory
comment and an additional unit test which covers the scenario with multiple
blocks.
The alternative would be a bigger refactoring of DataFileStream and the classes
which are already using the public {{nextBlock}} method. I refrained from doing
that since I think this does not fit the scope of this ticket.
> Add an avro-tool to count records in an avro file
> -------------------------------------------------
>
> Key: AVRO-1720
> URL: https://issues.apache.org/jira/browse/AVRO-1720
> Project: Avro
> Issue Type: New Feature
> Components: java
> Reporter: Janosch Woschitz
> Priority: Minor
> Attachments: AVRO-1720-with-extended-unittests.patch, AVRO-1720.patch
>
>
> If you're dealing with bigger avro files (>100MB) it would be nice to have a
> way to quickly count the amount of records contained within that file.
> With the current state of avro-tools the only way to achieve this (to my
> current knowledge) is to dump the data to json and count the amount of
> records. For bigger files this might take a while due to the serialization
> overhead and since every record needs to be looked at.
> I added a new tool which is optimized for counting records, it does not
> serialize the records and reads only the block count for each block.
> {panel:title=Naive benchmark}
> {noformat}
> # the input file had a size of ~300MB
> $ du -sh sample.avro
> 323M sample.avro
> # using the new count tool
> $ time java -jar avro-tools.jar count sample.avro
> 331439
> real 0m4.670s
> user 0m6.167s
> sys 0m0.513s
> # the current way of counting records
> $ time java -jar avro-tools.jar tojson sample.avro | wc
> 331439 54904484 1838231743
> real 0m52.760s
> user 1m42.317s
> sys 0m3.209s
> # the overhead of wc is rather minor
> $ time java -jar avro-tools.jar tojson sample.avro > /dev/null
> real 0m47.834s
> user 0m53.317s
> sys 0m1.194s
> {noformat}
> {panel}
> This tool uses the HDFS API to handle files from any supported filesystem. I
> added the unit tests to the already existing TestDataFileTools since it
> provided convenient utility functions which I could reuse for my test
> scenarios.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
