[
https://issues.apache.org/jira/browse/ARROW-17008?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17563995#comment-17563995
]
Charlie Gao commented on ARROW-17008:
-------------------------------------
Ok sure. R simply wraps C ints (4bytes) and C doubles (8 bytes). Just a bit
puzzling how the double version managed to compress to 4.3MB vs 4.6MB for int
on default settings.
I am actually seeing some compression savings now using parquet-tools with
use-dictionary = FALSE. I guess this is equivalent to Spark now and it is
probably the optimal setting for my particular data.
Thanks for your help.
> [R] Parquet Snappy Compression Fails for Integer Type Data
> ----------------------------------------------------------
>
> Key: ARROW-17008
> URL: https://issues.apache.org/jira/browse/ARROW-17008
> Project: Apache Arrow
> Issue Type: Bug
> Components: R
> Affects Versions: 8.0.0
> Environment: R4.2.1 Ubuntu 22.04 x86_64
> R4.1.2 Ubuntu 22.04 Aarch64
> Reporter: Charlie Gao
> Priority: Major
>
> Snappy compression is not working when writing to parquet for integer type
> data.
> E.g. compare file sizes for:
> {code:r}
> write_parquet(data.frame(x = 1:1e6), "snappy.parquet", compression = "snappy")
> write_parquet(data.frame(x = 1:1e6), "uncomp.parquet", compression =
> "uncompressed")
> {code}
> whereas for double:
> {code:r}
> write_parquet(data.frame(x = as.double(1:1e6)), "snappyd.parquet",
> compression = "snappy")
> write_parquet(data.frame(x = as.double(1:1e6)), "uncompd.parquet",
> compression = "uncompressed")
> {code}
> I have inspected the integer files using parquet-tools and compression level
> shows as 0%. Needless to say, I can achieve compression using Spark
> (sparklyr) etc.
> Thanks.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
