[ 
https://issues.apache.org/jira/browse/ORC-161?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16434759#comment-16434759
 ] 

ASF GitHub Bot commented on ORC-161:
------------------------------------

Github user wgtmac commented on the issue:

    https://github.com/apache/orc/pull/245
  
    @t3rmin4t0r @omalley @majetideepak @xndai 
    Any suggestion or concern? If we can finalize this, I can start working on 
it.


> Create a new column type that run-length-encodes decimals
> ---------------------------------------------------------
>
>                 Key: ORC-161
>                 URL: https://issues.apache.org/jira/browse/ORC-161
>             Project: ORC
>          Issue Type: Wish
>          Components: encoding
>            Reporter: Douglas Drinka
>            Priority: Major
>
> I'm storing prices in ORC format, and have made the following observations 
> about the current decimal implementation:
> - The encoding is inefficient: my prices are a walking-random set, plus or 
> minus a few pennies per data point. This would encode beautifully with a 
> patched base encoding.  Instead I'm averaging 4 bytes per data point, after 
> Zlib.
> - Everyone acknowledges that it's nice to be able to store huge numbers in 
> decimal columns, but that you probably won't.  Presto, for instance, has a 
> fast-path which engages for precision of 18 or less, and decodes to 64-bit 
> longs, and then a slow path which uses BigInt.  I anticipate the majority of 
> implementations fit the decimal(18,6) use case.
> - The whole concept of precision/scale, along with a dedicated scale per data 
> point is messy.  Sometimes it's checked on data ingest, other times its an 
> error on reading, or else it's cast (and rounded?)
> I don't propose eliminating the current column type.  It's nice to know 
> there's a way to store really big numbers (or really accurate numbers) if I 
> need that in the future.
> But I'd like to see a new column that uses the existing Run Length Encoding 
> functionality, and is limited to 63+1 bit numbers, with a fixed precision and 
> scale for ingest and query.
> I think one could call this FixedPoint.  Every number is stored as a long, 
> and scaled by a column constant.  Ingest from decimal would scale and throw 
> or round, configurably.  Precision would be fixed at 18, or made configurable 
> and verified at ingest.  Stats would use longs (scaled with the column) 
> rather than strings.
> Anyone can opt in to faster, smaller data sets, if they're ok with 63+1 bits 
> of precision.  Or they can keep using decimal if they need 128 bits.  Win/win?



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to