[
https://issues.apache.org/jira/browse/BEAM-469?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15768476#comment-15768476
]
Daniel Halperin commented on BEAM-469:
--------------------------------------
Sorry I missed this JIRA comment, [~mariusz89016]! A bit late, but...
Say a coder C does not have the nested context. Then we actually have the
guarantee that no one will put later elements.
So if {{NullableCoder}} does not have the nested context, then no one will put
more elements after whatever the {{NullableCoder}} puts. If the NC puts {{0}}
then that's it -- the element is null. But if the NC puts {{1}}, then we know
that all remaining bytes in the encoded string belong to the inner coder. That
is effectively saying that the inner coder also does not need to have the
nested context, so it does not need to write its own length.
In your example, the {{NullableCoder}} is used in an inner context. So the
inner coder needs to also use the inner context, because there may be more
encoded elements later.
In either case: the context of the nullable coder can be the same as the
context of the inner coder. This is why in the patch here, we simply pass the
NC's context down into the inner coder. All we have removed is the _additional_
nesting that was used.
https://patch-diff.githubusercontent.com/raw/apache/incubator-beam/pull/992.patch
> NullableCoder optimized encoding via passthrough context
> --------------------------------------------------------
>
> Key: BEAM-469
> URL: https://issues.apache.org/jira/browse/BEAM-469
> Project: Beam
> Issue Type: Improvement
> Components: sdk-java-core
> Reporter: Luke Cwik
> Assignee: Thomas Groh
> Priority: Trivial
> Labels: backward-incompatible
> Fix For: 0.3.0-incubating
>
>
> NullableCoder should encode using the context given and not always use the
> nested context. For coders which can efficiently encode in the outer context
> such as StringUtf8Coder or ByteArrayCoder, we are forcing them to prefix
> themselves with their length.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
