[
https://issues.apache.org/jira/browse/FLUME-2649?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14497969#comment-14497969
]
Francis commented on FLUME-2649:
--------------------------------
My understanding is that it's not an ES limitation. The ES API expects valid
JSON, which is kind of normal (its responsibility is to format and send data to
ES, not pre-validate it). The parser returned by the API is actually a Jacskon
JSON parser and it's been designed to parser data on the fly, without having
the possibility to reuse already parsed data. They had a feature request for
this and specifically said they didn't want to do this:
https://jira.codehaus.org/browse/JACKSON-186
So, there's no other simple way to validate JSON. As I said in the code review,
we could at least not instantiate the tmp jsonBuilder and validate the JSON
using this code instead:
while(parser.nextToken() != null) { }
and then, reset the parser:
parser = XContentFactory.xContent(contentType).createParser(data);
> Elasticsearch sink doesn't handle JSON fields correctly
> -------------------------------------------------------
>
> Key: FLUME-2649
> URL: https://issues.apache.org/jira/browse/FLUME-2649
> Project: Flume
> Issue Type: Bug
> Components: Sinks+Sources
> Reporter: Francis
> Assignee: Benjamin Fiorini
> Attachments: FLUME-2649-0.patch, FLUME-2649-1.patch,
> FLUME-2649-2.patch, FLUME-2649-3.patch, FLUME-2649-4.patch
>
>
> JSON attributes are treated like normal strings and are escaped by the sink.
> For example, if the body or a header contains the following value:
> {code:javascript}
> {"foo":"bar"}
> {code}
> It will be added like this in Elasticsearch:
> {code:javascript}
> {"@message": "{\"foo\":\"bar\"}}"
> {code}
> We end up with a plain string instead of a valid JSON field.
> I think I found how to fix this bug. The source of the problem is caused by
> the way a "complex field" is added. The ES XContent classes are used to parse
> the data in the detected format, but then, instead of adding the parsed data,
> the string() method is called and it converts it back to a string that is the
> same as the initial data! Here is the current code with added comments:
> {code}
> XContentBuilder tmp = jsonBuilder(); // This tmp builder is completely
> useless.
> parser = XContentFactory.xContent(contentType).createParser(data);
> parser.nextToken();
> tmp.copyCurrentStructure(parser); // This copies the whole parsed data in
> this tmp builder.
> // Here, by calling tmp.string(), we get the parsed data converted back to a
> string.
> // This means that tmp.string() == String(data)!
> // All this parsing for nothing...
> // And then, as the field(String, String) method is called on the builder,
> and the builder being a jsonBuilder,
> // the string will be escaped according to the JSON specifications.
> builder.field(fieldName, tmp.string());
> {code}
> If we really want to take advantage of the XContent classes, we have to add
> the parsed data to the builder. To do this, it is as simply as:
> {code}
> parser = XContentFactory.xContent(contentType).createParser(data);
> parser.nextToken();
> // Add the field name, but not the value.
> builder.field(fieldName);
> // This will add the whole parsed content as the value of the field.
> builder.copyCurrentStructure(parser);
> {code}
> I tried this and it works as expected.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
