[
https://issues.apache.org/jira/browse/NIFI-3248?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16152023#comment-16152023
]
Koji Kawamura commented on NIFI-3248:
-------------------------------------
[~jope] Thanks for the detailed proposal to improve GetSolr behavior. By
looking at the shared document about pagination and cursor, I agree it would
provide more reliable result to keep getting updated documents from Solr
(tailing updated docs) as written in the documentation.
1. Instead of using 'Date field', use '_version_' to sort. I agree. Just to
provide backward compatibility, we can make 'Date Field' optional and if 'Date
Field' is not specified, use '_version. Users would expect that existing
GetSolr processor to behave as it is.
2. Always iterating through result. I agree with this idea. It should provides
the best latency.
3. Use Cursor instead of 'start' parameter. Good idea. I looked at the GetSolr
and found that it does not use NiFi Managed State feature. Instead, it writes a
file under conf dir locally. We should use Managed State instead of local file.
Also, if we support backward compatibility, we need some logic to convert
existing last date to equivalent '_version_'. (probably by querying the stored
last date and use the last document '_version_' subsequently?)
4. I agree with the idea of using fq for better performance.
5. 'users should not be enabled to change the parameters sort and q'. I agree
with not letting user modify sort condition. But I imagine some use-cases are
only interested in particular set of documents to track updates. So I would
expect new implementation still support providing query options.
I have a question regarding to the '_version_' value. Is this globally unique
within a collection among all documents? The actual value looks like a unix
epoch in microseconds resolution. By using '_version_' as sort condition and
cursor, can we eliminate duplication or missing document completely?
{quote}
Cursor mark values are computed based on the sort values of each document in
the result, which means multiple documents with identical sort values will
produce identical Cursor mark values if one of them is the last document on a
page of results. In that situation, the subsequent request using that
cursorMark would not know which of the documents with the identical mark values
should be skipped. Requiring that the uniqueKey field be used as a clause in
the sort criteria guarantees that a deterministic ordering will be returned,
and that every cursorMark value will identify a unique point in the sequence of
documents.
{quote}
https://lucene.apache.org/solr/guide/6_6/pagination-of-results.html#using-cursors
> GetSolr can miss recently updated documents
> -------------------------------------------
>
> Key: NIFI-3248
> URL: https://issues.apache.org/jira/browse/NIFI-3248
> Project: Apache NiFi
> Issue Type: Bug
> Components: Extensions
> Affects Versions: 1.0.0, 0.5.0, 0.6.0, 0.5.1, 0.7.0, 0.6.1, 1.1.0, 0.7.1,
> 1.0.1
> Reporter: Koji Kawamura
> Assignee: Johannes Peter
> Attachments: nifi-flow.png, query-result-with-curly-bracket.png,
> query-result-with-square-bracket.png
>
>
> GetSolr holds the last query timestamp so that it only fetches documents
> those have been added or updated since the last query.
> However, GetSolr misses some of those updated documents, and once the
> documents date field value becomes older than last query timestamp, the
> document won't be able to be queried by GetSolr any more.
> This JIRA is for tracking the process of investigating this behavior, and
> discussion on them.
> Here are things that can be a cause of this behavior:
> |#|Short description|Should we address it?|
> |1|Timestamp range filter, curly or square bracket?|No|
> |2|Timezone difference between update and query|Additional docs might be
> helpful|
> |3|Lag comes from NearRealTIme nature of Solr|Should be documented at least,
> add 'commit lag-time'?|
> h2. 1. Timestamp range filter, curly or square bracket?
> At the first glance, using curly and square bracket in mix looked strange
> ([source
> code|https://github.com/apache/nifi/blob/support/nifi-0.5.x/nifi-nar-bundles/nifi-solr-bundle/nifi-solr-processors/src/main/java/org/apache/nifi/processors/solr/GetSolr.java#L202]).
> But these difference has a meaning.
> The square bracket on the range query is inclusive and the curly bracket is
> exclusive. If we use inclusive on both sides and a document has a time stamp
> exactly on the boundary then it could be returned in two consecutive
> executions, and we only want it in one.
> This is intentional, and it should be as it is.
> h2. 2. Timezone difference between update and query
> Solr treats date fields as [UTC
> representation|https://cwiki.apache.org/confluence/display/solr/Working+with+Dates|].
> If date field String value of an updated document represents time without
> timezone, and NiFi is running on an environment using timezone other than
> UTC, GetSolr can't perform date range query as users expect.
> Let's say NiFi is running with JST(UTC+9). A process added a document to Solr
> at 15:00 JST. But the date field doesn't have timezone. So, Solr indexed it
> as 15:00 UTC. Then GetSolr performs range query at 15:10 JST, targeting any
> documents updated from 15:00 to 15:10 JST. GetSolr formatted dates using UTC,
> i.e. 6:00 to 6:10 UTC. The updated document won't be matched with the date
> range filter.
> To avoid this, updated documents must have proper timezone in date field
> string representation.
> If one uses NiFi expression language to set current timestamp to that date
> field, following NiFi expression can be used:
> {code}
> ${now():format("yyyy-MM-dd'T'HH:mm:ss.SSSZ")}
> {code}
> It will produce a result like:
> {code}
> 2016-12-27T15:30:04.895+0900
> {code}
> Then it will be indexed in Solr with UTC and will be queried by GetSolr as
> expected.
> h2. 3. Lag comes from NearRealTIme nature of Solr
> Solr provides Near Real Time search capability, that means, the recently
> updated documents can be queried in Near Real Time, but it's not real time.
> This latency can be controlled by either on client side which requests the
> update operation by specifying "commitWithin" parameter, or on the Solr
> server side, "autoCommit" and "autoSoftCommit" in
> [solrconfig.xml|https://cwiki.apache.org/confluence/display/solr/UpdateHandlers+in+SolrConfig#UpdateHandlersinSolrConfig-Commits].
> Since commit and updating index can be costly, it's recommended to set this
> interval long enough up to the maximum tolerable latency.
> However, this can be problematic with GetSolr. For instance, as shown in the
> simple NiFi flow below, GetSolr can miss updated documents:
> {code}
> t1: GetSolr queried
> t2: GenerateFlowFile set date = t2
> t3: PutSolrContentStream stored new doc
> t4: GetSolr queried again, from t1 to t4, but the new doc hasn't been indexed
> t5: Solr completed index
> t6: GetSolr queried again, from t4 to t6, the doc didn't match query
> {code}
> This behavior should be at least documented.
> Plus, it would be helpful to add a new configuration property to GetSolr, to
> specify commit lag-time so that GetSolr aims older timestamp range to query
> documents.
> {code}
> // with commit lag-time
> t1: GetSolr queried
> t2: GenerateFlowFile set date = t2
> t3: PutSolrContentStream stored new doc
> t4: GetSolr queried again, from (t1 - lag) to (t4 - lag), but the new doc
> hasn't been indexed
> t5: Solr completed index
> t6: GetSolr queried again, from (t4 - lag) to (t6 - lag), the doc can match
> query
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
